Latest / 1 More Paper / Human-Centered Artificial Intelligence: Design Principles for Public Services — Schmager (2025)

Transcript

Human-Centered Artificial Intelligence: Design Principles for Public Services — Schmager (2025) · 1 More Paper

• You’re listening to “Human-Centered Artificial Intelligence: Design Principles for Public Services,” by Stefan Schmager, from University of Agder. Published in 2025. Human-Centered Artificial Intelligence: Design Principles for Public Services Stefan Schmager Human-Centered Artificial Intelligence: Design Principles for Public Services Stefan Schmager Human-Centered Artificial Intelligence: Design Principles for Public Services Dissertation for the degree philosophiae doctor (ph.d.) University of Agder Faculty of Social Sciences 2025 Doctoral dissertations at the University of Agder, No: 537 © Stefan Schmager, 2025 Printed by Make!Graphics Kristiansand Acknowledgements. “The philosophers have only interpreted the world in various ways; the point, however, is to change it.” Karl Marx in Eleventh Thesis on Feuerbach (1938) “Tears on my Gore-Tex” High Vis - Trauma Bonds (2022) As this doctoral thesis constitutes the end of my PhD journey, I would like to express some due words of gratitude to those who have accompanied and supported me throughout this challenging, revelatory, but rewarding path. First and foremost, I would like to thank my main supervisor, Polyxeni (Xenia) Vassilakopoulou. You have been much more than a supervisor to me. You were my guide, my motivator, and a counterpart who always met me at eye level. You provided empathy, recognition and understanding for my struggles, confusion and despair. You grounded me when I needed to belong, you built me up when I felt out of place and thoughts of alienation, loneliness and abandonment crept in. But you also always had an open mind to my ideas, backed my chosen paths, and endorsed my sometimes-unconventional approaches. Without your unconditional support, encouragement, and belief in my abilities, I would not have made it to the end of this journey. For that, I will forever be grateful. I also owe sincere gratitude to my second supervisor, Ilias Pappas. Your guidance, insight, and support have been crucial in shaping my research and refining my work. Your expertise and thoughtful feedback have made a significant impact on my work, and I am deeply appreciative of the time and effort you dedicated to my progress. Beyond academia, I have been fortunate to have an incredible circle of old and new friends, fellows, and companions, who stood by me through good times – and more importantly, in my darkest moments. There are too many to name, but I am eternally thankful to each one of you, for your support, solace, distraction, and sympathy when I needed it most. Lastly, I want to thank my family, who have been the backbone of my educational path from the very beginning. With patience and encouragement, you have always supported my choices and have been a steady source of strength for me. Thank you for always believing in me, even when I struggled to believe in myself. I owe this accomplishment to your love and reassurance. To all of you, thank you for making this journey possible. Abstract. Artificial Intelligence (AI) technologies increasingly integrate into all areas of everyday life, including healthcare, employment, and public services. But as much as AI is lauded as a great white hope for enabling humans to approach societal challenges, there are at least as many dystopian futures, describing gloom and doom scenarios. This ever-widening impact has spawned discussions, considerations, and warnings about responsible, ethical, and sustainable implementation of such advanced technologies. In light of the growing interest in comprehending AI's impact on society and its alignment with human values and needs, Human-Centered Artificial Intelligence (HCAI) has emerged as a promising approach to address such questions and concerns. HCAI is frequently used in the discourse on how to guide the development and deployment of AI in responsible and trustworthy ways. A human-centered approach to AI foregrounds human needs and values, fostering the understanding that technology is created to support humans, not rendering them superfluous. To operationalize specific recommendations about the fundamental nature and structure of an HCAI system, design principles have been established as a suitable tool across various disciplines such as engineering, economics, and information systems. Design principles can guide the implementation of AI systems ensuring that systems are developed with a clear understanding of their social implications, equipped to handle complex ethical dilemmas, and capable of delivering balanced benefits to both individuals and society as a whole. They aim to bridge the gap between abstract theoretical concepts and practical implementation, guiding the development and deployment of AI technologies within these companies' spheres of influence and beyond. Major technology actors including Microsoft, Apple, and Google are claiming to promote a responsible approach to AI by providing their own set of guidelines and recommendations. However, these guidelines inherit the profit-oriented pretext of their creators and are therefore mostly oriented toward commercial contexts. This commercial focus leaves a significant gap for guidance tailored to public services, where the stakes are uniquely high. This leads to the overarching research question of this doctoral research: What are Human-Centered AI Design Principles for public services? This doctoral research presents an Action Design Research study, approaching HCAI in public services through a Social Contract Theory lens. It draws on empirical investigations to explore both citizens' perspectives on AI usage in public services and the understanding, needs, and expectations of public servants. By acknowledging the multi-stakeholder nature of HCAI, this research highlights the complex dynamics between citizens, public servants, and AI systems in the public sector. The study recognizes an evolving understanding of prioritizing human values and well-being in AI technologies, specifically considering the relationship between citizens and governmental organizations in the context of AI systems in the public sector. It highlights the need to ensure that such AI systems are developed with a commitment to fostering societal and individual benefit and welfare. To reconcile abstract theoretical concepts with the practical realities of AI development and deployment in the public sector, HCAI design principles are formulated to contribute to theory and practice by advancing the scholarly discourse on HCAI and informing AI strategies and implementations within the public sector. Sammendrag Kunstig intelligens (KI) blir stadig mer integrert i alle deler av hverdagen, inkludert helsevesen, arbeidsliv og offentlige tjenester. Men like mye som kunstig intelligens hylles som et stort håp for å gjøre det mulig for mennesker å løse samfunnsutfordringer, finnes det minst like mange dystopiske fremtidsutsikter som beskriver dystre scenarier. Denne stadig mer omfattende påvirkningen har skapt diskusjoner, betraktninger og advarsler om ansvarlig, etisk og bærekraftig implementering av slike avanserte teknologier. I lys av den økende interessen for å forstå hvordan kunstig intelligens påvirker samfunnet og samsvarer med menneskelige verdier og behov, har Human-Centered Artificial Intelligence (HCAI) vokst frem som en lovende tilnærming til å håndtere slike spørsmål og bekymringer. HCAI brukes ofte i diskursen om hvordan man kan styre utviklingen og bruken av AI på en ansvarlig og pålitelig måte. En menneskesentrert tilnærming til kunstig intelligens setter menneskets behov og verdier i sentrum, og fremmer forståelsen av at teknologien er skapt for å støtte mennesker, ikke for å gjøre dem overflødige. For å operasjonalisere spesifikke anbefalinger om den grunnleggende karakteren og strukturen til et HCAI-system, har designprinsipper blitt etablert som et egnet verktøy på tvers av ulike fagområder som ingeniørfag, økonomi og informasjonssystemer. Designprinsipper kan veilede implementeringen av AI-systemer og sikre at systemene utvikles med en klar forståelse av de sosiale implikasjonene, at de er rustet til å håndtere komplekse etiske dilemmaer, og at de er i stand til å levere balanserte fordeler til både enkeltpersoner og samfunnet som helhet. De har som mål å bygge bro mellom abstrakte teoretiske konsepter og praktisk implementering, og veilede utviklingen og implementeringen av KI-teknologi innenfor og utenfor selskapenes innflytelsessfærer. Store teknologiaktører som Microsoft, Apple og Google hevder at de fremmer en ansvarlig tilnærming til kunstig intelligens ved å utarbeide egne retningslinjer og anbefalinger. Disse retningslinjene bærer imidlertid preg av det profittorienterte påskuddet til dem som har laget dem, og er derfor for det meste rettet mot kommersielle sammenhenger. Dette kommersielle fokuset etterlater et betydelig tomrom for retningslinjer som er skreddersydd for offentlige tjenester, der mye står på spill. Dette leder til det overordnede forskningsspørsmålet i denne doktorgradsstudier: Hva er menneskesentrerte AI-designprinsipper for offentlige tjenester? Denne doktorgradsstudier er en Action Design Research studie som tar for seg HCAI i offentlige tjenester gjennom en Social Contract Theory tilnærming. Den bygger på empiriske undersøkelser for å utforske både innbyggernes perspektiver på bruk av AI i offentlige tjenester og forståelsen, behovene og forventningene til offentlig ansatte. Ved å anerkjenne at HCAI involverer flere interessenter, belyser denne forskningen den komplekse dynamikken mellom innbyggere, offentlig ansatte og AI-systemer i offentlig sektor. Studien anerkjenner en utvikling i forståelsen av å prioritere menneskelige verdier og velvære i AI-teknologier, spesielt med tanke på forholdet mellom borgere og offentlige organisasjoner i forbindelse med AI-systemer i offentlig sektor. Studien understreker behovet for å sikre at slike AI-systemer utvikles med en forpliktelse til å fremme samfunnsmessige og individuelle fordeler og velferd. For å forene abstrakte teoretiske konsepter med de praktiske realitetene ved utvikling og bruk av kunstig intelligens i offentlig sektor, formuleres det designprinsipper for HCAI som skal bidra til teori og praksis ved å fremme den vitenskapelige diskursen om HCAI og informere om AI-strategier og -implementeringer i offentlig sektor. Table of Contents. List of Figures List of Tables 1. Introduction. There’s no question that the development of Artificial Intelligence (AI) moves at an unprecedented pace. Each week brings bold new announcements, product releases, or lofty claims that push the bounds of what was previously thought to be the frontiers of what is possible. Over the past years, there have been a myriad of predictions, prophecies and forecasts about what AI can and cannot do. Yet recent developments, both of General-Purpose AI but especially of Generative AI, have exceeded any expectations and have opened ever new pathways of imaginaries. AI systems become a fundamental part of day-to-day lives for millions of people. These technologies are now influencing domains such as healthcare, education, urban design, employment, and finance, among many others. One of the areas where AI is being more and more used is the public sector. In their comprehensive analysis, van Noordt and Misuraca (2022) found a vast amount of examples of AI being used in public sector organizations within the EU. A similar picture has been drawn in a recent report by the OECD. Approximately half of the public employment services in OECD countries are employing AI technologies to improve their services. AI is being integrated into various functions such as job matching, profiling jobseekers, and supporting administrative tasks. But AI is a two-edged sword. On the one hand, governments can use the potential of AI applications to improve public affairs and to increase the efficiency of internal processes. On the other hand, the threats indicate that AI needs policies and regulation based on principles and core societal values in order to bring benefits to everyone. In an era defined by rapid technological advancement, the intersection of AI and public services stands as a focal point of deep significance and complexity. The requirements for transparency, accountability, and inclusivity are more stringent, and the impact on civic life is more direct. And consequently, also the role of providing concrete guidance for the design, development, and implementation of AI becomes even more critical. As noted by Bekker (2021), the evolution of digital welfare states necessitates a robust set of tools, methods, and principles that are specifically crafted to meet the needs of the public sector. But despite AI's deepening integration into crucial areas of human and societal existence, the discussions surrounding its potential benefits and challenges has often sidelined the very human element at its core. Yet, there are also intellectual factions advocating for a different perspective. Concepts such as Responsible AI, Trustworthy AI and Human-Centered AI are gaining ever more attention within the discourse on AI. These concepts argue for a paradigm shift in the approach to technology, emphasizing the need to prioritize human values and well-being in the development and deployment of AI-enabled systems. This also implies a shift from viewing AI merely as a tool for efficiency gains to understanding it as a transformative force that requires careful consideration of its ethical, legal, and societal impact. This shift is not solely theoretical but is intended to influence how AI technologies are responsibly designed, developed, and implemented, ensuring to provide a genuine benefit for humanity. Human-Centered AI (HCAI) provides such a responsible approach toward AI design, enhancing and safeguarding human welfare, emphasizing AI systems that respect and amplify human values and capabilities. A HCAI approach demands the engagement with concerned stakeholders to align AI systems with the multifaceted spectrum of human needs and societal values. Stakeholder participation can ensure that AI systems are designed and implemented considering specific requirements and boundary conditions for AI adoption while preserving social functions. The perspectives and experiences of individuals affected by AI systems are invaluable, necessitating their active participation in the decision-making processes governing the application of AI technologies. Ensuring that these voices are heard and considered is not just beneficial but imperative for crafting AI policies and practices that are equitable, ethical, and truly reflective of the diversity of human needs and rights. Navigating the future of technological possibilities is a hazardous venture and it begins with learning to ask the right questions – questions that reveal the pitfalls of inaction, and more importantly, the passageways available for plotting a course to a safe harbor. One way to bridge the gap between abstract theoretical concepts and practical implementation, guiding the development and deployment of AI technologies can be found in the form of Design Principles. Design Principles can provide clear, actionable guidelines that help navigating the complexities involved in the ethical and effective implementation of AI technologies. Employing robust Design Principles already at the outset of the design and development process for an AI system helps to ensure that these systems are created with a keen awareness of their broader impacts. There are already many frameworks and guidelines that have been established to promote the responsible and human-centered design and development of AI systems. Leading technology corporations, including Microsoft, Apple, and Google, have participated in shaping the narrative and direction of AI development toward minimizing harm, by creating their own versions of HCAI guidelines and principles. However, the public sector's unique challenges and priorities necessitate a tailored approach to AI design and implementation, underscoring the need for frameworks that go beyond commercial aspirations and include the perspectives of citizen actors. Public service organizations are under the obligation to serve the public, as conceptually described in the Social Contract (Rousseau, 1762). This implies a variety of altering perspectives into the relationships between the involved actors. On the one hand, there is the requirement to weigh public good against individual benefit of citizens, balancing potentially conflicting interests. Furthermore, the concept of a market does not exist in the interaction between the governmental institutions and the citizens. 1.1. Research Questions. The sweeping advancements of AI technologies and the specific conditions and sensitivity of the public sector context gives rise to the following research question. RQ: What are Human-Centered AI Design Principles for public services? To develop the empirical grounding required to answer this overarching research question, three sub-questions are formulated (Figure 1). The first sub-question explores the applicability of existing commercial HCAI guidelines and how they would need to be adapted for the specific context of the public sector. This question is answered in publication #1. • Sub-RQ1: How can commercially oriented guidelines for human-centered AI be adapted for the context of public services? The second sub-question aims at understanding the perspective of citizens using AI systems in the public sector. This question is answered in publication #2. • Sub-RQ2: What are citizens’ perspectives towards the use of AI in public services in Norway? Sub-question three explores the perspectives of public servants towards using AI systems in their workplace. This question is answered in publication #3. • Sub-RQ3: What are public servants’ perspectives about the future of AI in their workplace? Furthermore, to answer the overarching research question, clarity on the concept of Human-Centered AI is also required. This includes an understanding of how it is defined in the literature, analyzing its constituting elements and the processes, methods and tools used to achieve HCAI. This HCAI conceptual clarification and grounding is presented in publications #4 and #5 (Figure 1). To fully answer the overarching research question, the conceptual and empirical grounding are brought together to formulate actionable HCAI Design Principles. These are included in the publications #6 and #7. Overall, this doctoral research has yielded several significant contributions to both practice and theory. The practical contributions are threefold: First, the formulation of HCAI Design Principles for public services, offering practitioners actionable guidance for AI systems in the public sector. Second, the development of an innovative artifact that provides the collaborating organization a concrete instantiation of human-centered design for communicating the use of AI with citizens. Third, it provides valuable insights into the perspectives of the affected stakeholder groups regarding AI adoption in public services. On the theoretical side, this thesis advances the HCAI discourse by offering conceptual clarity on the definition and application of Human-Centered AI. Additionally, it introduces Social Contract Theory as a novel theoretical framework for examining AI in public services and makes methodological contributions by showcasing the development of Design Principles on the basis of existing design guidelines, as well as proposing the use of Foresight approaches within Action Design Research to identify stakeholder perspectives and needs. 1.2. Structure of the Thesis. This doctoral thesis is comprised of an integrative essay (Kappa), consisting of ten chapters, as well as seven publications. After introducing the motivation and rationale of this research project in Chapter 1, the thesis is structured as follows (Figure 2). Chapter 2 presents the relevant literature and the current landscape of research. It covers the fields of AI in public services, Human-Centered AI as an approach for a responsible implementation of AI systems as well as Design Principles as a way to operationalize AI guidelines, considerations and best practices. In Chapter 4, the specific research context of the doctoral research is described. It provides a brief illustration of the role of national culture for AI transparency in public services, the Norwegian Guidelines on AI Transparency as well as a description of the collaborating organization, the Norwegian Labor and Welfare Administration (NAV). Chapter 5 outlines the employed research design of this thesis. The methodological approach of Action Design Research is described, followed by a detailed description of the two research strands, the creation of an innovative artifact for addressing the problem-instance and finally the Design Principles formulation process. Chapter 6 summarizes the seven included research publications of this thesis, their findings and individual contributions. It also provides an overview on how each publication contributes to the general aims of the thesis. In Chapter 7 the latest iterations of the defined Design Principles for both stakeholder groups are presented – the citizen-focused design principles as well as the employee-focused design principles. Chapter 8 discusses the emergence and definition of the design principles against the prior literature as well as the conceptual boundary conditions of the SCT framework. Subsequently, a research agenda in the form of an extended SCT framework is described. Chapter 9 provides an overview about the contributions of this doctoral research. Practical contributions include an innovative artifact to address the problem instance, insights for public services from stakeholder participation and HCAI Design principles for the public sector. Furthermore, the chapter presents theoretical contributions to the scientific discourse on HCAI, Social Contract Theory as a novel theoretical approach for AI in public services as well as contributions to the Action Design Research methodology. Finally, Chapter 10 provides a concluding summary of the thesis, together with the acknowledged limitations of this research. 2. Related Literature. In this chapter, related literature and the current research landscape is presented. The chapter covers the topics of AI within a public service context, Human-Centered AI as an approach for a responsible implementation of AI systems, and Design Principles as a way to operationalize AI guidelines, considerations and best practices. 2.1. AI in Public Services. In general, AI needs be understood as a complex socio-technical system, encompassing the complex interplay between technology, human actors, and societal structures. Such an perspective recognizes that AI systems are not merely standalone technologies, but are deeply embedded within and influenced by the social, cultural, and institutional contexts in which they operate. This is particularly true for the use of AI systems within the public sector. As Janssen and Kuk (2016) argue, the implementation of AI in public services involves a web of interdependencies between technological infrastructures, organizational processes, and human behaviors. These interactions shape and are shaped by the broader societal landscape, including the power dynamics within the public sector, as well as access to technological resources. In this section, the current state of AI in public services is outlined, followed by a discussion of the key concerns associated with this trajectory. To address these concerns, the concept of socially responsible AI is introduced, highlighting essential factors such as transparency and explainability. Central to this understanding is the recognition that AI systems in public services are not neutral tools, but rather reflect and potentially amplify existing social structures and power relations. As Eubanks (2018) demonstrates in her work on automated decision-making, the design and deployment of AI systems in social services can reinforce existing inequalities and biases if such systems are not carefully managed. This underscores the critical importance of considerations about who has the power to decide which processes will be automated, what data utilized for model training, and how interactions between decision-makers and affected populations are structured and maintained. Furthermore, the integration of AI into public services raises fundamental questions about governance, accountability, and democratic participation. Veale and Brass (2019) highlight the need for new models of public engagement and oversight to ensure that AI systems align with public values and serve the common good. By recognizing AI as a socio-technical system, we can better navigate the complexities and challenges of harnessing its potential for future developments while mitigating risks and ensuring its alignment with democratic principles and societal values. 2.1.1. Current State of the Use of AI in Public Services. Although the use of AI in public services is still considered to be in its early stages, there are multiple examples of potential uses of AI systems in the public service context. A recent analysis by van Noordt et al. (2024) documented 549 use cases of AI-based services in the public sectors among EU Member States and several other European countries. AI has also been introduced into the public sector beyond the prevailing western centric perspective, for instance in Malaysia, the UAE or local governments in China. The use case variety of different types of AI systems in the public sector is huge. There are systems used to improve efficiency by automating routine tasks to identify patterns for service delivery and enhanced decision making (de Sousa et al., 2019; van Noordt & Misuraca, 2022; Wirtz & Müller, 2019). AI is used for fraud detection, corruption, and whistleblowing, which can lead to more transparent and accountable governance. Another application of AI technologies are public service chatbots that can enhance citizen interaction and engagement by providing governmental services at any time. Furthermore, AI can also be used in the development of smart cities, managing traffic flow, optimizing energy consumption, and creating sustainable urban environments. Other examples can be found within the healthcare sector, where AI has also been explored to support transformation efforts, from assisting physicians with diagnoses to streamlining administrative processes, offering a range of applications to improved care as well as cost savings. And finally, law enforcement has been using AI to identify on which local areas they need to focus efforts to prevent crime. AI technologies are rapidly being developed and deployed by governments to enhance public services operations, and compliance and security activities. The use of AI has huge potential to enable public organizations getting a better understanding of citizens and provide better services by personalizing their offerings. There are significant opportunities provided by AI systems in public services, as they have potential access to extensive amounts of data which can be harnessed for the development of AI systems. However, research on the existing challenges of adopting AI for public service delivery has pointed to a variety of barriers. These barriers relate to AI-specific capabilities including the management of algorithmic performance and data governance, technical and managerial capabilities, and regulatory hurdles. 2.1.2. Concerns of Using AI in Public Services. While the potential benefits of AI are significant, there are also numerous challenges and risks. There are multiple examples of cases where the use of AI in the context of public services has received extensive criticism. AI systems are particularly susceptible to biases inherent in the data they are trained on, which can lead to discriminatory outcomes and erode public trust. Citizen concerns and controversies have halted several public service AI initiatives after their launch. In 2020, the Austrian labor administration (AMS) developed an AI system to categorize job seekers by their likelihood of reintegration into the workforce. The system was criticized for conducting algorithmic profiling of job seekers to increase the efficiency of its counseling process and the effectiveness of active labor market programs. Based on a statistical model of the job seekers' prospects on the labor market, the system was designed to classify them into three categories: those with high chances to find a job within half a year, those with mediocre prospects on the job market, and those clients with a bad outlook of employment in the next two years. Depending on the category a particular job seeker is classified under, they would be offered differing support in (re)entering the labor market. A coalition of civil society organizations started a campaign to rescind the system, and it has been declared illegal by the Austrian data protection agency. Another example can be found in the Dutch System Risk Indication (SyRI) which was an initiative to detect fraud with social services by profiling citizens. In response to that, civil society organizations and individuals convened and filed a freedom of information request asking questions about its workings and use. A court found that the SyRI Act is in violation of the European Convention on Human Rights (ECHR) thus calling for ceasing its further use. The plaintiffs contended that citizens were unaware of the risk model or algorithm that was employed and therefore unable to defend themselves against decisions made using SyRI. This lack of transparency in AI decision-making processes, often referred to as the black box problem, raises ethical concerns and challenges accountability. Therefore, it is crucial for citizens to know that AI is used in a transparent and accountable manner upholding all established rules for public service delivery. In Brazil, several legal organizations filed a lawsuit demanding a halt on the use of face recognition technology for public safety. They claimed that the technology violated citizens' fundamental rights to free speech, assembly, and association, and would worsen issues such as inequality and structural racism. Overall, key concerns expressed about the use of AI systems in the public sector context relate to training on inappropriate datasets that may perpetuate or even amplify biases, harming service impartiality, and black-boxing the inner workings of such systems. This urgently raises the issue on how we can ensure socially responsible and legitimate use of AI in public services 2.1.3. Socially Responsible AI. Although, the awareness about ethical importance of algorithms and AI has been growing, the general public is unaware of even the most fundamental aspects of AI. Applying AI in the public sector necessitates careful consideration of ethical implications, ensuring fairness, accountability, and human oversight. AI systems for public services should be designed and implemented in ways that do not infringe on core values and citizen rights. Public service organizations, as the name suggests, are under a duty to serve the public. They obtain their legitimization by the consent of the public, which implies their interests should be driven by the benefit of society and the collective good. Citizens expect government services to fulfill public missions and implement policies in the interest of the public. A responsible development and use of AI in the public service context therefore requires the balancing act of harnessing its power while minimizing risks for individuals and society. The special relationship between the public – the ruled – and public governance – their rulers – makes the use of AI technologies in the public sector particularly sensitive. Scholars from various disciplines have examined this predicament from different perspectives. Benedikt et al. (2020) examined the intricacies of having humans in-the-loop within the public sector context. They found that human interventions are needed if the goal is to find a balance between efficiency and data quality. In their examination of possible adverse consequences of algorithmic decision-making (ADM) within the public sector, Rinta-Kahila et al. (2022) investigated an ADM system implemented by the Australian government. The system automatically calculates and collects welfare overpayment debts from citizens. They found that the system inadvertently led to significant distress among both citizens and employees of the public service organization. The destructive effects ultimately harmed the government, both in the dimensions of public acceptance as well as trust of AI-driven systems used in the public sector. Factors such as transparency, explainability, and perceived control influence public trust in these technologies. 2.1.4. Transparency and Explainability. The concept of transparency can be defined as the availability of information about an organization or actor that allows external actors to monitor the internal workings or performance of that organization. Within the context of public services, transparency into the inputs and inner workings of AI is considered key. This goes beyond mere data availability, as it relates to the openness and clarity of decision-making processes, algorithms, data sources, and intentions behind AI systems. Transparency can therefore be described as an umbrella term encompassing several overlapping and associated concepts. In the context of AI within organizations, transparency in AI caters to the right and ability to obtain information about the use of AI. It encompasses the ability to understand, scrutinize, and monitor the internal workings and decision-making processes of AI systems. Transparency is also addressed in laws and regulations, such as the General Data Protection Regulation (GDPR) or the recently passed EU AI Act. At the same time, transparency pertains different meanings in different sectors. As Schulze-Weddige and Zylowski (2021) state, the first step of designing an AI system is to fully understand the context of use, i.e., users, stakeholders, and the environment of the system. In a framework to assess system transparency, Hepenstal et al. (2021) also found that an understanding of context is crucial. Within the context of public services, transparency is an important factor for maintaining public trust and accountability in government operations. As such, leveraging AI to ensure transparency in public services has the potential to enable personalization, support decision-making processes and a more efficient use of resources in public service delivery. Furthermore, it is also considered as a vital aspect of maintaining a balance of power between the public service authorities and citizens thereby increasing the likelihood of exposing wrong-doings or abuse of power. It can serve as a mechanism against potential pitfalls, such as poor governance, flawed decision-making, lack of accountability, and the spread of misinformation. Proponents of transparency stress that transparency fosters understanding, a tighter bond between citizens, and a better understanding of public services. To achieve such relationships, an AI system must provide explanations for outputs together with transparency of the underlying system processes and data, allowing inspection by its stakeholders. However, it is also acknowledged that achieving full AI transparency in public services and its practical realization is a complex challenge. One such challenge relates to the socio-technical constraints that complicate the process of creating meaningful transparency from open public service data. Additionally, ensuring data privacy and security while making AI systems transparent poses significant difficulties, as protecting citizens’ personal data may conflict with the goal of full transparency. Moreover, the effort to simplify AI decision-making for greater accessibility risks oversimplifying complex AI logic. Such reductions could strip away critical nuances, potentially misleading citizens or exposing vulnerabilities in the system that hackers or malicious actors could exploit. On the other hand, some scholars argue that an excess of transparency doesn’t necessarily increase citizens' trust but could potentially even delegitimize the administration. They claim that most of the explanatory material on AI systems provided by public services is difficult to understand even for subject-matter experts, the lack of AI literacy in the general citizenry just increases ambiguity and a lack of trust. A related concept to enable Transparency, can be found in explainability. The core idea of explainability is meaning making, concerning the ability to make sense of the inner workings of an AI system. Explainable AI (XAI) is primarily discussed within the field of computer science. Researchers in this field focus on developing algorithms that make complex AI models transparent and provide explanations for how they work. This becomes particularly important in situations when the understanding of failures and unexpected AI behavior is necessary. Doran et al. (2017) assert that to allow an evaluation of the ethical and moral standards of a machine, explanations should provide insight into the rationale of the AI system and enable users to draw conclusions based on them. Bove et al. (2022) show that the contextualization of explanations can improve user satisfaction and the user’s objective understanding. However, as found by Gajos and Mamykina (2022) in their study on incidental learning, it may not be sufficient to just include explanations together with AI-generated recommendations and assume that people engage carefully with the AI-provided information. Steels (2020) argues that a Human-Centered approach to AI, with its implications of explainability, transparency will only be possible when AI comes to grips with meaning and understanding. 2.2. Human-Centered Artificial Intelligence. The increasing integration of AI systems into public services has brought more and more attention to the discourse about the need for responsible and ethical development approaches. This echoes longstanding debates about the neutrality or political nature of technology. But the seemingly accelerating pace of development has heightened urgency and broader implications more than ever. In response to these urgent needs, an approach termed Human-Centered Artificial Intelligence (HCAI) has received increasing attention. Similar to the Human-Centered Systems approach described by Cooley (2000), or Human-Centered Design (HCD) (IDEO.org, 2023; Roumani & Both, 2020), HCAI rejects the mechanistic paradigm of technological and societal development. Instead, it advocates for placing human beings and their human experiences at the center of the design, development, and implementation of AI systems. The goal of HCAI is creating AI systems that are not only more inclusive and trustworthy but fundamentally aligned with human values and goals. And while the idea of creating technology for human benefit is not new, HCAI represents a contemporary effort to apply Human-Centered Design concepts specifically to AI systems. HCAI is built upon several interconnected foundational elements with Human-Centered Design as the conceptual and methodological bedrock. 2.2.1. Human-Centered Design. The idea of applying human-centered approaches in technology development has evolved significantly over the past decades. While initially a general term encompassing various research themes within the field of Human-Computer Interaction (HCI), it has emerged as a distinct field of inquiry, manifesting in areas such as human-centered computing (HCC), design (HCD), and systems (HCS). As the name suggests, the core of HCD is understanding the needs and perspectives of the people who are supposed to use a product, service or system. Cooley (2000) describes Human-Centered Systems as the appropriate context to leverage linguistic, cultural, and geographical diversity as a source of information. IDEO, a design and consultancy company that is often attributed as one of the organizations coining the term, describes HCD as being about cultivating deep empathy with the people one is designing for. It also implies a design process, from generating ideas, building different prototypes, testing and evaluating these co-design results to eventually putting a new solution out in the world (IDEO.org, 2023). The Stanford d.school understands HCD similarly, describing it as a process, a mindset, and an approach to identify meaningful challenges and creatively solve complex problems. HCD as an approach guides understanding and responding to the needs of specific people, questioning assumptions and reframing problems, and experimenting to advance their solutions. The term Human-Centered Design is often used synonymously with the idea of user-centered design. However, the human-centered paradigm goes beyond simply considering the user to fundamentally prioritizing human concerns, skills, and activities in the design process, emphasizing that it impacts several stakeholders, not just a single group typically considered as users. As argued by Gasson (2003) user-centered system development methods fail to promote human interests because of a goal-directed focus on the closure of predetermined, technical problems. Some scholars see human-centered design as an evolution from user-centered design. The multidisciplinary nature of HCD is evident in its historical roots, which span ergonomics, psychology, computer science, anthropology, engineering, and arts. This diverse foundation is reflected in the ISO 9241-210:2019 standard, which defines HCD as an approach aimed at enhancing the usability of interactive systems through the application of human factors, ergonomics, and usability principles. Each of these descriptions highlights the importance of putting humans at the center of the design process. They underscore that the understanding of needs and perspectives of affected stakeholders is essential for the creation, effective and meaningful solutions to promote individual and societal value. In addition, yields the adoption of HCD in system, service, and product development yields significant economic and social benefits. These include increased user productivity, improved organizational efficiency, reduced training and support costs, enhanced accessibility and user experience, minimized discomfort and stress, and contributions to sustainability and well-being objectives (ISO 9241-210:2019). 2.2.2. The Emergence of Human-Centered AI. Creating AI-driven technologies that serve human needs requires combining AI-based technologies with a human-centered design approach. The International Standard on Human-Centered Design describes that the extent to which products are usable and accessible depends on an understanding of the context of use (ISO 9241-210:2019; 3.7). The context of use is described as the combination of users, goals, tasks, resources, and environments. Dignum and Dignum (2020) describe that any task of an HCAI system is not done in isolation, but should be done for someone, in some context. This comprises multiple sensible aspects highlighting the importance of acknowledging the stakeholder-plurality for a system as well the technical, physical, social, cultural, and organizational circumstances (ISO 9241-210:2019; 3.2). Stakeholders are considered individuals or organizations with a right, share, claim, or interest in a system or in its possession of characteristics that meet the individual or collective needs, goals, and duties (ISO 9241-210: 2019; 3.11). As part of the overarching goal for a human-centered AI future, the High-Level Expert Group on AI, a group of experts appointed by the European Commission to provide advice on its AI strategy, raises the general question of necessity and beneficial intents. Furthermore, the group mentions that the purpose of an HCAI system should intend to improve human welfare and freedom, by identifying and understanding existing needs and values of different stakeholders. They are a major sources of information for establishing requirements and an essential input to the HCD process. Needs can be individual, considering the immediate users of a system, as well as collective. The development of HCAI systems entails the identification of social values, deciding on a moral deliberation approach, and linking values to formal system requirements and concrete functionalities. As articulated by the Stanford Institute for Human-Centered Artificial Intelligence, HCAI explicitly addresses societal needs. This perspective enables a nuanced evaluation of both direct and indirect impacts of AI systems on individuals and communities. The HAI framework by Xu (2019) introduces the notion of ethically aligned design, which emphasizes the development of AI solutions that endorses principles of non-discrimination, fairness, and justice. Similarly, Holzinger et al. (2022) define HCAI as a synergistic approach to align AI solutions with human values and ethical principles. Further dimensions within HCAI that require close attention are the interactions among humans and other elements of a system. This is encapsulated under the umbrella of human factors and ergonomics which focus on the effectiveness, efficiency, and satisfaction of use within the nexus of users, goals, and context of use (ISO 9241-210:2019; 3.13). Different Properties can contribute to create a usable HCAI system. In the aforementioned HAI framework, human factors design ensures that AI solutions are explainable, comprehensible, useful, and usable. Accordingly, are the purpose, the values and the properties of a product, service, or system crucial aspects of Human-Centered AI systems. 2.2.3. Purposes, Values and Properties. HCAI systems are fundamentally socio-technical systems, having the social context implied in which they are developed, used, and acted upon (Dignum & Dignum, 2020). It is fundamental to recognize that the technical component cannot be separated from the socio-technical system when considering the effects and the governance of AI technology, or the artifact that embeds that technology. A similar understanding is put forward by Riedl (2019), stating that HCAI systems must be designed with the awareness that they are part of a larger system consisting of human stakeholders, such as users, operators, clients, and other people in proximity, fulfilling a purpose within its context. Considering the Purpose of a product, service, or system is therefore a crucial element of Human-Centered AI systems and is therefore considered as the first element of an HCAI system (Table 1). The first type of purpose within the HCAI literature has been found in Augmentation. AI systems are often divided into intelligence imitation and intelligence amplification. Within public media depictions, many times intelligence imitation is the fuel for dystopian narratives, whereas intelligence amplification is usually portrayed in a more practical, humane light. Where intelligence amplification carries the notion of human augmentation and enhancement of human capabilities, intelligence imitation considers the ideas of AI autonomy and automation. Intelligence augmentation is an idea originating in the works by Douglas Engelbart at the Stanford Research Institute as well as the Defense Advanced Research Projects Agency (DARPA) in the early 1950s and 1960s. Instead of viewing AI as a standalone tool, the idea of symbiotic cooperation between humans and machines is a fundamental belief of HCAI with the shared precept of augmentation instead of replacement. This is based on the understanding that technology is created with the purpose of supporting humans, not rendering them redundant. Another type of purpose that has been identified in the Human-Centered AI literature is AI-Autonomy. It describes the capability of an AI system to make decisions and take actions without direct human intervention or control. In order to be fully autonomous, a system must have the capability to independently compose and select among different courses of action to accomplish goals based on its knowledge and understanding of the world, itself and the situation. Some features and systems can be made autonomous if they require rapid action (e.g. airbag deployment, anti-lock brakes, pacemakers, implantable defibrillators; Shneiderman, 2020c). In these applications, there is no time for human intervention or control. Because the price of failure is so high, these applications require extremely careful design, extensive testing, and monitoring during usage at scale to refine designs. A third possible purpose of an HCAI system identified within the HCAI literature is Automation. Automation is not an AI-specific purpose since it can be achieved also without AI. However, the novel possibilities unlocked by AI and especially, machine learning technologies, enable new forms of automation. According to Xu et al. (2023), automation refers to the outsourcing of repetitive, mundane, or time-consuming tasks that can be performed by AI systems more efficiently. The demarcation between Autonomy and Automation is rather blurry, as an autonomous system can also perform automated tasks. A useful distinction is provided by Shneiderman (2022), who states that systems are automated, but not autonomous, when they are governed by prescriptive rules that permit no deviations. Xu et al. (2023) align with such differentiation by describing Automation as being deterministic whereas Autonomous behavior as non-deterministic. The second element of HCAI identified in the literature is Values (Table 2). Values refer to fundamental beliefs and ideals that shape individuals' and societies' judgment and behavior. Values serve as a guiding framework for individuals and groups, to define what is important, just, and meaningful. Developing HCAI systems entails processes such as identifying social values, deciding on a moral deliberation approach, and linking values to formal system requirements and concrete functionalities. A prominent example of how to work with and for values can be found in the Value Sensitive Design approach by Friedman et al. (2013). Similar to Value-Sensitive Design, Renz and Vladova (2021) introduce Design for Values, which is based on a methodology aimed at incorporating moral values as part of technological design, research, and development. Within the analysis of the HCAI literature, three sub-categories of values emerged, namely Ethical Values, Protection Values and Performance Values. The terms Ethical AI or AI Ethics have become common terms used not only within the realm of HCAI but also much more broadly in the debate on AI technology. The idea of AI ethics has emerged as a response to the abundance of societal and individual perils caused by the abuse, misuse, poor design, or unintended consequences of AI systems and is closely related to concepts such as dignity, fairness, and justice for individuals and society. By incorporating ethical considerations into the design and implementation of AI systems, Human-Centered AI aims to safeguard against potential negative consequences and ensure that AI technology operates in a manner aligned with human values and societal well-being. Another type of values within the HCAI literature is the group of values that carry a protective meaning. Protection values such as privacy, safety, and prevention of harm concern the observation of fundamental human rights, the safeguarding of physical and mental health, as well as environmental protection (The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, 2019; United Nations General Assembly, 2015). In the context of HCAI, they play a pivotal role in promoting and warranting trust, societal acceptance, and a beneficial impact of an AI system. Finally, one of the most obvious and remarkable strengths of AI technology is the improvement of performance. Performance values contribute to HCAI systems by maximizing performance benefits for individuals and societies separately but also in interdependence as they are also interconnected and mutually reinforcing. The third element of HCAI systems relates to ensuring specific AI properties (Table 3). Some of these properties concern oversight aspects, addressing questions around accountability and responsibility. Other identified properties concern comprehension, relating to making AI interpretable and understandable to users, stakeholders, and regulatory bodies. Furthermore, the property of controllability is considered key and is linked to meaningful human control over AI systems. Transparency and traceability are also pivotal and concern disclosure of data usage, algorithmic logic, and decision-making to build trust and minimize the risk of hidden biases or unfair practices. Sustainability emerges as another crucial aspect, emphasizing the farsightedness of HCAI systems concerning environmental, social, and economic sustainability goals to ensure the longevity of HCAI applications. And finally, the properties of reliability and robustness relate to the consistent and dependable performance of HCAI systems under diverse conditions to instill user confidence and facilitate safe AI deployment. Overall, each of the identified properties contributes to the overarching objective of human and societal well-being. 2.2.4. Human-Centered Design practices, processes, and tools. Finally, practices, processes, and tools coming from a Human-Centered Design approach are building the backbone of human-centered design. They entail continuous and active involvement of users throughout the design and development (ISO 9241-210:2019; 5.3) to maintain a clear understanding of user and task requirements (ISO 9241-210:2019; 5.2). The ISO standard states that the involvement of users provides a valuable source of information, and that the effectiveness of involvement increases as the level of interaction between developers and users increases. According to Auernhammer (2020), the participation of stakeholders in the design and implementation of HCAI systems can provide essential perspectives in the re-design of social organization arrangements such as work systems. Including many diverse groups throughout the design and development process allows the design of the AI system to be accessible and meaningful for a wider group of people. User-centered evaluation (ISO 9241- 210:2019; 5.4) and iterative processes are essential, as many of the needs and expectations of users and other stakeholders that will impact the design of the interaction only emerge in the course of development, as the designers refine their understanding of the users and their tasks, and as users are better able to express their needs in response to potential solutions (ISO 9241-210:2019; 5.5). A common perspective on HCAI processes, methods, and tools in the literature is the relevance of the HCI discipline. The authors highlight specifically the intersection of AI and HCI and how HCI approaches and methods can be applied to ensure Human-Centered AI development. A similar proposal has been also brought forward by Harper (2019), claiming that “the future is HCI, not AI”. Capel and Brereton (2023) found that the design of HCAI systems has been approached through various methods, including participatory design, focus groups, interviews, usability studies, and observations. There exists already a myriad of methods and tools for the design of HCAI systems. Some of them are existing methods being transferred to the AI context as well as new processes and methods being specifically developed. Ozmen Garibay et al. (2023) ask whether existing HCI design methods and processes scale up to accommodate a wide variety of users’ characteristics and contexts of use in order to create AI-enabled systems that are universally accessible and universally usable? In their work on new challenges and opportunities for HCI professionals to enable HCAI, Xu et al. (2023) highlight that the context of use of single user-artifact interaction needs to be expanded to study the impacts of complex ecosystems of artifacts, services, and data in a distributed context of use. Continuous user involvement constitutes the procedural bedrock of HCAI. In a research panel discussing the role of HCI in the conception, design, and implementation of HCAI, Schmidt (2020) state that HCAI emphasizes the human side of the interaction between people and AI and focuses on creating a positive impact on human users and society, rather than highlighting algorithmic performance. In section 5.2, the ISO 9241-210:2019 standard defines that a human-centered design needs to be based upon an explicit understanding of users, tasks, and environments, i.e., the context of use. Products, systems, and services should be designed to take account of the people who will use them as well as other stakeholder groups, including those who will be directly or indirectly affected. For Dignum and Dignum (2020), the idea of a human-centered system design is manifested in having humans always in focus for deliberation. This means that an HCAI system should not be understood in isolation, but that it acts for someone, in a specific context of use. To grasp the vast complexity of an algorithmic system, the continuous inclusion of the people who are and will be affected by the envisioned HCAI system is crucial. Furthermore, evaluation is key to assess whether a proposed HCAI system addresses identified needs, adheres to elicited values, and provides the necessary properties. Evaluating HCAI systems against real-world scenarios with stakeholders and improving them based on feedback provides an effective means of minimizing the risk of a system not meeting stakeholder needs 2.3. Design Principles. To operationalize specific recommendations about the fundamental nature and structure of a socio-technological system, Design Principles have been established as a suitable tool across various disciplines such as engineering, economics, and information systems. Design Principles are a type of design knowledge, infused by a theoretical lens, often referred to as a Kernel Theory. Walls et al. (2004) referred to kernel theories as the bases of design theories coming from natural and social science theories. The authors state that the conditions of the natural and social worlds are defining the rules boundary conditions of creating information systems. Design principles can guide the implementation of AI systems ensuring that systems are developed with a clear understanding of their social implications, equipped to handle complex ethical dilemmas, and capable of delivering balanced benefits to both individuals and society as a whole. They act as intermediaries, bridging the gap between theoretical knowledge and its practical application. As AI technologies continue to evolve and become more integrated into public sector operations as well, design principles can contribute to advancements that are innovative, inclusive, and beneficial to society. 2.3.1. Design Principles in IS research. Design principles serve as a helpful tool for gain insights from past successful designs and applying them across various contexts to guide future design efforts. By identifying cross-case patterns, they contribute to the development of a broader theory for design and action. Codifying these principles is particularly significant as it allows for the transfer of design knowledge, making it accessible and reusable for others. This reuse minimizes the trial-and-error process, thereby accelerating the creation of effective designs. Specifically, we propose that design principles can facilitate the standardization of boundary resources in e-commerce ecosystems, lowering entry barriers for developers and fostering network effects. 2.3.2. Design Principles in Practice. Within the sphere of design practice, several approaches exist to catalyze the generation of creative, human-centered solutions. The most common can be found under the name of User-Centered Design (UCD). UCD describes a similar concept to Human-Centered Design yet possesses a subtly different understanding of people-centricity in technology creation. As described earlier, this doctoral thesis opts for the term Human-Centered Design as it captures the nature of human involvement beyond the prescribed role of being a user. Other approaches that aim to ensure affected humans are part of the design process can be found in participatory design or empathic design). The focus of these approaches is to ensure involvement, participation and empowerment of people who are not trained in design to inform, ideate, and conceptualize products and technology together with professional designers. A widely used set of design principles can be found in Nielsen’s Usability Heuristics (1994). These heuristics are a set of 10 general principles for designing user-friendly interfaces, including guidelines on system visibility, clear feedback, the prevention of errors, and maintaining consistency to help users achieve their goals efficiently. These heuristics serve as a foundation for evaluating and improving the usability of digital products. Compelled by the rapid spread of AI technologies, there are already numerous AI design principles, aiming to bridge theoretical concepts with development and deployment practices, focusing on enhancing user-technology relationship through various recommendations and best practices. Leading tech companies like Microsoft, Apple, and Google have actively contributed to shaping a more ethical and human-centric AI development approach. In their comparative analysis between different existing HCAI guidelines and recommendations, Wright et al. (2020) found Google's PAIR guidelines to be the notably most comprehensive, addressing various socio-technical aspects more holistically compared to others. Google's PAIR guidelines were praised for fostering trust, interpretability, and transparency in AI. Despite their efforts, these guidelines predominantly cater to commercial interests, and the lack of domain-specific guidelines for understanding AI processes in distinct contexts remains. This underscores the need for a public sector-specific approach due to its unique relational and dependency dynamics. 2.4. Summary of the Chapter. In this chapter, an overview about the foundational concepts and the current research landscape related to AI in public service, Human-Centered AI, and Design Principles is provided. Specifically, the chapter explores the state of AI adoption in the public sector, highlighting its potential, associated challenges, and concerns. The chapter also introduces Human-Centered AI as a framework for responsible AI implementation, detailing its origins and key components. Finally, Design Principles as a practical approach for operationalizing Responsible AI are proposed. 3. Theoretical Lens – Social Contract Theory. Chapter 3 introduces Social Contract Theory (SCT) as the theoretical lens for this study. It outlines the nature of SCT as a diverse body of conceptual streams, rather than a single theory, explaining why it offers a suitable lens for analyzing responsible AI in the public sector. By focusing on the relationships and power dynamics among involved actors, SCT provides a robust foundation for examining AI in public services. Furthermore, the chapter examines existing uses of SCT within Information Systems research, as well as proposes a three-level framework, derived from SCT, to guide the analysis in this research. 3.1. The Concept of the Social Contract. The term Social Contract Theory encompasses a large body of theories and should be understood as a collection of conceptual streams rather than a single specific theory. According to D'agostino (1996), a Social Contract is concerned with determining whether a ruling power is legitimate and therefore worthy of loyalty. In general, the Social Contract describes a hypothetical agreement of moral and political rules of behavior in which the members of a society endorse and comply with these fundamental rules, institutions, and principles of that society. It delineates a structure of power relationships between governments, institutions, and the people (Figure 3). These power relationships assign legitimacy to governments and institutions stemming from the consent of the governed (Rousseau, 1762). It is an agreement made between citizens and the state, through which they accept the authority of the state in return for benefits that only a sovereign power can provide. In consequence, this engenders a responsibility for the rulers to act in the interest of the ruled but also ascribes certain rights and obligations to the latter. It entails the necessity for governments to uphold the general will, ensuring public services serve the collective interests and promote equity among citizens. It formalizes the effort to create consensus on shared values and ideals and ensure ethical practice. Gauthier (1986) argues that any system of moral constraints must be justifiable to those subject to it, emphasizing the importance of mutual acceptance. A Social Contract approach seeks to establish that social, moral, or political rules can be rationally defended and accepted by all members of society. As Rawls (2001) explains, the Social Contract serves as a model of rational justification by reframing the challenge of justification as a deliberative process. This framework, with its general parameters and dynamics, provides a basis for understanding the reasons individuals may have for endorsing and adhering to shared social principles, norms, and rules. Yet justification does not rely on exogenous reason or truth but is generated endogenously by rational agreement, namely consensus. In Rousseau’s understanding consensus is built on complex social processes and is not merely an intellectual task but requires social, political, and organizational transformation. 3.2. Social Contract Theory in Information Systems Research. Having its origin in political philosophy, SCT has been less discussed within the Information Systems field compared to other theories. However, although less common, SCT has served as a theoretical vantage point in several studies. Smith and Hasnas (1999) examine SCT as a potential guiding theory to form normative business ethics. Describing that when following a SCT approach, the involved actors need to be aware of their contractual roles. Managers should aim to increase social welfare above what it would be in the absence of the existence of corporations, yet without violating basic canons of justice. And in turn, society will be willing to authorize corporate existence only if corporations agree to remain within the bounds of justice. In an attempt to create an IS-compatible definition, Grazioli (1998) defines a Social Contract as an interaction between two parties characterized by a conflict of interest: one party is obligated to satisfy a requirement of some kind, usually at some cost, in order to receive a benefit from the second party. In their work on developing a framework for the formulation and governance of AI regulation, Almeida et al. (2020), state that frameworks taking an SCT perspective, rank among the most open to the participation of society in a co-production with the government. Because a moral responsibility is distributed among designers, regulators, and users, they took inspiration from SCT to adjust the man-in-loop or human-in-the-loop model to a society-in-loop model, which they claim would benefit addressing conflicts among safety-, privacy-, and justice-related tensions. A similar proposal has been put forward by Rahwan (2018), who describes the idea of society-in-the-loop as a consolidation of the human-in-the-loop mechanism and the fundamentals of a Social Contract. Cardon et al. (2023) used a Social Contract perspective to examine potential issues about consent, data use, and privacy in the context of work meeting recordings. Following an SCT perspective, privacy is a contextual—and relationship-dependent norm that is negotiated as a microsocial and macrosocial contract within a particular community or situation. Malhotra et al. (2004) used SCT as a theoretical framework to investigate online consumer privacy concerns. Vial (2019) suggested a turn toward normative theories of ethics, including SCT to research the challenges in the context of technology design and use by multiple parties. Donaldson and Dunfee (1994) used the tenets of SCT to develop their Integrative Social Contract Theory (ISCT) which outlines that shared norms in an industry act as a foundation for organizations to behave ethically. By that, ISCT molds SCT from a philosophical level into a more applicable construct that can be applied within different domains. According to Xu et al. (2006), ISCT is the most widely used ethical theory in the context of information privacy and has been used to strengthen the bond of trust between corporations and consumers. They state that ISCT posits that members of a given community or industry behave fairly if their practices are governed by Social Contracts. According to this ISCT perspective, a Social Contract is held to occur when consumers provide personal information to certain corporations, and the corporations in turn offer some benefits to the consumer. For the corporations, one generally understood obligation accruing from entering this Social Contract is that they will undertake the responsibility to manage consumers’ personal information properly. Thus, the social contract, dictating how corporations handle consumers’ personal information in an implicit form, not in an economic or a legal form, involves unspecified obligations and requires consumers’ trust in the corporation’s compliance with this social contract. Perreault (2015) has used ISCT as one of the foundational elements for the hypothesis formulation around control and awareness concerning Big Data and Privacy. She hypothesizes that consumers may perceive the collection and use of personal data as fair as long as they are made aware of the intended use as well as granted control over the data. Further, if the personal data collection and use has been granted, there is the expectation that it will be used for its intended purposes. Perreault states that this expectation of the use of personal data by organizations can be seen as an implied social contract. 3.3. A Social Contract Theory Perspective on AI in Public Services. Public service organizations are the executive branch of governmental administrations, fulfilling governmental rights and obligations to provide services to the citizens within its jurisdiction. The purpose of public service organizations is to mediate the relationships between government organizations and citizens and make positive contributions to society, by providing their services to citizens as well as instruments for implementing public policies. These dynamics can be understood as interrelations between citizens and public service organizations, which renders SCT a suitable lens for this research into the responsible design, development, and deployment of AI and machine learning technologies in the public service domain. By law and definition, public service organizations are obliged to provide the same services to the same extent and in the same quality to all citizens regardless of social, ethnic, or religious background. They cannot choose whether they want to offer or deny a specific service to a citizen. At the same time, neither the public service organization nor the citizen can pick and choose with whom they want to engage with. This de-facto monopoly leaves citizens bound to this one public organization, without any option to choose or change between different service providers. Such mutual dependencies, boundedness and power relations create a special relationship between governmental institutions and the people. An equal treatment of citizens through ensured neutrality and impartiality is therefore a critical public value. From the perspective of SCT, power structures between citizens, governments and governmental organizations are derived from an implicit or explicit agreement among individuals to form a society and accept certain obligations and limitations on their freedoms for the sake of order, protection, and the common good. This theoretical approach has been elaborated by philosophers such as Thomas Hobbes (1651), John Locke (1689), Jean-Jacques Rousseau (1762), and more recently, John. In a more contemporary perspective, Weale (2020) argues that for a government to be legitimate, it must operate within the confines of justice as determined by agreements or principles that citizens can reasonably accept. This includes the recognition that power must be exercised in a manner that respects the rights and interests of citizens, ensures their participation in the political process, and distributes benefits and burdens in a fair manner. Based on these considerations of power structures, inter-actor dynamics and various negotiation practices, the SCT approach can be conceptualized in a three-level analytical framework consisting of a Macro, a Meso and a Micro level (Figure 4). 3.3.1. Macro Level. Citizens expect that their data will be used by the authorized organization for its intended purposes and to their benefit, which can be seen as an implied Social Contract. This points toward an implicit expectation from the individual to be able to exercise their right as the sovereign to question governmental decisions. Furthermore, this also examines how a social contract, which is not a formalized contract but rather a conceptual one, can be broken. For example, Xu et al. (2006) found that a Social Contract in the case of technology use in the public sector is considered breached if citizens are unaware of unauthorized data collection or data processing, or data transfer to other parties without their explicit consent, which illustrates a form of power abuse. This level of examination can be understood as the Macro level of SCT. 3.3.2. Meso Level. At the next level, (the Meso level), the interaction between the individual and society is considered. This relationship has been touched by almost all political debates. Applied within the context of socially responsible AI, it can help to better investigate the balancing act between individual rights and benefits, and the common good. Sticking with the example of personal data, what incentive do individuals have to share personal data, to provide benefit for everyone within a society? Why are they prioritizing either societal cooperation or personal gain? The reasons individuals have for agreeing to some rules or principles need to be understood as their own reasons, not necessarily as good reasons from the impartiality perspective. Individuals may care about what they perceive to be impartially good or some other non-individualistic notion, but what they care about will differ from one another. Rawls (1996) highlights that a society cannot reasonably be expected to have similar conceptions of the good, and heterogeneity needs to be considered. And the same is true for the perspective of the government organization. Does the organization prioritize individual freedom, or does it exercise its power for the benefit of the collective? If a Social Contract is to endorse interrelated normative desiderata (e.g., liberty, equality, and welfare, which are all guiding principles for a working society), a deliberative process that draws on a diversity of perspectives will outperform one based on a strict normalization of perspectives. This aligns with a human-centered approach that acknowledges and incorporates the tensions between the individual benefit and the collective good. 3.3.3. Micro Level. The final tier of examination is found in the Micro level. This level explores the intricate processes and mechanisms that underpin the attainment of consensus among the various involved stakeholders. It places particular emphasis on the dynamics and interactions between citizens and public organizations, scrutinizing the immediate communication and exchanges among these different actors to construct an agreement. Among the most widely discussed mechanisms are consent (Hobbes, 1651; Locke, 1689; Rousseau, 1762), bargaining, aggregation, and equilibrium. These mechanisms delineate various forms of normative authority for self-binding, predicated on the notion that individuals possess fundamental normative powers over themselves. If parties can indeed bind themselves by wielding this normative power, then the outcome of the Social Contract results in an obligation (Hobbes, 1651; Hume, 1963). It is worth noting that some of these agreement mechanisms have faced scrutiny from contemporary Social Contract theorists, with consent, in particular, standing out as a key example. 3.4. Summary of the Chapter. Within this chapter, Social Contract Theory was presented as the theoretical lens for this study, as it provides a comprehensive perspective on the relationships and power dynamics among the different stakeholders. Additionally, the chapter presented a three-level analytical framework based on SCT foundation for empirical data analysis. 4. Research Context. This chapter outlines the research context of the doctoral study. It provides an introduction about the Norwegian Labour and Welfare Administration (NAV), which has been the collaboration partner within this Action Design Research. Further, the chapter elaborates on sick-leave handling, which has been the specific empirical context. It offers a brief description of NAV’s current sick leave assessment procedures and outlines an AI-enabled sick leave prediction tool, which has been developed by NAV for testing purposes. Additionally, to provide more contextual information, the chapter presents the Norwegian Guidelines on AI Transparency and elaborates on the role of national culture which is reflected in these guidelines. 4.1. The Norwegian Labour and Welfare Administration (NAV). The Norwegian government promotes the use of AI in public administration, aiming to pioneer in the development of human-friendly and trustworthy solutions (Norwegian Ministry of Local Government and Modernisation, 2020). The national Norwegian AI strategy states that in the future, the public sector will use AI to provide more targeted and user-adapted services, enhancing the societal value of its activities, streamline operations and work processes, and mitigate risks (Norwegian Ministry of Local Government and Modernisation, 2020). The goal is to have 80% of government agencies to have adopted AI in 2025, and 100 per cent by 2030 (Norwegian Ministry of Digitalisation and Public Governance, 2024) A central role within this strategic undertaking is played by the Norwegian Labour and Welfare Administration (NAV). NAV is the largest public service organization in Norway, managing approximately one-third of the national budget by providing services such as unemployment benefits, sickness benefits, pensions, child benefits, and cash-for-care benefits. NAV employs around 19,000 staff, with approximately 14,000 working for the central government and 5,000 employed by local authorities. Its extensive network includes 456 municipal and city borough offices, alongside over 100 specialized units handling various back-office operations. NAV’s overarching mission is to deliver services tailored to users’ individual needs and circumstances, support a well-functioning labor market, promote active participation, and ensure an effective and comprehensive labor and welfare administration. To improve their overall service provision, delivering better, more efficient and more robust service, NAV is actively exploring the potential use of AI systems. Already in 2016, the organization established a large data lab to explore the possibilities that new data analytics methods may offer for labor and welfare services in various areas. In 2017, NAV established its own AI lab and hired its first data scientists to secure cutting-edge expertise in the field. However, NAV is also very careful with the deployment of any AI tools as the receivers of welfare services and benefits are often in vulnerable situations and any worsening of these situations should be avoided. As a major public organization, NAV is aware of its societal role and aims for a responsible use of AI within their public service provision. The use must not only be lawful but also arranged in accordance with societal ethical norms in a way that meets citizens with respect and dignity. 4.1.1. Sick Leave Handling. In their endeavor to prospect potential uses of AI systems in their organization, NAV explored multiple different use cases. Examples of ongoing AI initiatives include an AI solution for efficient routing of cases with automated content classification, a payment error detection system, and a prediction model to support sick leave management. The latter project was chosen as the use case for this doctoral research project. Based on the legal requirements accounted for by the Working Environment Act and the Norwegian National Insurance Act (2005), NAV is responsible for following up with employees being on sick leave. The general assumption is that the employee and their employer will discuss and develop a plan to get the person on sick leave to return into the workforce when the health situation allows it. The employer and the employee together must prepare a follow-up plan containing concrete measures to achieve this goal. Furthermore, the Norwegian National Insurance Act also requires the two parties to carry out a first appointment, a so-called dialogue meeting, to discuss the pursuit of the agreed follow-up plan. For the first appointment, NAV can be involved on request by the parties if there is a need for external assistance, but there is no legal obligation for NAV to be included during this first period. However, if the person on sick leave is not able to return to work within 16 weeks, it is the duty of NAV to request a second dialogue meeting. In this second meeting, the employee on sick leave, the employer, and the healthcare professional that notified about the sick leave, need to take part. This second meeting must be conducted no later than within 26 weeks of the beginning of the sick leave. This second dialogue meeting has been introduced to strengthen the follow-up of long-term sick leavers, and to increase their chances to return to work. To plan the second dialogue meeting, a NAV caseworker sends messages to the person on sick leave, the employer and the healthcare professional during week 17 and schedules the meeting for week 26. If any of the parties requests the dialogue meeting to take place earlier than week 26, the caseworker needs to act accordingly. Another option is for the NAV caseworker to decide that a dialogue meeting is obviously unnecessary and not to schedule it at all (Figure 5). For instance, a reason for skipping the meeting would be if people on sick leave are in the hospital for prolonged in-patient stays and therefore will not be able to return to work regardless of the dialogue meeting. Also, an expected timely recovery and a return to work by week 28 is a potential reason to go without a dialogue meeting despite the normally legal obligation. 4.1.2. AI-enabled Sick Leave Duration Prediction Tool. The decision to waive such meetings requires careful consideration of previous medical information, the physician’s assessment, and the prepared progression and treatment plan. NAV identified this decision-making process as a potential use case for leveraging AI technology for decision support. Specifically, an AI-enabled sick leave duration prediction tool to assist caseworkers in assessing the necessity of dialogue meetings was envisioned. The rationale behind developing this tool stems from the recognition that many dialogue meetings are obviously unnecessary while consuming significant time and resources for all parties involved. The concept for AI use is to have a tool which can provide a prediction of the sick leave duration based all available information. This can be used by the caseworkers to assess the necessity of a dialogue meeting (Figure 6). By enabling caseworkers to make more informed and efficient decisions, the tool aims to free up resources and alleviate the organizational burden of scheduling, attending, and evaluating these meetings, for cases where they are not needed. The introduction of the sick leave duration prediction tool is expected to yield multiple benefits for citizens, caseworkers and the overall organization. Citizens can expect a better follow-up for sick leaves who really need a dialogue meeting due to more focused efforts, as well as saving time by avoiding unnecessary dialogue meetings. Caseworkers are envisioned to reduce their time spent assessing the need for a dialogue meeting and being able to focus more on relevant cases. On an organizational level, the tool is supposed to increase efficiency, following a recurring theme in political demands that NAV's services must be more efficient and designed for user-adapted follow-up. At the time of writing of this doctoral thesis, the AI-enabled sick leave duration prediction tool has been developed as a sandbox case but has not yet been deployed. However, NAV plans to use the experiences from this particular project to shape internal AI practices, clarify legal bases for developing and deploying ML models and to develop insights about the provision of transparency for different users to cater for their needs. 4.2. The Notion of AI Transparency in Norway. As described by Robinson (2020), national strategic policies for AI strengthen notions of openness and trust while fostering sociotechnical understanding and resulting in a deeper comprehension of AI transparency. The concept of transparency has been frequently debated in the fields of AI and Human-Computer Interaction. With several contrasting viewpoints about AI transparency in public services. Several authors have contended that the public's lack of faith in public services is mainly because they are not provided with real information regarding the operations and efficacy of public services. Proponents of transparency, on the other hand, stress that transparency fosters understanding, a tighter bond between citizens, and a better understanding of public services. Though several cultures share the same values of transparency, trust, openness, accountability, and fairness, they do not prioritize those values equally. These variations may cause AI policies from various nations to enshrine the same values in different ways. Any research in the public sector context should consider sociocultural, moral, and economic differences between countries and regions (van Berkel et al., 2020). This is especially valid for research informing National AI policies and when employing a human-centered approach. Within the Norwegian National Strategy for AI, Transparency is explicitly highlighted as a key value in AI development. Even the strategy paper itself sets an example of transparency. Compared to other national policy documents that provide merely truncated technical overviews, the Norwegian policy guide provides detailed explanations of AI technologies, providing basic AI literacy. 4.3. Summary of the Chapter. This chapter provided an overview of the research context. It introduced NAV as the collaborating organization and the rationale for a sick leave duration prediction tool. The chapter also highlights the influence of national culture on AI transparency in public services. 5. Research Design. The following chapter outlines the research design employed in this thesis. First, the philosophical stance underpinning the study is stated. Next, the methodological approach of Action Design Research is introduced, including detailed descriptions of its key stages: Problem Formulation, Building, Intervention and Evaluation, Reflection and Learning, and Formalization of Learning. The chapter then provides an in-depth account of the research process followed and concludes by detailing the process of formulating the defined Design Principles. 5.1. Positioning the Research Approach. This research project follows an interpretivist research philosophy and is supported by an inductive reasoning research approach. An interpretivist philosophy emphasizes the meaningfulness of people's participation in social and cultural life. Interpretive research assumes that people create and associate their own subjective and intersubjective meanings as they interact with the world around them, through social constructions such a language, consciousness, shared meanings, documents, tools, and other artifacts. It attempts to understand phenomena through the meanings that participants assign to them. Due to the exploratory character of this research project, and its guiding research questions, an inductive approach was followed, as meaning was generated from the data collected in the field. With an inductive approach, research starts with a set of observations moving from particular experiences to a more general set of propositions about those experiences. The Action Design Research approach followed allowed the development of insights via real world empirical experiences. 5.2. Action Design Research. Employing an Action Design Research (ADR) approach provides the unique opportunity to facilitate continuous engagement with involved stakeholders in a real-world setting. In general, Design Science Research (DSR) aims to address problems through the creation of solutions that serve human purposes. These solutions emerge via the creation and evaluation of artifacts, which can be understood as any designed object with an embedded solution to an understood research problem, such as constructs, models, methods, or instantiations. However, it is noted that within the field of Information Systems, Design Science Research often focuses on technological rigor but may overlook organizational relevance and the impact organizational context has on the design of an IT artifact. Unlike other Design Science research approaches, Action Design Research (ADR) views the research process as an inseparable integration of iterative cycles of building, intervention, and evaluation (Figure 7). This approach emphasizes that the ensemble artifact is continuously shaped by the organizational context during its development and use. By this, design is understood as a context-specific, multi-stage process within an organization as well as a reflective endeavor that yields design principles for solving specific problems. Table 4 provides an overview about the iterative stages of the ADR process, detailing how the ADR principles were implemented throughout the project, conducted activities, actors, data collection and the artifact emergence. In this research project, the public service conditions served as a rich environment for producing generalizable knowledge about Human-Centered AI design. 5.2.1. Problem Formulation. In the canonical ADR process by Sein et al. (2011), the first stage of the ADR methodology is described as Problem Formulation, in which the fundamental organizational problem acts as the catalyst for the research undertaking. The problem formulation is grounded in practical concerns which are identified together with the collaborating organization, while also being ingrained by theory driven by the research context. Furthermore, the problem formulation is also understood as the single entry point for every ADR project, as every action research should emphasize the notion of a guided emergence. However, the reality of conducting research in the sphere of practice sometimes necessitates adaptations deviating from theoretical ideals, with this ADR project deviating from the orthodox approach to adapt to structural circumstances as well as organizational changes. The first deviation within the ADR approach of this doctoral thesis is of structural nature, caused by the offset timelines between the overarching research project, and the doctoral research project. This specific doctoral research project started in December 2021 but is part of a larger, encompassing research project called AI4Users, which runs from 2020 until 2025. The AI4Users project adopts a human-centered approach to AI in public services. The project is funded by the Norwegian Research Council (Project number: 311680) and collaborates with the Norwegian Labour and Welfare Administration (NAV) as the partner from practice as well as involving researchers from the University of Oslo (UiO), NTNU in Trondheim, and the University of Agder (UiA). Within the project, critical issues such as intelligibility and accountability in AI systems, the black box problem as well as transparency and human-in-the-loop approaches are explored. An initial problem identification to understand the general research problem of the collaborating organization, as well as a definition of the solution space happened already during these early stages of the overarching AI4Users project, and thus before the start of this particular doctoral research. In these initial iterative reconciliations, the underlying root problem of how to introduce AI systems responsibly within the public sector context, who are the main stakeholders within this context, and which are the envisioned goals of the overarching project, were defined. As a result of this initial coordination, the agreed scope and defined goal to contribute to the responsible use of AI in public sector services then informed the objective of this specific doctoral research. The elaborated process model by Mullarkey and Hevner (2019) provides the conceptual capacity to enter an ADR project through different entry points depending on the maturity of the project or problem domain (Figure 8). Each of the formalized entry points is defined by specific questions to be answered and activities to perform to find answers to these questions. Following the taxonomy of their elaborated ADR research process model, this doctoral research therefore entered the ADR process through an objective-centered entry point, as it aims to explore the design options based on defined project objectives and to generate design knowledge in the solution space. It needs to be acknowledged, that the original authors oppose the notion of multiple entry points into an ADR project, arguing that it could contradict ADR’s foundational premise that all research starts with a Problem Formulation. They argue that such inconsistency risks shifting ADR into a deductive design science research, (re-)introducing predefined stages, and potentially undermining the emergent nature of ADR. However, it should be clarified that entering the overarching project at the objective-centered entry point does not mean that the stage of problem formulation was omitted in any way. Critical engagement with the organizational problem moved toward a more granular problem formulation, to explore the specific research problem within the defined solution space. This problem formulation step is informed by the negotiations and understandings of the general research problem but focuses its efforts on the exploration of design options with the goal to generate feasible design knowledge. Besides being grounded in the practical concerns identified together with the collaborating organization, the defined solution space is also informed by a theoretical perspective that suits and reflects the organizational context. Within this doctoral research, SCT builds the theoretical foundation. The intersection of the practical challenge of designing, developing, and deploying AI systems within the public service context, with the theoretical foundations of SCT creates a fertile ground for innovation, and the creation of ensemble artifacts as well as contributing to wider knowledge in the field. Over the course of the whole doctoral research project, the research team held regular and iterative coordination meetings with the public service organization to facilitate a constant exchange of insights and warrant the project remains rooted in the actual needs of the organization. This close interaction between the project partners fostered the creation and rigorous evaluation of the research strategy as well as the artifacts developed during the ADR process. 5.2.2. Building, Intervention, and Evaluation. In the second stage of the ADR methodology, the initial creation of the artifact is informed by the defined problems and theoretical lenses. The process unfolds through iterative cycles of building, intervention, and evaluation (BIE), which are fundamental for the continuous adaptation and enhancement of the artifact in partnership with the collaborating organization. Each cycle begins with the building phase where an artifact is designed or revised based on the insights gained from the previous iterations and grounded in the theoretical underpinnings established during the problem formulation stage. This stage also defines the focus and scope of innovation. Based on the foundational organizational problem, observed through the lens of SCT, two main stakeholder groups – citizens, as well as public servants – were identified. Induced by this multi-stakeholder condition, the need for two initially separated but complementary research strands were established (Figure 9). Conducting two separate research strands allowed for a stakeholder-specific yet conceptually aligned research design as well as flexibility regarding potentially evolving interests and priorities within the collaborating organization or other unanticipated changes to the project timeline. Despite the initially separated and time-wise shifted research strands, the findings and insights from both stakeholder groups are merged into a holistic view of the different perspectives regarding the use of AI systems in the public sector. In coordination within the ADR team, the initial focus of the research project was put on the citizen perspective from the Norwegian sociocultural context, as questions concerning transparency, personal data usage, and potential ways of public participation were deemed as the most relevant and pressing. For the citizen-focused research strand of the study, the author of this dissertation created an interactive prototype based on the provided example use-case of an AI-enabled sick-leave prediction model, described in the organizational context. The caseworker-focused research strand started later in the project, which allowed for potential ad-hoc adjustments regarding the research design. Within this research strand, a Foresight approach was followed to explore the needs and perceptions of caseworkers. Besides discussions on how AI already is, or can be, involved in daily work, the sessions also explored different kinds of futures of AI use in their workplaces. The innovative artifact was used as a discussion prompt during the explorative discussion about the different future scenarios. The evaluation phase subsequently assesses the artifact informing the next BIE cycle, ensuring that the artifact not only meets the initial design criteria but also evolves in response to the insights identified during its evaluation. This step is crucial as it provides a practical test bed for the artifact’s functionality and its alignment with the organizational needs and user expectations. 5.2.3. Reflection and Learning. During the third stage of the ADR methodology, conscious and deliberate reflection on the problem understanding, the applied theoretical frameworks, and the development of the emerging artifact take place. This is a critical process for assessing the potential generalizations and contributions to knowledge that can arise from the research. The goal of this stage is to transition conceptually from constructing a solution tailored to a specific instance to generalizing this learning to address a broader class of problems. The evaluated artifact of a sick-leave prediction system depicts a representative instance within the broader category of Human-Centered AI Design in Public Services. This perspective is not limited to the singular application but extends to understanding how an HCAI approach can inform and improve public service delivery more generally. The process of broadening the learning involves rigorous analysis of how the artifact was shaped through its use within the organization, including the examination of the perspectives of citizens, the employees of the public service organization, as well as the input of the project stakeholders involved in the iterative refinement of the solution. 5.2.4. Formalization of Learning. In the fourth and final stage of ADR, the focus shifts to the systematization and articulation of the knowledge and insights gained throughout the project. This stage aims to elevate the situated learning derived from the specific case of sick-leave prediction into generalizable insights that can be applied to a broader class of field problems. This stage involves the documentation of achievements and advancements coming from the research findings of the studies with citizens and public service employees, alongside detailing the consequential organizational outcomes. These outcomes are not merely anecdotal but are systematically categorized and defined as design principles. Translating practical insights into broader design principles involves a rigorous methodological approach. It requires critical engagement in a detailed analysis of how the artifact addressed the specific challenges encountered during the intervention stages and reflecting on the scalability and applicability of the solution across different contexts. The formalization of the design principles represents the result of a dynamic collaboration with the public service organization. The findings from both research strands were systematically analyzed through the SCT framework and evaluated and refined in close, iterative coordination and collaboration. 5.3. Citizen-focused Research Strand. Different researchers investigating public sector AI have called for research that examines specifically citizens ́ stances towards AI. To elicit fundamental values that underlie perceptions and perspectives, ensuring public engagement, as well as discourse with members of society, is a guiding principle of Human-Centered AI. 5.3.1. Prototype. Within ADR, innovation comes from two sources. From generating knowledge that can be applied to a class of problems, as well as from the actual design of the innovative artifact used to address a problematic instance in an organizational context. The innovative artifact enables learning through intervention and provides the basis for generalization. In a series of collaborative iterations, an innovative ensemble artifact embodied in an interactive prototype, was developed together with the public organization. The prototype served a dual purpose. One the one hand, the prototype was used as a conceptual vehicle to facilitate the research study, providing a tangible interface to the participants. On the other hand, the prototype also constitutes an innovative artifact by itself. It provided the organization with specific design-suggestions regarding critical aspects such as data transparency, consent practices and declarations of value propositions to address real world problems. Besides being informed by SCT, the development of the prototype was also guided by Google’s PAIR AI design guidelines. The established and comprehensive PAIR design guidelines provided a helpful starting point for the creation of the first iteration of the artifact. Applying an existing set of field-tested AI design guidelines ensured a responsible basis for the artifact, without the need for reinventing the wheel. The different aspects of the practical guidelines, structured into separate chapters, influenced the creation of the artifact on multiple levels. According to the first chapter of Google’s PAIR framework (User Needs + Defining Success), it is crucial to identify areas where AI's capabilities align with user requirements. Ehsan and Riedl (2020) emphasize that for AI systems to be considered trustworthy and ethically sound, explanations must be clear enough to enable users to make informed judgments based on these understandings. Additionally, within the specific public service context, it is also important to consider and convey the rights and responsibilities of all parties involved. The prototype emulates a public service portal, presenting a scenario where users are asked about their willingness to consent to the use of an AI-enabled sick leave duration prediction system. The scenario starts with a notification in the portal, that directs users to an introduction page of the AI prediction tool. On this page, explanations about the government's mandate on mandatory dialogue meetings, the possibility of waiving these meetings, and the citizens' right to dissent from the use of the AI system, are presented. The information is provided in a textual format (Figure 10) and includes a statement regarding the collective as well as the individual benefit derived from consenting to the use of the AI-enabled sick leave duration prediction system. For the individual citizen, the benefit would result in the potential avoidance of unnecessary appointments, and organizations could improve processes, and use their resources more efficiently. This piece of information aims to address considerations of social ethics rather than imposing moral obligations on citizens. Doran et al. (2017) assert that to achieve trustworthiness, and to evaluate the ethical and moral standards inscribed on a machine, explanations should provide insight into the rationale of the AI system and enable users to draw conclusions based on them. In this specific use case, the purpose as well as the anticipated benefits of such a system for the general society, but also for the individual need to be described. As per chapter 3 of the Google PAIR guidelines (Mental Models), visual explanations can also help to set realistic expectations about the role of an AI system, its capabilities, limitations, and the value it offers. To enhance transparency regarding the process, data usage, and the model's functionality, the prototype allowed participants to explore different information elements and descriptions of the AI-enabled sick leave duration prediction tool. One such element is a process diagram, illustrating the role of the AI system within the overall process, and by that emphasizing the case handler's critical involvement (Figure 11). The process diagram places the AI system within the context of the full process flow and delineates areas where the AI system supplements human decision-making. Furthermore, it also depicts the procedure without the usage of the AI-enhanced prediction tool, in case a citizen would not consent to its use. By providing an in-context overview of the process, the chart aims to foster transparency and a deeper understanding and enable citizens to better understand their rights and role in the process. In chapter 2 of the PAIR recommendations, the importance of explaining the usage and benefits of specific data to users is emphasized. The prototype also includes a data table to provide information about the data used by the AI model and the rationale behind its need (Figure 12). The data table consists of an introductory paragraph clarifying that no new data will be collected, but existing data will be reused. The table itself consists of two columns, one for the type of data used and the other for why the data is necessary for the AI system to operate correctly. This transparent approach to data aims to enhance participants' understanding of the system and provide insights into the used personal data. Chapter 4 of the PAIR framework (Explainability + Trust) states that while providing detailed explanations of the model's logic may not always be possible, the reasoning behind predictions can sometimes be made accessible. Following this recommendation, an interactive chart visualizing feature importance through Shapley Values was integrated. This chart highlights the importance of various variables across different age groups, providing a visual representation of the model's logic, enhancing exploratory transparency, and helping citizens to calibrate their trust in the system (Figure 13). According to chapter 5 of the PAIR framework (Feedback + Control), it is essential to convey users' control over their data usage and the decisions they make. At the end of the interaction sequence within the prototype, participants reach a consent decision point where they need to decide whether they would consent to the use of the AI system or not (Figure 14). It also was highlighted, that the right to object to the use of this AI system exists without any negative consequences or need for justification, emphasizing the voluntary nature of their agreement and the option to not provide consent. Chapter 5 of the PAIR guidelines also stresses the importance of feedback mechanisms, which can help refine the AI model’s performance and user experience. While opting out can be considered implicit feedback, this stage also presents an opportunity to request more detailed feedback from citizens. Within the prototype, users were encouraged to express their opinions and objections (Figure 15). This not only facilitates the direct participation of citizens but also upholds their right to object to personal data usage without negative consequences or justification requirements. This approach ensures transparency and upholds the principle that the right to object is an inherent right of citizens within the process. 5.3.2. Citizen Participation. For the citizen-focused research strand, a total of 40 participants participated, distributed over two rounds of data collection à 20 participants. Participants were in between 18 and 65 years old, reflecting the distribution of the general population on sick leave based on the official Norwegian statistics office. The number of participants was defined considering the relevance and diversity of participants’ backgrounds and experiences as well as the qualitative research design, which entails engagement with participants to elicit rich qualitative insights. Another criterion for determining the number of participants is theoretical saturation. Theoretical saturation occurs in data collection when new information ceases to emerge and themes become repetitive. The key point is reaching the point of data saturation, where new participants are not providing substantially new insights. Data collection continued even after initial similarities in responses and to balance depth of insight with practical considerations, although reaching out to new participants stopped after 40 interview sessions with prototype walkthroughs were completed. This number has been shown to be sufficient to obtain meaningful results through user studies. As themes became repetitive, the interviewed 40 participants provided sufficient diversity and depth. The amount of collected data was deemed appropriate since the aim of the study was to gain a deeper understanding of citizens’ perspectives towards the specific context of AI use in public services. The recruitment criteria included gender and education, to match official Norwegian statistics. For the gender criteria, the research team was only able to differentiate between two genders (female and male) as no further identifications were available in the official data. Moreover, it was also aimed to match the statistical distribution of the educational level, including participants with high school education (Videregående), vocational school level (Fagskolenivå), and university education (Universitets- og høgskolenivå). The sessions were conducted both online via video conferencing and screen-sharing applications as well as in person, depending on the availability of the participants. At the beginning of each study session, participants received information regarding the aims of the particular study and the overarching goals of the research project and were also briefed on the voluntary nature of their participation. Additionally, they were provided with an overview of the study’s procedures and were reminded of their right to withdraw from the study at any point. This research study was examined and approved by the Norwegian Center for Research Data (Norsk Senter for Forskningsdata—NSD) under the reference number: #931033. This approval process involved a comprehensive review of the research protocol, the recruitment process, informed consent procedures, and data handling to confirm that the study adheres to ethical guidelines and legal regulations. The approval confirms that the study has undergone a thorough ethical review and met the necessary standards and requirements, ensuring the protection and well-being of the participants throughout the research process. 5.3.3. Empirical Data Collection and Analysis. The data collection included three consecutive stages. In the initial stage, general data about the participants, such as age and gender, current occupation, and highest educational level was collected. Further, participants were asked to provide two self-assessments on a scale from 1 to 5 (low–high). The first dimension was defined as prior knowledge about AI with an average self-reported rating of 2.20. Overall, participants reported a rudimentary knowledge of AI but lacked a deeper technical understanding. The second dimension was their frequency of technology use, with participants providing an average self-reported rating of 4.65, demonstrating exposure and general familiarity with the use of technologies such as computers or smartphones. Next, data about the level of trust toward the Norwegian government and governmental organizations and on the use of AI within public services was collected. These questions addressed the Macro as well as the Meso level of the SCT framework. The data collected through these questions provided indications about the current state of the government-citizen relationship and about participants’ understanding of the current balance between individual and collective benefits from public services. In the second stage, a moderated user study in the form of a task-based interaction with the prototype was conducted. Participants were presented with a short scenario and task that led them to a decision on whether they would consent to the use of a new AI-enabled sick leave duration prediction system. While performing the task, the participants were encouraged to share their thoughts and feelings with the study moderator. Such think-aloud protocol describes the concurrent verbalization of thoughts while performing a task. This helped the moderator to better follow the participants’ line of thinking and achieve a clearer understanding of their reasoning, thoughts, and concerns In the third and final stage, the researcher followed up with questions about the experience with the interactive prototype. After completing their interaction with the prototype, the participants were again asked to rate their level of agreement with a set of predefined statements in relation to the scenario, from 1 (strongly disagree) to 5 (strongly agree). The predefined statements were related to the perceived competence and efficiency of the AI-enabled system, the anticipation of negative consequences of its use in public services, and the understandability of the system. Further, participants were also asked about their levels of comfort and trust toward the specific system they tried. Lastly, a set of open questions about the different explanatory elements used in the prototype were asked. To ensure a comprehensive analysis, the collected data was categorized into three classes: scale ratings, spoken word feedback, and interactions with the prototype. Each class of data was first analyzed individually followed by a synthetization with the other two classes. The goal of the analysis was to identify themes that would help us to better understand the stances and concerns of participants (Braun & Clarke, 2006; Castleberry & Nolen, 2018). The data analysis was performed through the SCT framework, including an evaluation of power dynamics, the nuanced assessment of collective versus individual benefits, and an exploration of how participants perceived and experienced the various agreement mechanisms. First, the answers given by the participants to the predefined questions were evaluated. The scale ratings were aggregated in a spreadsheet, which provided a comprehensive overview of the data’s diversity and trends. This also enabled the identification of distinct groups, gauging participants’ perceptions, and helped to quantify and follow changes occurring throughout the study. In the next step, an in-depth analysis of the data obtained from both the think-aloud protocol, consisting of spoken responses, and the feedback participants provided in response to open-ended questions regarding various types of explanations was conducted. To carry out this analysis, the study recordings and interview transcripts were thoroughly examined in multiple rounds, to identify common patterns, perceptions, and recurring themes. Throughout this analytical process, noteworthy incidents, quotes, and specific expressions from the participants were highlighted and discussed among the research group members to collectively interpret these findings. This category of data yielded a wealth of valuable information concerning the concerns and understandings of the study’s participants. In a final step, the investigation extended to the examination of participant interactions with the prototype, based on the recordings as well as notes taken during the research sessions. A focus was set on determining whether specific information elements were accessed and how participants engaged with the specific information. By merging this analysis with the findings from the previous stages, the analytical examination was enriched and enabled the research team to uncover emerging concepts and recurrent themes through cross-referencing the observations and findings. Following the ADR principles of Reciprocal shaping (P3) and Authentic and concurrent evaluation (P5), findings from the empirical research session and the data analysis were continuously presented, discussed and evaluated within the research team, as well as the wider ADR team, including the organizational partners. Within these evaluations, the findings related the innovative artifact, such as participant remarks about specific information elements, perceptions of transparency and levels of understanding were discussed. Based on the evaluations, the prototype was continuously iterated to ensure and improve user utility. The prototype iterations and the discussions informed the formulation of the HCAI Design Principles. 5.4. Employee-focused Research Strand. The following section describes the employee-focused research strand, focusing on exploring the perspectives of public servants. 5.4.1. Foresight Approach. Due to external influences, the priorities of the organizational needs changed over the course of the project. The specific sick leave duration prediction system came under the scrutiny of the Norwegian Data Protection Authority (Datatilsynet), which raised concerns about the use of citizen data. The further development of the system was halted and no further plans for the introduction of such a system were in place. To maintain practical relevance, the caseworker-focused research strand was conducted with a more abstract research focus. To cater to the evolved organizational need, and to ensure the participation and collaboration of employees, it was decided to employ foresight methods for this research strand. The Foresight approach was deemed suitable for understanding public servants’ perspectives on AI as it focuses on anticipating and shaping uncertain futures. It allows people to think ahead and to consider, model, and respond to future eventualities. It provides the opportunity for insights into imagining and shaping a social, fair, and just future. It excels at capturing the complexity of uncertain future environments, by encouraging participants to envision multiple scenarios and anticipate potential challenges and opportunities from the employees' perspectives. 5.4.2. Caseworker Participation. To explore the perspectives of public servants, a total of 19 caseworkers from three different local NAV offices were interviewed as part of the employee-focused research strand. The caseworkers were found through a liaison between the university and the local branches of the public service organization. Initial interviews were conducted during November 2022 and happened digitally via video-conferencing software. The following interviews were conducted in person and held between December 2022 and March 2023. The initial interviews in November and December 2022 were conducted as group interviews, with two or more caseworkers participating together in a group interview. The interviews in January and March 2023 were conducted with individual caseworkers. Participating employees were selected due to their role as caseworkers within the public service organization. Within the structures of the welfare system, caseworkers constitute a group of frontline workers, who regularly encounter sensitive data, i.e. personal health records of citizens, and are likely to be affected by the introduction of AI systems in their workplace. 5.4.3. Empirical Data Collection and Analysis. The Foresight method was embedded within semi-structured interviews allowing for additional questions and inquiries depending on the flow of the conversation. Similar to the citizen-focused research strand, an interview guide including three parts was developed. In the first part general information about the participants, such as age, official job title, and duration of their tenure were collected. Like in the citizens study, the caseworker participants were also asked about their self-reported prior knowledge about AI. This was asked to provide the research team with a better context when analyzing the data and to allow for adjustments in the use of level of technical detail when engaging with the participants in the interview process. The second part contained questions about current processes in their roles as caseworkers within the public service organization. Participants were asked to explain what their daily work tasks entail, e.g., decisions they have to take, dialogue meetings, and types of support. As a follow-up to the described tasks and procedures, they were asked about how they could imagine AI being useful for their tasks. These questions were asked to better understand the general work of caseworkers, as well as an introduction to the Foresight section of the interview. In the third part, different types of futures were explored together with the participants. To probe participants into different possibilities and to provide some initial guidance, they were asked to imagine four different directions. The first direction was the imagination of best-case and worst-case scenarios concerning the use of AI systems in their work. This initial future scenario helped to get a first impression of the caseworkers' sentiment and to define some initial boundary conditions for the imaginary futures. Following the initial future exploration, participants were also asked to imagine the most probable, the most preferable, and the most likely futures. If it was noted that participants struggled with imagining the use of AI systems in their workplace, they were provided with the hypothetical scenario based on the AI-enabled sick leave duration prediction system and the interactive prototype. Providing a graspable example helped the caseworkers to explore potential futures. The interviews were then transcribed, coded, and grouped into themes, using thematic analysis. This method enables the identification, analysis, and reporting of identified themes. The method facilitates the generation of rich, detailed, and complex accounts of the data while allowing for flexibility as well as interpretative depth. The analysis followed a three-step process, including data familiarization, code generation, and theme identification. In the first step, the research team went through all interview transcripts to become familiar with the collected data and began to collect initial observations. During the second step, the data was systematically coded to condense the interview content and prepare for the emergence of common themes. In the final step, the codes were sorted and aligned according to the different lenses of the SCT framework, and the emergent themes were undergirded with coded data extracts. Themes were constantly and iteratively refined and redefined, moving between the raw data, the coded extracts of data, and the identified themes. The thematic analysis helped to organize findings and identify patterns within the data. 5.5. Formulation of the Design Principles. The goal of this doctoral research is the formulation of HCAI design principles based on the situated learning achieved throughout the ADR project. It seeks to transform specific findings into broadly applicable guidelines that can address similar challenges across various contexts. The envisioned design principles aim to bring the perspectives of citizens as well as public sector employees together to encourage a more wholistic perspective during the design, development and deployment of AI systems within the public service context. This inclusive approach helps to address the unique challenges and ethical considerations that arise when implementing AI systems in government services. It ensures that the resulting HCAI systems are not only technologically advanced but also aligned with public values, democratic principles, and the diverse needs of the citizenry. Since the pace of groundbreaking technological advances and even loftier forecasts for the field of AI have dramatically picked up in recent years, the formulation and proliferation of frameworks, guidelines, and checklists aimed at promoting a responsible AI implementation is only consequential. However, it is crucial to recognize that most existing guidelines carry an inherent commercial bias due to their corporate origins. But the purposes, roles, and materials constituting a public service are very distinguished from services originating from a commercially-centered context. As identified in the empirical studies with citizens as well as with employees of a public service organization, the understanding of the relationship between governments, governmental institutions and citizens is different than for a commercial company. For Human-Centered AI Design Principles to be impactful and practical, these differences need to be recognized. 5.5.1. A Formulation Framework for Design Principles. A framework for the formulation of Design Principles in Information Systems has been proposed by Gregor et al. (2020), delineating five essential components that should be incorporated into each design principle: 1. The roles of the actors involved in developing and using these design. principles, 2. Descriptions of complex IT-based artifacts that require design principles to. allow for decomposition, 3. Representation of various types of artifact-based actions, from affording. user action to performing actions, 4. The means to achieve the ends envisioned by the design principles, and. 5. The option of providing rationales that justify their formulation. This structure highlights the crucial difference between the implementer, who actualizes the abstract design principle, and users, who utilize this realization to meet specific objectives. This differentiation demands that design theorists meticulously develop principles that guarantee that these principles are not only theoretically robust but also practically significant and relevant in real-life contexts. Table 5 illustrates the architecture of the design principles established in this study, by an exemplary design principle following the template outlined by Gregor et al. (2020). In total twelve design principles have been formulized, which are presented in chapter 7. 5.5.2. Evaluation of Design Principles. Evaluation is a crucial aspect within any Design Science Research (DSR) project. A strategic approach to evaluating DSR studies entails clarifying objectives, selecting an evaluation method, identifying the properties to assess, and creating specific evaluation episodes. By explicitly analyzing the roles of key stakeholders in the formulation of design principles, the approach enables DSR researchers to assess the usability of generated design principles across different user groups. Two main considerations are whether the design principles are clear and useful for implementers and how well they help users achieve their goals considering the resultant instantiations from applying the design principles. The formulated design principles underwent a multi-venue evaluation process to ensure their relevance, robustness, and applicability. Iterative discussions within the research team were held continuously to refine the principles. These also included collaborative meetings together with organizational partners to align the principles with the organization’s perspective and operational realities. Further, the Design Principles were presented at an internal event organized within the collaborating organization, where employees from diverse technical, design, and legal backgrounds provided feedback on the first iteration of the Design Principles. Finally, the next iteration of the Design Principles was presented at a nationwide internal NAV conference, involving practitioners from managerial and front-office roles from all over Norway. This venue featured an interactive assessment, allowing participants to engage directly with the principles and contribute practical insights from their varied professional experiences. 5.6. Summary of the Chapter. This chapter provided a detailed description of the employed research design. It outlined the philosophical stance of the study and introduced the ADR methodology. Additionally, the chapter explained the two different but connected research strands and concluded with the Design Principles formulation and evaluation. The next chapter focuses on the dissemination of findings, detailing the research publications that emerged from the research, and how they contribute to this overall doctoral study. 6. Publication of Findings. This doctoral thesis is comprised of seven publications (Table 6). In this chapter, a summary as well as an overview about the individual findings is presented, but the full text of each publication is available in the Appendix. Publications #1, #2 and #3 address the sub-question SRQ-1, SRQ-2 and SRQ-3, which build the foundation to answer the overarching main research question. Publications #4 and #5 provide the conceptual foundation for the understanding of Human-Centered AI. The main research question is then answered in publications #6 and #7. Further, the chapter also provides an overview on how each publication contributes to the general aims of the thesis. 6.1. From Commercial Agreements to the Social Contract: Human-. Centered AI Guidelines for Public Services (#1) Summary: HCAI is a concept more and more used in the discourse on how to guide the design, development and deployment of AI in responsible and trustworthy ways. Organizations such as Microsoft, Apple and Google providing their own HCAI guidelines, operationalizing theoretical concepts to inform the practice of AI design. Yet, they share the commercial contexts of their origin organization. This paper focuses on AI for public services and on the special relationship between governmental organizations and the public. Approaching human-AI interaction through the lens of SCT it proposes amendments to improve the suitability of an existing HCAI framework for the public sector. Following the methodological approach of ADR, Google’s PAIR guidelines were applied, assessed, and adapted in collaboration with a public organization. The guidelines informed the design of an interactive prototype for AI in public services and the process also revealed gaps and potential enhancements. Findings: The paper describes three identified amendment themes to extend established design principles by context-specific considerations and recommendations. Specifically, a) to articulate a clear value proposition by weighing the public good vs. the individual benefit, b) to define boundaries for repurposing public data given the relationship between citizens and their government, c) to accommodate user group diversity by considering the different levels of technical and administrative literacy of citizens. Contributions: The identified amendments contribute to the responsible design, development, and deployment of AI systems in the public service context. They should be understood as additional considerations, supporting practitioners in the design of AI systems. Conceptually, the paper introduces the lens of SCT to the design of AI systems in the public sector. The amendments aim to shift the perspective within human-AI interaction, acknowledging that exchanges are not always subject to commercial agreements but can also be based on the mechanisms of a social contract. Finally, the paper also contributes methodologically by demonstrating how generalized outcomes need not always start from a blank page but can also be performed by revisiting existing principles and recommendations and amending or adapting them based on empirical learnings. 6.2. Exploring Citizens’ Stances on AI in Public Services: A Social. Contract Perspective (#2) Summary: This paper explores citizens’ stances toward the use of AI systems in public services in Norway. Utilizing a Social Contract perspective, the study analyzes the government « citizen relationship at Macro, Meso, and Micro levels. A prototype of an AI-enabled public welfare service was designed and presented to 20 participants who were interviewed to investigate their stances on the described AI use. Findings: The analysis identified a generally positive attitude and three factors contributing to this stance: (a) the high level of trust in government (Macro level); (b) the balanced value proposition between individual and collective needs (Meso level); and (c) the reassurance provided by having humans in the loop and providing transparency into processes, data, and model’s logic (Micro level). Contributions: The findings expand extant research on AI adoption in the public sector by foregrounding the government–citizen relationship that has received limited attention in this body of literature. Further, this study provides insights into how important it is for public organizations to ensure that public’s goodwill is not eroded. The study shows the importance of articulating a clear value proposition that considers both the individual and the collective interests and the provision of clear information about the data uses, the logic of models, and the processes followed, including the role of humans in these processes. Conceptually, the study proposes a SCT framework for researching socially responsible AI in the public sector. The framework can help to analyze AI in public services at three different levels: Macro, examining the power relationships among governmental organizations and citizens; Meso, investigating the gauging between the collective good and individual benefit; and Micro, exploring different agreement modeling mechanism 6.3. Human-AI Collaboration in Public Services: The Case of Sick. Leave Case Handling (#3) Summary: This research paper explores Human-AI Collaboration in public services, focusing on the needs and understanding of AI among caseworkers in a public welfare organization dealing with sick leave cases. Following a foresight approach with the participation of 19 caseworkers, the study delves into the tasks involved in their daily work, expanding beyond the digital touchpoints between humans and AI. It encompasses the broader scope of collaboration between humans and machines as joint cognitive systems and suggests thinking of AI as a collaborator that occupies specific roles. Findings: Caseworkers see potential for AI to improve prioritization, data retrieval, reporting and simple tasks, thereby simplifying their work and increasing efficiency. On the other hand, they emphasize that AI cannot fully replace humans due to the complexity of individual cases and the need for empathy, discretion and personal connections. Caseworkers envision AI playing an assisting rather than substitutive role, where artificial agents augment rather than compete with humans. This highlights both opportunities and challenges for introducing AI in welfare administration. Contributions: The defined taxonomy of different roles for AI systems contributes to the broader discourse on human-AI collaboration in the workplace. The study also contributes to practice, as it provides valuable insights into caseworkers’ expectations and needs for the design and implementation in the public service context. 6.4. Defining Human-Centered AI: A Comprehensive Review of HCAI. Literature (#4) Summary: This paper investigates the evolution of Human-Centered Artificial Intelligence as an emergent perspective on the design, development, and deployment of AI. It provides an overview of HCAI definitions, from the most established to the less common definitions found in the literature, highlighting the variety of emphases as well as the shared understandings among them. The proposed HCAI definition highlights the necessity to understand the involved and affected people. To identify and understand their needs and values, the new definition highlights the use of Human-Centered Design methods. Findings: In an HCAI context, needs and values are mainly manifested through the concepts of Augmentation, and Control. Augmentation refers to the idea of using AI to enhance human capabilities and performance, rather than replacing human beings with machines. Control, on the other hand, deals with the governance and management of AI systems to ensure that they operate ethically and safely. The paper highlights the importance of collaboration between AI and IS researchers to advance the HCAI agenda and ensure that AI serves the interests of society. Contributions: Based on the review, the paper proposes a new comprehensive HCAI definition, synthesizing the main features of the different definitions, which aims to facilitate a meaningful debate and contribute to advancing the scientific discourse about the responsible introduction of AI technologies. 6.5. Understanding Human-Centered AI: A Review of its Defining. Elements and a Research Agenda (#5) Summary: The article builds on and expands the work of publication #4, by delving deeper into the understandings, elements and methodological aspects of HCAI. In the first part of the review, existing HCAI definitions are analyzed and compared, resulting in a novel, more comprehensive HCAI definition. In the second part, the review explores the identified constituents of HCAI, namely Purpose, Values, and Properties. The third part of the review explores Human-Centered Design processes, methods, and tools and their applicability for HCAI. Lastly, the review discusses the characteristics and critiques of HCAI and provides an extensive research agenda. Findings: Within the corpus of analyzed literature, the elements of Purposes, Values and Properties have been identified. Purposes of HCAI can be categorized into Augmentation, AI autonomy, and Automation. The element of Values relates to Ethics, Safety, and Performance, and the element of Properties covers aspects of Oversight, Comprehension, and Integrity. Furthermore, a Human-Centered Design (HCD) foundation was identified, consisting of established as well as novel HCD tools, methods and processes. Contributions: This Systematic Literature Review contributes to the conceptual clarity around the definition, conceptualization, and implementation of HCAI. It provides a synthesis of prior research on HCAI, a comprehensive definition, outlining its key elements, and points to the fundamental role of Human Centered Design methodological approaches for HCAI. Additionally, it presents an extensive research agenda, pointing out major themes to fuel further investigation and to advance knowledge on creating AI systems that genuinely benefit humanity. 6.6. Citizens-Focused Design Principles for Human-Centered AI in. Public Services (#6) Summary: To reconcile abstract theoretical concepts with the practical realities of AI development and deployment in the public sector, this study formulates citizen-focused HCAI design principles, recognizing the unique aspects of the public sector. The study argues for bespoke AI design and implementation strategies that prioritize societal needs over commercial interests and incorporate citizen perspectives into the design principles. The design principles are based on the findings from an Action Design Research project, interpreting the findings through a SCT lens. Findings: By incorporating citizen stances and perspectives into design knowledge, design principles do not only instruct but align AI systems in the public sector with a clear understanding of their objectives, application contexts, and the mechanisms and rationale underpinning their design. Moreover, by considering the diverse roles of stakeholders and the intricate relationship between technology features and user needs, the formulated design principles ensure that AI systems are responsibly crafted to serve the public good, underscore the importance of ethical considerations, and respect the boundaries within which these technologies operate. Contributions: The defined design principles in this research contribute to the information systems discipline by providing insights into citizen stances on a socially responsible design, development, and implementation of AI systems in the public sector. They aim to expedite the discourse on how to merge theoretical knowledge with actionable guidelines, creating new knowledge for scholars and practitioners alike. The DPs are building upon established and proven guidelines but are going beyond commercial interests of profit-oriented companies by incorporating critically consideration of citizen and public interests. 6.7. Human-Centered AI Design Principles for the Public Service. Domain: An Action Design Research Study (#7) Summary: Most existing design and development guidelines for Human-Centered AI primarily cater to a commercial context, they are not tailored to the specific needs of public services. This paper presents an Action Design Research study proposing public service-specific design principles for Human-Centered AI. The design principles are informed by multiple iterations of empirical research with citizens and public service employees acknowledging the multi-stakeholder nature of Human-Centered AI. The study recognizes an evolving understanding towards prioritizing human values and well-being in AI technologies. Furthermore, it considers the relationship between technology features in the public sector and citizens' needs to ensure that AI systems are developed with a commitment to fostering public trust and welfare. The study contributes to theory and practice by advancing the scholarly discourse on Human-Centered AI and informing AI strategies and implementations within the public sector. Findings: One of the primary values of applying design principles in AI development is their capacity to operationalize considerations for existing power dynamics, which becomes particularly important within the specific structures of the public sector. This involves balancing public good and individual benefits, recognizing and mitigating risks of data misuse or inadvertently introducing and reinforcing forms of bias. Effective design principles for the public sector help to ensure that AI systems do not exploit or disadvantage any group but instead work to empower and elevate the collective well-being. Furthermore, these principles are instrumental in balancing the benefits between individual needs and societal welfare. This balance is crucial in public sector applications, where the impact of AI can be profound and far-reaching. Contributions: The proposed design principles contribute to the information systems discipline by providing a combined perspective into citizen as well as employee stances on a human-centered design, development, and implementation of AI systems in public organizations. They aim to advance the discourse on how to merge theoretical knowledge with actionable guidelines, creating new knowledge for scholars and practitioners and providing helpful guidance for introducing AI in public services. 6.8. Synthesis of the Publications. The presented publications embody the overall narrative presented in this doctoral dissertation. The first three publications serve as the empirical foundation of this research study. Publication #1 examines existing commercially oriented HCAI guidelines and identifies necessary adaptations for the public sector context. In publication #2 the perspectives of citizens regarding the use of AI in public services are explored and analyzed, identifying a generally positive stance including three factors contributing to this stance. Publication #3 investigates the perspectives of caseworkers towards the introduction of AI systems into their work. The caseworkers expressed optimistic yet circumspect perspectives about the introduction of AI, seeing potential benefits but also raising concerns. Publications #4 and #5 are focused on Human-Centered AI, providing conceptual clarification and grounding for this dissertation. To be able to answer the overarching research question, more clarity on the concept of Human-Centered AI was required. This involved exploring how HCAI is defined in the literature, analyzing its constituent elements, and examining the processes, methods, and tools employed to achieve it. The findings from were brought together, formalized, constantly reflected on and iterated over the course of the project. Different iterations of the defined Design Principles have been published in the publications #6 and #7. 7. Design Principles. This chapter presents the defined Design Principles for both stakeholder groups: the citizen-focused design principles as well as the employee-focused design principles. 7.1. Citizen-focused design principles. 7.1.1. cDP-1 - Balancing AI Benefits for the Public Good and Individual. Freedom To address potential conflicts of interest between the public good and individual freedoms, AI system designers and developers must be transparent about different needs and benefits in a comprehensible and accessible manner. This will help stakeholders comprehend the implications for the other involved actors. Recommendation: Communicate clearly and briefly the expected impact for the public and individual citizens. 7.1.2. cDP-2 - Comprehensive, Inclusive, and Necessary Data. To ensure the responsible use of citizens’ data, AI system designers, and developers who decide on which data to source for AI systems in the public sector, need to align system requirements with citizens’ rights. This will help to critically assess data needs and evaluate the availability of existing data, or the need for new data collection to adhere to the principle of data minimization. Recommendation: Explain which data is used, where it comes from, and the rationale behind its use. 7.1.3. cDP-3 - Reconciliation of Governmental Mandates and Process. Integration for AI Systems To ensure citizens are aware of the role of the AI model within the overall operational process and the link to governmental mandate this should be clearly defined. The link to legal or regulatory documents can be provided but the language needs to be kept simple. Recommendation: Present how AI fits in the overall process to fulfill the governmental mandate. 7.1.4. cDP-4 - Gradual Information Provision. AI system designers and developers should ensure citizens understand the system logic and its data use by gradually providing clear information allowing people to view what is relevant for them on the level of detail they're comfortable with. Citizens must receive relevant and timely information about how the AI functions, its inputs and outputs, and its effects on them. To help citizens, information provision should consider the diverse technical literacy of the wider public. Recommendation: Present an overview first and allow users to access details and drill down if desired. Use familiar terminology and design patterns and provide contextual explanations. 7.1.5. cDP-5 - Voluntary Feedback Mechanisms. AI Systems designers and developers should enable direct citizen participation to facilitate continuous learning about citizen needs, experiences, and system shortcomings. This can help to foster a participatory environment where citizens feel valued and heard, strengthen public trust and engagement with citizen input directly informing enhancements and adjustments. Recommendation: Integrate accessible feedback mechanisms and communicate how the feedback will be used to help further improve the AI system or the overall service provided. 7.1.6. cDP-6 - Appropriate Consent practices. To facilitate participation and human agency, AI system designers and developers need to gauge whether the consent of citizens toward the use of an AI system, or their personal data needs to be obtained or if it can be implied. Further, appropriate consent mechanisms need to be explored. This will help to address concerns about data ownership and the balancing of individual and public good. Recommendation: Obtain citizen consent where data ownership is unresolved. 7.2. Employee-focused design principles. 7.2.1. eDP-1 - Sensible Resource Allocation. To support employees of public service organizations focusing on people who need the most support, AI systems designers and developers should focus on features and functionalities that support the prioritization of work. Recommendation: Enable employees to focus more on selected subgroups of people that need extra support. 7.2.2. eDP-2 - Automating Repetitive Administrative Tasks. To support employees within public service organizations in using their resources most efficiently, AI systems designers and developers should enable them to delegate repetitive administrative tasks to automated systems. Recommendation: Enable employees to hand over repetitive tasks to automated tools. 7.2.3. eDP-3 - Consistency over Personalization. To ensure equal treatment, AI systems designers and developers should ensure consistent behavior in the AI system to prevent unjust and discriminatory outcomes through learned adaptation from individual employee behavior. Recommendation: Prevent employee personalization of AI system behavior. 7.2.4. eDP-4 - Providing Legal Reassurance. AI systems designers and developers should provide employees with transparency on data sources, to ensure integrity, lawfulness, and traceability of the data used within AI systems. Recommendation: Explain which data is used, the statutory basis, and the rationale behind its use. 7.2.5. eDP-5 - Voluntary Feedback Mechanisms. AI Systems designers and developers should enable direct employee participation to facilitate continuous learning about employee needs, experiences, and system shortcomings. This can foster a participatory environment where employees feel valued and heard, strengthen trust and engagement with employee input, and directly inform enhancements and adjustments. Recommendation: Integrate accessible feedback mechanisms and communicate how the feedback will be used to help further improve the AI system or the overall service provided. 7.2.6. eDP-6 - Enabling Human Decision-Making Power. To facilitate cooperation between employees and AI systems in the public sector, AI systems designers and developers should create AI systems that are perceived as helpers, with humans staying in control of the decision-making process. Recommendation: Provide humans with the ultimate power of decision when teaming with AI-enabled systems. 8. Discussion. This chapter discusses the design principles against the prior literature and the conceptual boundary conditions of the SCT framework introduced in chapter 3. Finally, a research agenda in the form of an Extended SCT framework is proposed. The public sector's mandate to serve the collective interest, ensure equity, and maintain democratic values creates a complex landscape for AI implementation. Unlike commercial entities primarily driven by profit and market share, public institutions must balance efficiency with fairness, transparency, and accountability. The concept of public value also plays a crucial role in distinguishing public sector AI from its commercial counterparts. While commercial AI often focuses on user satisfaction and engagement, public sector AI must consider broader societal benefits, long-term sustainability, and the equitable distribution of services. This balancing act is further complicated by the diverse needs of the citizenry and the long-term societal impacts of governmental decisions. The existing power dynamics inherent in the government-citizen relationship add another layer of complexity to public sector AI, highlighting questions about consent, data privacy and ownership, and the potential for AI to influence democratic processes. Moreover, the public sector operates under unique constraints, including legislative mandates, budgetary limitations, and heightened public scrutiny. And finally, AI solutions need to be not only technologically sophisticated but also aligned with public policy objectives, legal frameworks, and ethical standards specific to governmental operations. To facilitate such fundamental organizational changes, we must engage in visionary and strategic thinking as well as constant reflection on the future societies we are about to create. Following design principles for AI can be useful because they operationalize considerations for existing power dynamics, recognizing and mitigating risks of data misuse or inadvertently introducing and reinforcing forms of bias. Effective design principles for the public sector help to ensure that AI systems do not exploit or disadvantage any group but instead work to empower and elevate the collective well-being. In other words, these principles are instrumental in balancing the benefits between individual needs and societal welfare. This balance is crucial in public sector applications, where the impact of AI can be profound and far-reaching. In this context, well-defined design principles serve as crucial guideposts, facilitating a comprehensive consideration of AI systems across Macro (Trust & Transparency), Meso (Balancing Public Good vs. The Individual Benefit & Participation), and the Micro (Diversity, Equality & Agency) levels, considering the diverse perspectives of the affected stakeholders. They aim to ensure that an AI system will serve not just individual users or one specific group but also contribute positively to the community and society at large. Liikkanen (2019) describes that human-centered design will be crucial in further defending humans, particularly underprivileged users at risk of being mistreated by AI. By incorporating diverse stakeholder perspectives, these principles not only provide practical how-to guidelines but also emphasize the critical importance of clearly defined objectives, nuanced understanding of application contexts, and well-articulated rationales underlying HCAI design decisions. 8.1. SCT – Macro Level. Using AI systems within the public sector has the potential to enable a wide array of improvements and benefits – for governments, the deploying organizations as well as citizens. However, perceptions and expectations of stakeholders are varied, specifically among citizens and public servants. A mix of optimism and concern acknowledges AI’s potential to drive innovation and efficiency, but also fuels a wariness regarding its risks, including bias, indisputability, and a loss of autonomy. Analyzing the empirical data collected for this doctoral research through a Social Contract perspective identified generally positive stances toward the use of AI in public services, both from citizens as well as from the employees of the public service organization. However, both groups also have their specific concerns which need to be considered and addressed when designing, developing and deploying such systems. Engaging different stakeholders in the process of shaping the future and cultivating deep empathy is a fundamental element of Human-Centered AI. It can lead to better policymaking, more effective socialization of emerging technologies, and greater legitimacy for hard decisions, as well as creating constituencies for long-term change. Within the studies of this doctoral research, a high level of trust in the Norwegian government has been identified as a contributing factor. Trust in a government and its executive bodies to act in the best interests of its citizens forms a foundational aspect of the Social Contract and is a vital prerequisite for the proper functioning of a society. This trust empowers leaders to carry out their governmental responsibilities and maintain the legitimacy they derive from the populace. Yet, there are often prevailing prejudices against the public sector. As Jos (2006) describes, the vision of public administration as a politically neutral arbiter of the common good, a professional community whose connection with the public is both more direct and more authentic than that of elected leaders, has been fueled by the scapegoating of bureaucrats and the proliferation of multiple mechanisms of oversight and control. However, within the Nordics, and especially within the sociocultural context of this work – Norway – the public sector has been spared from being the focus of public scolding. 8.1.1. Legitimate Data Usage and Lawfulness. One of the most important aspects when implementing AI systems for public services, is to ensure data security and privacy as they rely heavily on personal data. The fact that public service organizations have potentially access to vast amounts of personal, partly sensitive data doesn’t necessarily imply that this data can be freely repurposed, e.g. to train AI models to improve administration and decision-making processes. This devolves a huge responsibility to governments and requires particular scrutiny about potential misuse or the discrimination of marginalized groups. The Social Contract is considered breached, if citizens are unaware of unauthorized data collection or data processing, or data transfer to other parties without their explicit consent, constituting a power abuse from the governmental organization. In the context of the studies conducted within the research, the trust in government and governmental organizations expressed by the participants describes the expectation of responsible use of AI and the processing of personal data. Participants expressed a clear foundation of trust in the Norwegian welfare organization about the data handling. Together, the means of consensus and contestation also constitute a measure of power for the citizens, as it provides a level of sovereignty for their data ownership and usage. But this can only happen within the boundary conditions the Social Contract, meaning the individual freedom needs to be weighed against the public benefit. This mutual dependence yields responsibilities and creates new requirements of comprehension for the processing of sensitive citizen information and the ability to provide an informed consent or contest AI outcomes (de Sousa et al., 2019; Misuraca & Van Noordt, 2020) To be able to provide an informed consent, citizens need to understand what data will be used, how it will be used and what their rights are to object AI informed decisions. To achieve such an understanding transparency into data ownership and means of objection can be a helpful property. The design principle cDP-2 -Comprehensive, Inclusive, and Necessary Data has been formulated to operationalize this need. Having deliberations for which data is needed, which data is available, and which data can be used with or without additional consent as part of the early design process can ensure that no unnecessary data is collected, stored or processed and the value of privacy is considered. Furthermore, such questions can also facilitate a level of accountability, if justifications on the intended use of citizen’s data are required during the inception of AI systems. At the same time, employees of the public service organization also require certain levels of transparency. Caseworkers for instance are envisioned to use AI systems as decision-support tools, which will make suggestions for an appropriate service or repercussions for citizens. These decisions, or decision suggestions are based on the data the system has used or was trained on, but caseworkers are bearing the accountability for the decisions. And although caseworkers are still free to exercise discretion, they need to be certain of the legal authority for the use of citizen’s data to fully unlock an AI systems potential. As stated by Thompson (1987), public servants are accountable for their personal conduct as well as for the conduct of the organizations they represent. But AI systems can be complex and opaque, making it difficult to determine how a certain outcome has been calculated. Caseworkers might hesitate to use any AI systems or follow suggestions coming from such systems, if there is a risk for illicit data use. Caseworkers should be able confidently use AI systems to share resulting information to citizens in a clear and comprehensible way to maintain accountability and traceability. Being reassured that the AI system and its components are legally sound to be used can help to foster trust as well as efficiency in the work of caseworkers. Furthermore, this can also support employees in spotting potential issues or causes for misinterpretation. It allows for employee scrutiny and helps to identify any biases or unfairness in the AI systems, calling on the values of responsibility and integrity. It is therefore pertinent to provide a legal reassurance for the employees to provide transparency in the use of personal citizen data. This is reflected in the design principle eDP-4 - Providing Legal Reassurance, which aims to raise awareness of this dimension during the inception of an AI system in the public sector, to support AI adoption among public sector employees. As an additional benefit, providing information about the used data, its sources and the legal grounds to do so can enhance the knowledge and understanding of caseworkers, addressing the knowledge gap that has been found to hinder the effective implementation of AI in government. The dimension of transparency also pans out in multiple directions. And when it comes to transparency around AI used by public agencies, the concept is even more complex than traditional governance transparency due to the opaque nature of AI systems. For instance, the recently accepted EU AI Act proposes more comprehensive documentation and transparency requirements for high-risk AI systems. But even the heavily praised EU Act is incomplete. The Act mandates that such systems be transparent to one type of users, but not necessarily to those affected by the decisions (Art. 13, EU AI Act, 2024). Also, the General Data Protection Regulation (GDPR) provides some transparency rights, such as the right to obtain meaningful information about the logic of automated decisions (European Parliament & Council of the European Union, 2016). However, this is limited to cases involving personal data and fully automated decision-making and does not cover decision-support systems. 8.1.2. Procedural Transparency. But even just shedding light onto the algorithm won’t be sufficient, as it also needs to be explained how an AI systems is integrated into the administrative processes and how their decisions affect citizens. Providing such transparency provides information and clarity on the legitimate use and reuse of citizen data. But at the same time, it creates vectors of accountability, so that citizens are assured they have the legal foundation for any cause of action against the public service organization. Without such clear accountability structures, it becomes challenging to determine liability in case of faulty or wrong decisions made by AI technology. Therefore, the governmental mandate under which specific tasks and processes are conducted need to be clearly stated and explained as well. But as Olsen et al. (2024) state, there is a need to balance transparency with legitimate interests such as proprietary information and internal deliberations. This balance must be carefully managed to avoid undermining transparency goals. One approach for this balancing can be found in an early consideration within the AI lifecycle. Being part a design principle that encourages comprehensive deliberation can help to address this situation as early as possible. This informed the design principle cDP-4 -Reconciliation of Governmental Mandates and Process Integration for AI Systems. The role of the AI system within the overall operational process should be linked to the benefits such systems are intended to unlock, for individual citizens as well the collective good. Explaining the need, expected outcomes, and improvements can help to provide an overview and better understanding of the bigger picture. Locating the use of the AI system within the organizational process it is a part of, can enhance trust, transparency and foster a perception of fairness and justice in an AI system used in the public sector. Such process transparency can take many forms, from text-based explanations to flow diagrams or other means to convey the context purpose an AI system is used. This also provides an opportunity to highlight the role of humans within the process and how it relates to the use of the AI system. The explanations can provide visual aids that help to differentiate between the supporting role of the AI system and the ultimate power of decision of the caseworker. This aspect will be further discussed in the Micro level of the SCT-Framework. While the attention for HCAI systems is often placed on augmenting human capabilities, it's important to recognize that automation can also be a useful purpose if the context provides a sensible and appropriate use case. For instance, in the context of case management, AI can serve as a tool allowing humans to benefit from automation without relinquishing control. The conducted empirical studies identified legitimate concerns about AI systems performing autonomous evaluations, assessments, and decision-making, both from citizens as well as from public sector employees. However, potential benefits of AI systems in other capacities have been acknowledged. Caseworkers see beneficial use of AI technologies, if considered to be a supportive tool rather than a replacement for human judgment and interaction. This perspective also introduces an interesting, nested approach to HCAI purposes, where Augmentation is enabled through automation or autonomy. The public sector caseworkers indicated that a potential benefit they imagined could be to hand off certain time-consuming and administrative tasks to AI systems. Examples mentioned were scheduling, data retrieval or reporting. Such tasks were perceived as time-consuming and little target-aimed towards the main functions of providing aid, counseling and nurture in their role as public servants. In accordance with, this aligns with the concept of rationalization, which aims for the highest degree of efficiency. This indicates that caseworkers feel they could provide better and more impactful services, if relieved from administrative burdens and being empowered to focus on their core governmental mandate. This potential is manifested in the design principle eDP-2 - Automating Repetitive Administrative tasks. Additional benefits of automating certain tasks can be improved reliability and efficiency, if such processes are conducted in a faster and more precise way. But to understand which tasks should be automated and how AI systems can help caseworkers without violating aspects of control and agency, involving stakeholders in the design and development process is vital. Close participation of different roles can provide useful and necessary perspectives into the decision of automating public services. This can include caseworkers, to learn about their needs when supporting them to focus on their core primary duties, citizens to inquire about their level of comfort and consent towards automated processes, as well as other managerial, financial or organizational perspectives. 8.2. SCT – Meso Level. Interpreting the findings from the empirical studies through the Meso level of the SCT-framework, reveals crucial insights into the delicate balancing act between the public good and individual benefits in HCAI implementation. Furthermore, it also discusses how participation of directly and indirectly affected stakeholders goes beyond the initial stages of design and development of AI systems. 8.2.1. Public Good vs. The Individual Benefit The intellectual foundation for this analysis can be found in Jean-Jacques Rousseau's seminal work, The Social Contract (1762). Rousseau states that governmental legitimacy stems from the consent of the governed, which implies a fundamental obligation for governments to uphold the general will. Consequently, public service organizations, as the executive arm of governments, inherit this duty to serve the public interest. In the context of AI applications in public services, this translates to ensuring that these technologies serve collective interests and promote equity among citizens – the public. In the understanding of Rousseau, the public consists of many individuals which share the interest of being a member of this political community and by that forming a general will. This perspective assumes that all citizens are capable of adopting a moral stance aimed at the common good, potentially reaching unanimous decisions. However, this idealistic perspective can be inconsistent with the behavior and interests of individual members of the public body. Even though humans are aware of being part of a bigger societal group, personal integrity and benefit can trump altruistic actions. As it is stated by Harari (2024), democracy is based on the understanding that the people is never a unitary entity and therefore cannot possess a single will. Every society consist of many different groups, with a plurality of opinions, wills and interests. Harari describes democracy as a conversation which presupposes the existence of several legitimate voices. It is therefore a constant balancing act of managing potentially conflicting interests which has fundamental implications to the design of algorithmic systems. Within the empirical studies, the importance of a clear value proposition for AI systems being deployed within the public sector, to explain how it can benefit collective and individual needs was identified. Such formulation of a promise of value raises expectations around transparency, explainability and trustworthiness but also sustainability, fairness and justice. This pondering act act is formalized in cDP-1 - Balancing AI Benefits for the Public Good and Individual Freedom. Another notion related to the theme of balancing public good and individual benefits that was imagined during the foresight interviews, was support on prioritizing which cases have the highest urgency or demand. Such integration of AI systems to support resource allocation and case prioritization presents a compelling opportunity to enhance both individual service delivery and overall public benefit. Caseworkers stated that AI systems could be used to improve case prioritization, helping them to devote their time and attention more equitably and effectively to people that really need attention and care. By reducing time spent on less critical tasks and unnecessary meetings, caseworkers can dedicate more attention to individuals requiring intensive support, which benefits particularly vulnerable citizens. Furthermore, leveraging AI systems can be used to streamline workflows and reduce waiting times which improved overall service delivery for the entire public. There even is potential for preventative interventions, if high-risk cases can be identified early. This could possibly help to improve public health and reduce long-term costs. This aim for a prudent use of available resources is addressed in eDP-1 – Sensible Resource Allocation. Analyzing these findings on the Meso level of the SCT framework reveals a profound awareness among public service employees of their obligations to serve citizens effectively. As stated by Thompson (1987), public servant should act as guardians of the public interest. This also rhymes with the understanding of the ethic of responsibility defined by Weber in his works on political sociology. The commitment to fulfilling their governmental mandate underscores the importance of designing AI systems that respect, reflect, and support employees in performing their duties. 8.2.2. The Right to Say. A crucial aspect of HCAI is the incorporation of the perspectives, needs and concerns of stakeholders as it constitutes the bedrock of a human-centered approach. In the context of AI systems in public services, this means, among others, the citizens who will be affected by the AI systems as well as public servants who are going to use them. Such formative involvement can be interpreted as a form of direct participation in governance processes. In the understanding of Rousseau (1762), governmental decisions like laws and legislation must be ratified by the people to be legitimate by participation. Enabling ways to provide such feedback aligns with the ideals of participatory democracy, where individuals have the right and responsibility to contribute to decision-making processes. The same procedure needs to apply for the use of advanced technology when used by public institutions. But feedback doesn’t cease importance once AI systems have been deployed and are actively used in the field. It is still important to listen and incorporate the feedback from the affected people, public servants as well as the public, who is considered the sovereign within a SCT understanding. According to Rousseau (1762), sovereignty is the exercise of the general will. And by providing feedback, citizens avail themselves to their right of sovereignty and participate in shaping what affects them. This means, the Social Contract requires citizens to be active participants in governance, ensuring that the general will is accurately represented. This process of continuous feedback constitutes the evaluative and iterative nature of HCAI. Learning and improving based on the provided insights and reflections allows HCAI systems to adapt to ever evolving circumstances as well as addressing initial misconceptions and errors. Provided feedback can be explicit or implicit. For instance, making use of an opt-out option can be considered implicit feedback, as the dissent communicates a lack of agreement. Within the empirical studies, users were encouraged to express their opinions and objections, which not only facilitates the direct participation of citizens but also addresses the potential question of data ownership and upholds their right to object to personal data usage without negative consequences or justification requirements. Therefore, citizen feedback is a direct expression of the populace’s needs, preferences, and concerns regarding public services and policies, including the deployment of AI systems. At the same time can public sector employees provide valuable insights based on their professional experience that can inform the design as well as the continuous development of AI systems. Incorporating such specialized knowledge aligns with the understanding of a public servant in the Social Contract understanding, who is under the obligation to provide feedback to enhance efficiency, effectiveness and moral foundations. This mechanism is formalized in the design principles cDP-5/eDP-5 - Voluntary Feedback Mechanisms. Both citizens and public servants can participate in shaping policies and the implementation of AI systems by providing feedback that reflects their experiences and needs. Their input helps to define functionality, scope and boundaries of AI applications to better serve the public by ensuring that services are efficient, equitable, and responsive. To encourage direct participation and continuous refinement of AI systems through feedback, AI system designers and developers of AI systems should clearly communicate the value of feedback and provide transparency in how the feedback can contribute to the improvement of the AI. This can foster a participatory environment where stakeholders feel valued and heard to directly inform enhancements and adjustments. Therefore, AI systems designers and developers should facilitate direct participation, empower stakeholders to their right to say. This can enable continuous learning about stakeholder needs, experiences, and system shortcomings. Integrating accessible feedback mechanisms and communicating how the feedback will be used will help further improve the AI system or the overall service provided. 8.3. SCT - Micro Level. 8.3.1. Human Agency. The integration of AI systems in public services requires a careful balance between technological efficiency gains and human oversight. Within the empirical studies, the expectation that AI systems in the public service sphere will incorporate human safeguards was highlighted both by citizens as well as public service employees. Such safeguards include having humans in the loop during decision-making processes, providing possibilities to reach out to human actors for help or contesting a specific outcome. Citizens affected by AI-based decisions often lack the means to challenge these decisions due to the opacity of AI systems, making it difficult to exercise their rights. This has been also seen in the empirical studies, where individual citizens are concerned about their right to appeal. Human involvement provides citizens with a tangible means to question and contest AI-driven decisions, preserving their fundamental right as sovereign individuals in a democratic society (Rousseau, 1762). Furthermore, the availability of human contact persons is perceived as a mechanism to seek clarification, assistance, or empathy in complex situations. This is based on the assumption that AI systems might not fully take into account the consequences of specific situations or might lack the necessary sensitivity. The presence of a human counterpart can alleviate such prevailing anxieties and provide assurance and comfort. This also relates to the belief that human oversight can ensure better alignment with societal values and the general will, since human actors can provide necessary ethical judgments in complex scenarios where AI systems may lack nuanced understanding. By that, the assurance of human oversight can bolster public confidence in AI-driven public services, crucial for their long-term acceptance and success. Interpreting the theme of Human-in-the-Loop on the Micro level of the Social Contract framework reveals a profound insight: human actors are often perceived as more appropriate counterparts in agreement modeling processes than machines. This understanding touches upon Max Weber's concept of bureaucracy and the role of public servants. In a Weberian sense, public servants are expected to act with a sense of expertise, duty, ethical responsibility, and in service of the public good. This perception stems from a deeper trust in human judgment, especially when it comes to understanding and acting in accordance with the general will. At the same time, this expectation creates a potential for tensions if the general will is not aligned with the interests of the individual. Interestingly, some participants in the citizen studies mentioned that they expect AI to be more objective than human agents which they see as positive. They stated that they are expecting fairer and more equal treatment being the result of an impartial AI system. But this raises the question to what extent AI systems can embody these qualities without human oversight. Additionally, the caseworkers that were interviewed in the empirical studies share a similar view, indicating that AI’s role should be to enhance, not supplant their efforts to effectively support citizens. They felt AI could aid certain tasks and processes but emphasized that human judgment, empathy and personalized attention would remain crucial for fulfilling NAV’s mandate. They imagine leveraging AI capabilities and using generated insights for their own decision-making process, while continuing to apply their judgment, potentially leading to more informed and nuanced outcomes. The expectations and concerns raised by the interviewed employees highlight the irreplaceable role public servants will also play in the future, as they combine professional expertise with the strengths of AI systems, rather than being replaced completely by technology. By retaining the ultimate decision-making power, being available for assistance, and serving as a point of contact for objections, public service employees fulfill a vital role in the contractual relationship between citizens and a governmental organization. When pondering about a long-term future, caseworkers did mention concerns about job displacement, the future of work and the challenges of reskilling the workforce, as well. But in general caseworkers are seeing the potential for AI in a much more positive light, rather than as a dystopian future. An important premise for enabling human decision-making power is the concept of controllability. Various understandings of HCAI share a common notion that Augmentation goes hand-in-hand with the concept Control. The maxim of Augmentation instead of replacement is based on the understanding that technology is created with the purpose of supporting humans, but keeping them in control. Consequently, control is another dimension that needs to be considered when designing and developing AI systems for the public sector. To address this expectation, the design principle eDP-6 - Enabling Human Decision-Making Power has been formulated. Shneiderman argues that control and automation are not necessarily two ends of the same spectrum, but rather, two separate dimensions. In his framework, high levels of control and high levels of automation are not mutually exclusive. Shneiderman claims that both control and automation are needed for HCAI systems. Xu et al. (2022) highlight a shift from human-centered automation to human-controlled autonomy. The same understanding is implied in the HCAI definitions by Holzinger et al. (2022) and in Xu’s earlier work. However, it's crucial to note that the implementation of Human-in-the-Loop (HITL) approaches in AI-driven public services is not without challenges. It requires careful design to avoid creating bottlenecks, ensure consistency in human interventions, and maintain the efficiency gains offered by AI. Additionally, there's a need for ongoing training and support for public service employees to effectively work alongside AI systems. Another form of human agency identified relates to considerations around consent. In the light of the Social Contract, every person has a property in their own person to which nobody has any right to but themselves (Locke, 1689). Especially in the context of AI in the public sector, empowering data ownership through legally binding agreement modelling processes is critical. Such agreement processes ensure that citizens have control over their data, deciding who can access it and for what purposes. Consent as an agreement modelling process facilitates participation, and touches on to concepts like data ownership, human agency and the balancing of individual and public good. Apart from consent, there are also other agreement modelling processes, such as bargaining, which describes the negotiation of terms among citizens that reflect their interests and concerns, aggregation as the process of collecting individual preferences to inform decisions, and lastly achieving equilibrium, which describes the balancing act of reaching a fair and sustainable agreement. In this doctoral research, consent was the only form of agreement modelling. Consent can take various forms, from active consent to a specific inquiry– often described as Opt-In mechanisms, the absence of an explicit dissent – referred to as Opt-Out, the existence and inherited consent from an earlier inquiry, or the implicit consent through the usage of a particular service. During the interaction with the prototype, citizen participants were reaching a decision point where they had to actively decide whether they would consent to the use of their data for the AI-enabled sick-leave prediction system. In the interaction with governmental organizations, citizens trust that their consent will only be used by the authorized organization for its intended purposes and to their benefit. This implicit understanding can be interpreted as being part of the Social Contract of the citizen-government interactions in the digital age. This expectation is upheld, even if the provided transparency and explanations are not fully comprehended by the citizen. This phenomenon unveils a discrepancy between the need of transparency and the needed maturity to give an informed consent. Further, it raises the question to what extent an understanding of the context, technology and processes is necessary for citizens to make informed decisions. Within the empirical studies, the sole existence and availability of information and explanations seemed to contribute toward a positive stance toward AI. This positive stance is rooted in the perception of transparency, which, combined with a pre-existing trust in public institutions, leads citizens to accept a certain level of vulnerability stemming from their incomplete understanding of the technology. At the same time, Bayer et al. (2022) reported evidence indicating that an elevated level of expertise and comprehension in a domain can potentially erode trust in AI-driven decision-making systems. This presents a paradox, as transparency is crucial for building trust, but increased understanding might in some cases lead to decreased trust. It becomes therefore a matter for consideration whether increased transparency can help to enhance understanding, or if it stifles trust in the AI system. This makes agreement modelling another critical area where design principles prove invaluable, which led to the formulation of the design principle cDP-6 - Appropriate Consent practices. These mechanisms are designed to facilitate consensus among diverse stakeholders, reflecting a range of perspectives and interests. A recent study by Google DeepMind described the use of an AI model trained to identify common ground in democratic deliberation and facilitate consensus building. By incorporating such mechanisms, AI systems can better adapt to and reflect the values and norms of the communities they serve. This is particularly important in the public sector, where AI systems must often navigate complex regulatory and ethical landscapes. 8.3.2. Diversity & Equality. The purpose of public service organizations is to provide essential functions such as healthcare, education and social welfare to the populace. But the increasing reliance on digital platforms for public service delivery can exclude individuals without digital skills, exacerbating existing inequalities, as these organizations are most likely the sole providers of specific public services (da Costa Alexandre & Pereira, 2023; Drobotowicz et al., 2023). As Junginger (2016) describes, the concept of a market doesn’t exist in the interaction between the governmental organization and citizens. Neither in the form of a target group, which usually allows institutions to provide services only to those who fit a predefined profile, nor as competition, providing a customer with a variety of service providers to choose from. Having public services accessible to everyone, regardless of language skills, technical expertise or administrative literacy is a fundamental requirement for a just and equitable society. In Rousseau’s (1762) point of view, each member of a society puts their person and all their power in common under the supreme direction of the general will, which makes each member an indivisible part of the whole. The contemporary Social Contract theorist John Rawls describes in his Theory of Justice, that a fair equality of opportunity means everyone is having a fair chance to attain them. Therefore, it is not only a functional requirement for public service organizations to serve all members without discrimination or exclusion, but a moral imperative. In the context of the empirical citizen studies, providing clear information to all was specifically aimed by varying information representations regarding data usage, algorithmic transparency and explanations of the governmental mandate. For an HCAI system to be intelligible by everyone, entails accommodating for a diverse audience with differing skills, knowledge and expertise. The prototype provided different types of information with increasing levels of detail to allow participants to find and process the information to a degree they felt comfortable with. This was achieved through a nested information architecture, which allowed interested participants to drill-down into the information until they found a satisfactory level of detail. This informed the formulation of the design principle cDP-3 – Gradual Information Provision. On the other side, to ensure equal access to specific services, AI systems used by public servants should perform robustly, reliably and consistently. A key strength of AI systems in general is their capability to learn and adapt. Commercial applications of AI gain their value from learning the preferences and habits of users, suggesting and predicting next actions. However, in the context of public services, this alleged powerful feature can create inequalities and reinforce biases. The learning and adaptation to caseworker behavior can undermine the responsibility of public services for consistency and impartiality. Furthermore, when AI systems adapt to individual employee behavior, their decision-making process becomes less transparent, making it difficult to explain decisions and determine responsibility for potential errors or biases. To ensure an equal treatment, the design principle eDP-3 - Consistency over Personalization has been formulated. 8.4. Extended SCT framework – A Research Agenda. The study of the empirical findings through a SCT lens has revealed the potential and the need for a more detailed analysis of the different aspects, involved actors and how their relationships change through the use of AI in public services. The following section presents a research agenda, which emerged from combining the different Citizen « Government « Organization relationships (Rousseau, 1762) with the three levels of the SCT-framework. This extended framework provides a structure for deeper analysis of the different relationships on each level of the SCT-framework, creating new, bi-directional ties between the different actors, e.g. between citizens and public service organization on the Meso level (Figure 16). 8.4.1. Macro Level Tie between Citizen « Government. As has been now discussed in detail, the implementation of AI systems within public services possesses great potential for improvements that would benefit multiple stakeholders. But at the same time, the new and advanced capabilities require a closer investigation concerning how this affects the balance of power between its actors, namely citizens and governments. On the Macro Level, new questions regarding the responsibility and accountability when deploying AI systems are emerging. Does the use of advanced AI technologies shift the power dynamics in favor of governments, providing them with more sway and facilitate overreach as the introduction of such systems centralizes power within governmental institutions? Or does it rather empower citizens with a new voice of opportunity to more directly impact and affect democratic processes, and the access to information and services become more democratized? Will a shift in traditional power dynamics require a renegotiation of the Social Contract? A closer inspection of such developments is crucial, since the understanding of the impact can help to ensure that AI adoption enhances democratic values rather than undermining them. AI systems have the potential to either alleviate or exacerbate social inequalities and governments are under the obligation to not only maintain, but also to promote social justice. By examining these mechanisms, researchers can identify strategies and policies that harness AI for equitable outcomes, ensuring that technological advancements benefit all segments of society rather than reinforcing existing disparities. Also from the citizens’ perspective, it is vital to explore how public participation can be fostered so that citizens can influence governmental policies regarding the ethical use of AI. There are various avenues into how citizens can engage with policymakers to contribute to a responsible design and development of AI systems in the public sector. Are existing approaches such as public consultations and citizen panels effective measures, or are there also other participation mechanisms? Through such examination, citizens can be empowered to shape ethical AI governance frameworks and align them with societal values. 8.4.2. Macro Level Tie between Government « Organization. Another emerging tie from the extended SCT-framework is the relationship between a government and its organizations on the Macro level. This includes whether policies and mandates issued by governments towards public service organizations will change with the introduction of AI systems. And if they change, how will these changes pan out in practice? Policies and mandates establish the ethical and regulatory framework of organizations, but also the radius of operation of public service organization. On the one side, it is crucial to analyze how governmental oversight can influence operational aspects of providing public services, for instance for an equitable distribution of such AI-enabled services. By investigating this, potential systemic issues which may lead to unequal service provision can be identified and inform strategies to promote fairness and inclusivity in AI deployment. At the same time, exploring how existing power dynamics between the legislative character of a government, and its executive body of a public service organization can influence AI deployment is also important. It can help to understand friction points such as resistance from stakeholders or misalignment of organizational goals. Uncovering such obstacles can have important effects on how AI projects are prioritized, funded, and managed. 8.4.3. Macro Level Tie between Organization « Citizen. The third tie on the Macro level is between the public service organization and the citizenry, which is also the original scope of the Macro level in the SCT-framework. But this relationship contains many more highly relevant questions that should be investigated. For instance, AI systems have the potential to bridge or widen existing societal gaps. There has been a first attempt to address aspects of diversity and equality in the defined design principles based on the empirical findings. However, an equitable and fair treatment of citizens is multifaceted and can benefit from further investigation to ensure inclusive and unbiased AI implementations. Biases are a common and widely acknowledged issue of AI systems. Investigating biases in the specific context of the relationship between public sector organizations and citizens on the Macro level, and potential measures to address them to avoid discriminatory outcomes is a much-needed research avenue. This also includes ways of participation and involvement of citizens in the process of sourcing and preparing AI training data. Furthermore, exploring mechanisms and ways to incorporate citizen feedback during the conceptual phase of designing AI services is vital for fostering participatory governance. 8.4.4. Meso Level Tie between Citizen « Government. On the Meso level, there are new, more nuanced ties emerging which prompt further investigation. In the understanding of the Social Contract, governments and citizens both have rights and obligations towards each other. But further elaboration is needed to understand what new rights and duties are emerging due to the use of AI systems in the public sector and how do they affect different stakeholder values. For instance, how does a government protect individual privacy when training or deploying AI systems? And how is this privacy weighed against societal benefits? Another value that could be examined is the value of security. Also, which options exist for citizens to voice concerns about AI applications that affect them? Such single-value investigations could help to better understand the challenges of balancing personal interests with societal ones. Further, they help identify whether current mechanisms are adequate or if new are needed. Such more differentiated analysis can help to further unpack the intricate process of balancing individual and societal interests. 8.4.5. Meso Level Tie between Government « Organization. A deliberation process on the Meso level is happening between public organizations and governments when balancing ethical responsibility and operational efficiency. Organizations face pressure to deliver services quickly and cost-effectively but are expected to uphold ethical standards at the same time. This potential tension requires particular attention in the light of AI use, which bears the promise of improved efficiency, but also necessitates a re-evaluation of risks and misconduct. Understanding how public service organizations interpret governmental guidelines on AI ethics is a necessary research avenue as it directly influences how ethical principles are embedded into the AI systems. This also involves the exploration of how governments motivate and incentivize the enforcements of such standards. Are existing incentive structures, such as recognition programs, compliance benefits and funding sufficient, or does the introduction of AI systems require novel approaches? Research can inform policy as well as operations to ensure that conferred responsibilities are performed and followed, rather than being compromised in the pursuit of efficiency or sacrificed by pressure of austerity. This also necessitates a deeper consideration of frontline workers, who will engage with AI systems in their day-to-day work. How much AI literacy is appropriate, or required to ensure an informed usage of such systems? Are the AI systems perceived as helpful tools to augment employee capacities, or will time-pressure entice public sector employees to neglect necessary due diligence when handling individual cases? 8.4.6. Meso Level Tie between Organization « Citizen. Breaking down the Meso level of the SCT framework further, another essential tie to examine is between the governmental organization and citizens. One of the core considerations when designing, developing and deploying AI system in the public sector, is the balancing act of individual rights and freedom, both when considered next to other individual cases as well as the wider public. Reaching a better understanding of this weighing process is crucial for a equitable and fair use of AI technologies. It can help to prevent scenarios where AI services cater only to individual preferences, potentially neglecting the needs of the whole society. At the same time, the use of AI systems should not infringe on any personal rights and freedoms individuals within a society have, breaking the Social Contract. Citizen influence and participation relate to this balancing process. The design principles of this doctoral thesis already give guidance on pathways for raising concerns and providing feedback (cDP-5/eDP-5 – Voluntary Feedback Mechanisms). But such feedback mechanisms are reactive, as they are anchored in the context of an existing system. The bedrock of Human-Centered AI is the early involvement of different stakeholders in the process, which necessitates the exploration of other mechanisms that enable citizen participation. 8.4.7. Micro Level Tie between Citizen « Government. Exploring the specific tie between citizens and governments on the Micro level revisits agreement modelling in a conceptual sense. One research avenue is the analysis of how citizens consent to the use of their personal data in the context of AI systems within the public sector. This touches upon the use of individual personal data for actual operations processes, (e.g. service provision), but also for training purposes. It also opens questions whether personal data loses the property of being personal during a process of anonymization, and how does this affect the need for explicit citizen consent? Does data ownership shift or even dissolve at a certain point and does personal data becomes a communal commodity? But there are also other agreement modelling concepts that could be valuable to be used within decision making processes. SCT posits that the consent of individuals, either implicit or explicit, to accept certain rules for mutual benefit, forms a society. Harsanyi’s aggregation theorem (1955) aligns with this understanding by showing that rational individuals would endorse a system where social welfare is determined by aggregating individual utilities. Could such a calculative approach be leveraged to calculate societal benefit when considering the use of AI systems? Furthermore, are there any bargaining processes that could allow citizens to contribute to the negotiation of AI policy terms, such as public consultations, participatory policymaking platforms, or citizen assemblies? Understanding when consent is required, how it is negotiated, obtained and maintained, whether through explicit or implied agreement, or via legislative mandates, can inform a responsible data handling. 8.4.8. Micro Level Tie between Government « Organization. An extended SCT framework also spawns a tie between governments and public service organizations on the Micro level. In this relationship, there are negotiations processes taking place concerning the goals and constraints of introducing AI systems, that public sector organizations are going to use. This provides a highly relevant research direction to understand such negotiation process, as they shape the direction, scope, and objectives of AI initiatives within the public sector. Gaining an understanding of this interaction can help to reveal how strategic priorities are aligned between policymakers and implementers. Related to the negotiation process, this also opens the debate on how performance, impact and success of such AI initiatives can be measured? Performance metrics will be defined as well as the necessary strategies and operations to benchmark them. Examining how these metrics are negotiated helps ensure that they are comprehensive, transparent, and aligned with both governmental objectives and public expectations. Another aspect worth investigating in the relationship between governments and public organizations is how different such organizations cooperate, particularly on data sharing. Collaboration and data exchange between organizations can enhance the efficiency and effectiveness of AI applications even further but raises new questions on which data can be shared, and which cannot. On the one side, investigating the mechanisms of inter-agency cooperation and identifying challenges such as data silos and technical incompatibilities is important to enable seamless data sharing. But this also needs to take concerns about privacy and governmental overreach into account. An issue that concerns the relationships between public sector organizations and citizens, as well as between governments and its public service organizations on the Micro level. 8.4.9. Micro Level Tie between Organization « Citizen. Inter-organizational data-sharing is a concern in the tie between public sector organizations and citizens of the extended SCT-framework. Consent to the use of personal data with one public service organization doesn’t necessarily extend to other organizations. Therefore, it is necessary to investigate when and how consent is required for sharing citizen data between different public service organizations. Generally, explorations of agreement modelling processes related to interfaces and interactions are also required. The defined design principle cDP-6 – Appropriate Consent Practices addresses the need for appropriate consent mechanisms but doesn’t specify how these mechanisms could look like. 8.5. Summary of the Chapter. This chapter examines the design principles in relation to prior literature and within the conceptual boundary conditions of the SCT framework. It discussed the identified concepts and empirical findings and how the defined Design Principles are addressing them. Additionally, the chapter outlines an extensive research agenda through the proposal of an Extended SCT framework. The suggested framework aims to further break down the relationships among the actors on the different SCT framework levels. 9. Contributions. In the process of this doctoral research, multiple contributions have been produced, both to practice and to theory. The practical contributions of this doctoral research are three-fold. First, HCAI Design Principles for public services are formulated which can be used by practitioners when designing AI systems in the public sector context. Second, an innovative artifact has been created, which provides the collaborating organization with a concrete instantiation of human-centered design for communicating the use of AI with citizens. Thirdly, the thesis contributes to practice by providing insights into the perspectives of different stakeholder groups regarding the use of AI in public services. This thesis also provides theoretical contributions, including the advancement of the HCAI discourse by providing conceptual clarity on Human-Centered AI, introducing Social Contract Theory as a novel theoretical approach for AI in public services as well as methodological contributions. 9.1. Contributions to Practice. Practical impact is of utmost importance and a crucial contribution of this research endeavor. The defined design principles aim to be useful for practitioners working on introducing AI in public sector contexts. They should be understood as an additional set of considerations, supporting practitioners when designing, developing, or deploying AI systems used within the public sector. The design principles emerged from the theoretical foundation of SCT to operationalize practical design deliberations considering the relationships between citizens, public servants, and public service organizations. Applying these design principles within the context of public sector organizations won’t alone ensure a completely beneficial and risk-free application of AI systems, but it can contribute to more responsible AI strategies and implementations within the public sector. To uphold practical relevance, the design principles draw from established and proven guidelines, namely Google’s PAIR guidelines. These guidelines provided a comprehensive foundation and starting point for the definition of more context-specific design principles. Furthermore, the extended set of design principles for the public sector context goes beyond the commercial interests of profit-oriented companies by incorporating the critical considerations of citizens and public servants. The identified amendments can contribute to the responsible design, development, and deployment of AI systems in the public service context. Another contribution of this research project is made through an innovative artifact in the form of an interactive prototype. The innovative artifact was reciprocally shaped in multiple iterative cycles. Iterations included empirical studies with stakeholders as well as close coordination with NAV to ensure user utility by addressing a real-world problem in the organizational context. The artifact presents concrete proposals for solutions for the example use case of the sick leave duration prediction system. It includes suggestions on how to communicate value propositions in the public sector context, how to provide procedural transparency, and on how to present the inner workings of an AI system and data usage. Furthermore, it also provides design solutions for consent and feedback mechanisms. These proposed solutions have been included as tangible implementation examples for the more abstract Design Principles, which can help NAV to create human-centered AI interfaces, that provide the necessary level of explanations, transparency, and participation. Furthermore, this thesis contributes to practice by providing insights into how important it is for public organizations to ensure that public goodwill is not eroded. This can happen if public sector projects overstep the boundaries of introducing AI while preserving social functions. The erosion of people's goodwill may limit the ability of public organizations to deliver their services effectively in the future. By identifying the factors that contribute to positive attitudes toward AI, practitioners are provided with actionable guidance on the responsible and sustainable design and development of AI systems in public service contexts. The findings expand extant research on AI adoption in the public sector by foregrounding the government-citizen relationship that has received limited attention in this body of literature. The overall positive attitudes of citizens in this study align with prior research which shows that although AI can be a source of public anxiety, informing citizens and workers about the characteristics of the AI system and of the humans’ involvement in decisions has a significant positive influence on the public's attitude towards AI which is important for its adoption. However, by being able to probe participants to share their reflections, this doctoral research project goes beyond prior survey-based research, providing insights into the reasoning behind their attitudes. The information provided to citizens and workers fortifies their trust in the government. 9.2. Contributions to Theory. The first theoretical contribution concerns the conceptual understanding of Human-Centered AI within the literature. Within this doctoral thesis, a comprehensive definition of the concept of HCAI has been proposed. The definition aims to emphasize the fundamentally humane character of HCAI while also encompassing its contributing constituents. The provided HCAI definition underscores the constant consideration of whether an envisioned AI system is in accordance with human needs and values. Further, the definition highlights context sensitivity, including the acknowledgment of stakeholder diversity, a comprehension of the context of use, and the awareness that an AI system is not an isolated entity, but rather a part of a larger structure. Understanding the characteristics of an AI system, including scope, usage implications, and sociocultural context is crucial for HCAI. Finally, the definition addresses the consideration of ethical and legal requirements to ensure responsible deployment of AI systems. In essence, the definition delineates the overarching objective of HCAI to consider and promote the well-being of individuals as well as the whole of society. Furthermore, this doctoral work provides a contribution towards conceptual clarity of the constituent elements of HCAI. The work outlines its key elements and points out the fundamental role of Human Centered Design methodological approaches for HCAI. Established Human Centered Design methods do offer a constructive starting point for navigating the complex issues posed by AI. With deliberate refinement and expansion, these methods can guide the design of AI systems that benefit humanity by strengthening rather than eroding what makes us human. Moving forward, the outlined research agenda points to major themes warranting further investigation to advance knowledge on creating AI systems that genuinely benefit humanity. The doctoral thesis contributes to the conceptual discourse by introducing Social Contract Theory as a novel lens to the contemporary topic of responsible and human-centered AI development specifically for the public sector. SCT recognizes different relationships between governments, their institutions, and citizens compared to the relationship between a commercial company and its customers and therefore provides a novel and relevant theoretical approach. Within the work of this thesis, a novel SCT framework as a theoretical approach for researching socially responsible AI in the public sector is proposed. This framework can help analyze AI in public services at three different levels: Macro, examining the power relationships among governmental organizations and citizens; Meso, investigating the gauging between the collective good and individual benefit; and Micro, exploring different agreement modeling mechanisms. Building on the initial SCT framework, an extended SCT framework (Figure 16), merging the traditional SCT relationship model (Figure 4) with the three defined levels of the initial SCT framework (Figure 4) is proposed. This extended SCT framework provided the structural input for a research agenda for a more detailed analysis of the different aspects, involved actors, and how their relationships change with the introduction of AI in public services. Finally, this doctoral thesis also offers methodological contributions. Utilizing an existing set of HCAI guidelines from a different context as a starting point, demonstrates that generalized outcomes need not always originate from a blank page. Instead, they can be developed by revisiting established principles and recommendations, and then refining or adapting them based on insights gained through the ADR process. This approach enables the integration of best practices while tailoring them to address specific needs. Secondly, the two-strand ADR approach employed in this study represents a novel methodological contribution to the Building, Intervention, and Evaluation cycles of Action Design Research. By initially separating the research strands based on stakeholder groups, this approach enabled a more nuanced investigation and offered flexibility to adapt the research design in response to shifting priorities within the collaborating organization, ensuring sustained practical relevance. Furthermore, the integration of a foresight-driven approach allowed exploring and shaping uncertain futures while emphasizing participatory engagement. This facilitated the elicitation of nuanced and context-specific insights from the participants helping to formulate relevant HCAI Design Principles. 10. Conclusion & Limitations. When change is easy, the need for it cannot be foreseen; when the need for change is apparent, change has become expensive, difficult and time consuming. The situation described in Collingridge’s Dilemma is a key quandary in the area of responsible technology development and governance. During the early stages of a technological advancement, the potential as well as consequences of the new technology are often not fully understood. Therefore, the need for efforts to address these aspects is not considered to be worthwhile or important. Over time, the effects of the technology on individuals and societies become more apparent. At that point however, governing and regulating the technology becomes more challenging due to established approaches, applications, and structures. Consequently, changes become more difficult, more time-consuming, and more expensive. Human-centered and context-specific Design Principles can facilitate necessary critical thinking about the short- and long-term effects already in the early stages of the AI system lifecycle. The goal of this doctoral thesis was to define such HCAI Design Principles for the specific context of the public sector. 10.1. Summary. In the context of public services, and especially with the rise of ever newer AI technologies, understanding responsible technology use has a particularly large scope of relevance. Deploying AI systems in public service creates the opportunity for governmental organizations to generate collective benefits for the society by enhancing services, increasing efficiency, and improving processes. However, while aiming to enhance efficiency and service, governmental organizations also face the responsibility of mitigating the risks of algorithmic harms, such as bias and lack of transparency. Incorporating different stakeholders’ perspectives in shaping the future is the bedrock of human-centric development of technology. Applying this understanding as a general approach for AI systems can lead to better policymaking, democratic processes, and the creation of constituencies for long-term change. The empirical studies of this doctoral thesis brought forward the perspectives, expectations and needs of key actors within the social contract relationships underlying public services, namely citizens and public servants. In the citizen-studies, a generally positive attitude toward the use of AI was found. This is linked to a high level of trust into governmental organizations, expectations for a balanced value proposition between individual and collective benefits, and assurance for human oversight. Within the employee-studies, the participants were optimistic, yet careful and prudent. They acknowledge the vast possibility space for AI bringing improvements but also emphasized that AI shouldn’t fully replace humans as public services require empathy, discretion and personal connections. These multi-stakeholder insights have been translated into actionable HCAI Design Principles based on a SCT understanding of the relationship between citizens and public servants. By considering the intricate web of relationships between the diverse roles of stakeholders, their needs and technology features, the defined HCAI Design Principles may not only bridge existing gaps in the intellectual discourse but also support practitioners in their pursuit to develop AI systems that hold a genuine commitment to foster public trust and welfare. The defined Design Principles can guide the AI development ensuring that the systems serve not just individual users but also contribute positively to the community and society at large. This balance is crucial in public sector applications, where the impact of AI can be profound and far-reaching. 10.2. Limitations. Despite every intention for comprehensive, diligent and meticulous research, this study has its limitations, which need to be taken into account when looking at this work as a whole. A first limitation is caused by the nature of many empirical studies that are researching frontier technologies. Any investigation of the latest technological advancements, especially with the aim of creating prescriptive knowledge, can easily feel like chasing a moving target. This study is no exception. This work has its beginnings in a time prior the widespread adoption and tremendous success of Generative AI. The overarching AI4Users project started in 2020, with this specific doctoral research beginning at the end of 2021. The breakthrough of Large-Language Models (LLMs) such as OpenAI’s GPT, Anthropic’s Claude, or Google’s Gemini, represents a hallmark moment in technological advancement. Some scholars, journalists and pundits even comparing it to the transformative effects of the printing press or the advent of the internet. But the focus of the research conducted in this thesis lies mainly on AI systems for predictions and classifications (discriminative AI) not on Generative AI. Also, the ethical implications of AI in public services are dynamic and continually evolving. The thesis addresses several ethical considerations, such as transparency, accountability, and bias, but it may not fully capture emerging ethical dilemmas associated with the latest AI technologies. Another limitation can be found in SCT as the theoretical lens applied for this research. Feminist- and race-conscious philosophers have highlighted how power and inequalities are also enacted without the agreement of those affected. Another aspect that highlights the lack of universal applicability of SCT is it origin in the western enlightenment period, therefore primarily reflecting European societal structures and values. Critics argue that it may not adequately address or represent non-Western cultures and societies, making it less applicable in a global context. And finally, in SCT and especially in Rousseau’s (1762) understanding of it, individuals are described as fully rational actors capable of understanding and consenting to the social contract. This oversimplifies the complexities of human behavior, cognitive biases, and social influences that affect decision-making. A shortcoming that is addressed in part by grounding the conceptual development on a human-centered design approach. Another limitation of the thesis relates to the scope of the empirical investigations. The research focuses on the Norwegian context and reflects the needs and perspectives of citizens and employees in Norway. Human-centered design is always contextual, influenced by cultural, social, and political factors that may vary significantly from one region to another. Investigating differing perspectives across nations could provide the basis for adaptations to local nuances. The principles were developed through iterative engagement with the partner organization undergoing evaluation cycles with both practitioners and scholars. Further evaluation would be beneficial, and this could also include evaluations outside Norway. The structure suggested by Gregor et al. (2020) for formulating Design Principles also grants various evaluation options for researchers. Appendix. Publication #7 Title Human-Centered AI Design Principles for the Public Service Domain: An Action Design Research Study Authors Stefan Schmager, Ilias O. Pappas and Polyxeni Vassilakopoulou Addressed Research Question RQ: What are Human-Centered AI Design Principles for public services? Publication Outlet 58th Hawaii International Conference on System Sciences (HICSS), 2025 Human-Centered AI Design Principles for the Public Service Domain: An Action Design Research Study Stefan Schmager University of Agder the email address Ilias Pappas University of Agder the email address Abstract. Most existing design and development guidelines for Human-Centered AI primarily cater to a commercial context, they are not tailored to the specific needs of public services. This paper presents an Action Design Research study proposing public service-specific design principles for Human-Centered AI. The design principles are informed by multiple iterations of empirical research with citizens and public service employees acknowledging the multi-stakeholder nature of Human-Centered AI. The study recognizes an evolving understanding towards prioritizing human values and well-being in AI technologies. Furthermore, it considers the relationship between technology features in the public sector and citizens' needs to ensure that AI systems are developed with a commitment to fostering public trust and welfare. The study contributes to theory and practice by advancing the scholarly discourse on Human-Centered AI and informing AI strategies and implementations within the public sector. 1. Introduction. Recently, the discourse on artificial intelligence (AI) has increasingly emphasized concepts such as Human-Centered, Responsible, and Trustworthy AI. They advocate for a paradigm shift in the approach to technology, emphasizing the need to prioritize human values and well-being in the development and deployment of AI-enabled systems. This shift is not solely theoretical but is intended to influence how AI technologies are designed, developed, and implemented, to ensure a genuine benefit for humanity. Polyxeni Vassilakopoulou University of Agder the email address In the public sector, where AI systems can significantly influence societal welfare, the role of providing concrete guidance for the design, development, and implementation of AI becomes even more critical. The requirements for transparency, accountability, and inclusivity are more stringent, and the impact on civic life is more direct. As noted by Beckert (2021), the evolution of digital welfare states necessitates a robust set of tools, methods, and principles that are specifically crafted to meet the needs of the public sector. Design principles can provide clear, actionable guidelines that help navigate the complexities involved in the ethical and effective implementation of AI technologies. Employing robust design principles for AI helps to ensure that these systems are developed with a keen awareness of their broader impacts. Human-Centered AI (HCAI) provides a responsible approach toward AI design that can enhance and safeguard human welfare, emphasizing AI systems that respect and amplify human values and capabilities. HCAI foregrounds the engagement with a range of stakeholders to align AI technologies with the multifaceted spectrum of human needs and societal values. Many frameworks and guidelines have been established to promote the Responsible and Human-Centered development of AI systems; however, they predominantly cater to commercial entities and products. This commercial focus leaves a significant gap for guidance tailored to public services, where the stakes are uniquely high. This leads to the Research Question: What are public service-specific Design Principles for Human-Centered AI? The paper suggests design principles informed by human-centered empirical research with citizens and public service employees, aligning with the multi-stakeholder nature of HCAI. This alignment ensures that AI systems in the public sector are developed with a clear understanding of objectives, application contexts, and rationales. By considering the diverse roles of stakeholders and the intricate relationship between technology features and user needs, design principles may not only bridge the existing gap but also ensure that AI technologies are developed with a genuine commitment to fostering public trust and welfare. The remainder of this paper is structured as follows. First, the concepts of Human-Centered AI, Human-AI Interaction as well as Design Principles are described. Next, the theoretical lens of Social Contract Theory is introduced, followed by an overview of the organizational context and the methodological approach of Action Design Research. Finally, the specific Design Principles for Human-Centered AI in public services are presented and their contribution and limitations are discussed. 2. Related Literature. 2.1. Human-Centered AI. Digital technologies - such as AI - can either exacerbate existing challenges or provide innovative solutions, depending on how they are integrated into societal frameworks. A human-centered approach to AI is not merely about integrating AI into the daily lives of human beings but doing so in a manner that respects and enhances the human condition. HCAI aims to enhance and safeguard human welfare, emphasizing the integration of AI systems that respect human values and amplify their abilities. The foundation of HCAI is deeply rooted in Human-Centered Design (HCD) principles, highlighting that engagement with a broad range of stakeholders during the AI development process is crucial. Riedl (2019) notes that involving stakeholders helps to align AI technologies with the multifaceted spectrum of human needs and societal values. Similarly, Schmager et al. (2023) highlight that HCAI is committed to understanding and incorporating human purposes, values, and the desired properties of AI systems through HCD methodologies. Especially in the context of the public sector, HCAI not only addresses the augmentation of human capacities but can also contribute to the ethical governance of AI applications, balancing technological advancements with societal norms and individual rights. 2.2. Human-AI Interaction. Many recent research initiatives exploring how AI systems are integrated within organizational frameworks focus specifically on human-AI interaction. While these interactions form the foundational layer of how humans and AI systems interoperate, they represent only the initial step towards a more integrated and cooperative future. Basic interaction with AI systems involves exchanges and operational commands, but true synergies encompass a more cooperative relationship. This implies a mutual effort where both humans and AI systems work towards common objectives, leveraging their unique strengths. This shift from mere interaction to substantive cooperation redefines the role of AI in the workplace from a tool of automation to a partner in decision-making. As AI systems become increasingly sophisticated, the critical question shifts from whether AI will replace human roles to how AI can augment human capabilities. Hence, within the design process, AI systems need to be understood not as replacements but as enhancements that work alongside human intelligence. Such understanding enables a workplace where AI systems and humans bring together their respective strengths – humans offering contextual insights, creativity, and ethical judgment, while AI contributes computational power, data processing, and pattern recognition capabilities. By harnessing these complementary abilities, organizations can achieve higher levels of efficiency, accuracy, and innovation. 2.3. Design Principles. Design principles act as intermediaries, bridging the gap between theoretical knowledge and its practical application across various disciplines such as engineering, economics, and information systems. As noted by Gregor et al. (2020), there is a growing need within the Information Systems field to transition from merely explaining and predicting phenomena to prescribing actionable solutions that deliver tangible societal benefits. This reflects a broader academic shift, advocating for knowledge that is not only theoretical but also practically applicable (Bichler et al.‚ 2015; Seidel & Watson, 2014). Design principles can guide the implementation of AI systems ensuring that systems are developed with a clear understanding of their social implications, equipped to handle complex ethical dilemmas, and capable of delivering balanced benefits to both individuals and society as a whole. As AI technologies continue to evolve and become more integrated into public sector operations as well, design principles can contribute to advancements that are innovative, inclusive, and beneficial to society. 3. Theoretical Lens: Social Contract Theory. Within this research, Social Contract Theory (SCT) is used as the theoretical framework to examine the power dynamics and relationships essential for implementing HCAI in public services. Social Contract Theory, a broad spectrum of ideas rather than a singular theory, posits a hypothetical agreement where society agrees to abide by certain rules and principles for mutual benefit. This agreement grants the state authority in exchange for security and services only a sovereign can provide. SCT underscores the unique role of governmental organizations, which are tasked with mediating between the state and citizens, thus enhancing societal welfare through service delivery and policy enforcement. D’Agostino (1996) emphasizes that the social contract not only confirms the legitimacy of governmental power but also mandates that rulers act in the interests of the governed, establishing reciprocal rights and obligations. This introduces a myriad of new questions in relation to the mandate of using private and public data, the balancing act of individual and collective interests as well as the details of how agreements are made within the structures of the social contract. To analyze these affairs the SCT framework defined by Schmager et al. (2024) is used as the analytical tool, which conceptualizes SCT into three lenses – Macro, Meso, and Micro. The Macro-Lens of the SCT framework analyzes the critical role of power dynamics, social justice, and equity within AI applications in public services. Power, derived from public legitimacy under the social contract, involves decision-making capabilities from the organizational side, that significantly impact citizens. An HCAI approach to the use of AI systems in this sociotechnical construct must, therefore, include mechanisms for human oversight, ways to appeal, and rights to contest decisions, which reinforce citizens' rights to challenge governmental actions and ensure adherence to principles of social justice. The Meso-Lens considers the weighing between individual interests and collective good, focusing on the negotiation between personal rights and societal benefits. A particular focus of the Meso Lens lies in the evaluation of how the promotion of personal freedoms might affect societal cohesion versus fostering independence. Governments must navigate these differences when deploying AI systems, deciding whether to prioritize individual rights or the “common good”. This approach highlights that societal and public morality should guide AI applications in public services, emphasizing a shift from personal obligations to a broader ethical perspective. The Micro-Lens analyzes various agreement mechanisms within a Social Contract Theory approach, including Consent, Bargaining, Aggregation, and Equilibrium. These mechanisms describe how individual citizens as well as public service organizations exercise normative power to bind themselves to obligations within a social contract. 4. Context & Method. 4.1. Organizational Context. This research is performed in collaboration with a major organization within Norwegian public administration. In Norway, the government actively promotes the use of AI in public services with the goal of leading toward human-friendly and trustworthy solutions. The collaborating organization dedicates internal resources to exploring data analytics and the use of AI systems to enhance service delivery, aiming for greater efficiency, robustness, and responsibility. They recognize their special responsibility to design and develop AI systems with a Human-Centered approach by carefully considering citizen perspectives, as well as employee needs before deploying AI. As one of the initiatives, they developed a predictive model to estimate the duration of citizen sick leaves from work, intended to assist case handlers in prioritizing their efforts for more effective "user-adapted follow-up” (Figure 1). By the time this research began, the sick-leave prediction system had already been internally developed involving caseworkers as the main users of this system, although it was not yet deployed. However, a citizen-focused interface for the sick-leave prediction system has neither been designed, developed, or deployed. 4.2. Action Design Research. The overall research project employs an Action Design Research (ADR) approach, providing the unique opportunity to facilitate continuous engagement with the various involved stakeholders (Figure 2). The research started in January 2022 and went through iterative cycles of collaboration with the public service organization. These cycles include constant reflection on the problem formulation, the creation and evolution of prototypes, and feedback from the empirical research. The approach recognizes the importance of including the people concerned, i.e. citizens as well as the employees of the collaborating public service organization (Table 1). ADR enables direct participation, and collaboration and acknowledges the unique challenges the public service organization must face as avenues for generating general knowledge. By this, design is understood as a context-specific, multi-stage process within an organization as well as a reflective endeavor that yields design principles for solving specific problems. 4.2.1. Problem Formulation. In the first ADR stage, the problem is defined as the catalyst for the research undertaking. It draws inspiration from practice but also emphasizes that the ensemble artifacts creation and evaluation is informed by theory. This problem identification is grounded in practical concerns identified together with the collaborating organization, while also being informed by the Social Contract Theory framework, as well as Google's PAIR guidelines. These established industry guidelines are considered the most comprehensive and balanced best practices when designing AI systems. The intersection of the practical challenge of designing, developing, and deploying an HCAI system within the public service context, with the theoretical foundations of Social Contract Theory creates a fertile ground for innovation, and the creation of ensemble artifacts as well as contributing to wider knowledge in the field. The research team had regular and iterative consultations with the public service organization. These coordination meetings facilitate a constant exchange of insights and warrant the project remains rooted in the actual needs of the organization. In addition, conceptualizing the practical problem through the theoretical lens of SCT ensures that the research addresses not only the practicalities of the issue but also places it in broader societal context. Within this collaboration, we identified the need for two complementary research strands addressing a dual problem formulated in collaboration with the public organization. To examine the perspective of citizens, we identified the problems of transparency, personal data usage, and ways of public participation. Since no citizen-facing interface had yet been designed, we created a tangible artifact to demystify the abstract notion of AI systems in public services. On the other hand, the employee-focused research strand builds upon existing knowledge and familiarity with different forms of AI systems. This allowed the research to focus on the perceptions, and expectations of caseworkers on the use of AI. Together, both research strands provide a holistic view of the ecosystem, encompassing both immediate user feedback and strategic organizational planning (Figure 2). This close interaction between the project partners fostered the creation and rigorous evaluation of the research strategy as well as the artifacts developed during the ADR process. This strategy is articulated through and embedded within iterative cycles of building, intervention, and evaluation. 4.2.2. Building, Intervention, and Evaluation. In the second stage of the ADR methodology, the initial creation of the artifact is informed by the defined problems and theoretical lenses. This stage unfolds through iterative cycles of building, intervention, and evaluation (BIE), which are fundamental for the continuous adaptation and enhancement of the artifact in partnership with the collaborating organization. Each cycle begins with the building phase where the artifact is designed or revised based on the insights gained from the previous iterations and grounded in the theoretical underpinnings established during the problem formulation stage. For the citizen-focused research strand of the study, we created an interactive prototype in close coordination with the collaborating public service organization, which was also informed by Google’s PAIR guidelines. The public service organization provided the example use-case of a sick-leave prediction model, described in the Organizational Context. The prototype depicts a predefined interaction sequence starting from a notification about the optional use of an AI-based prediction, different types and levels of information about the prediction and consent options. Tangible ensemble artifacts provide beneficial traits like shareability within the organization to collect broader feedback, the exploration of a variety of versions as well as a changelog to track progress with the ability to roll back where necessary. After the building phase, an intervention occurs where the artifact is tested within the actual use context. This step is crucial for observing the artifact in use and understanding its impact in a real-world setting. It provides a practical test bed for the artifact’s functionality and alignment with the organizational needs and user expectations. To examine citizen perspectives, the intervention took place in scenario-based test sessions, in which participants interacted with the artifact. Over the course of multiple iterations, 40 semi-structured interviews including prototype interactions were conducted. All participants were between 18 and 65 years old, and the sample reflects gender distribution according to the general population on sick leave based on the official Norwegian statistics (Statistisk sentralbyrå, 2022). The sessions captured in-depth insights into the users' experiences and expectations. Data collection within the sessions happened in three stages. Initially, basic participant data such as age, gender, current occupation, and highest educational level was gathered. Participants were also asked about their prior knowledge about AI and their frequency of technology use as well as their trust in the Norwegian government and its use of AI in public services. These initial inquiries aimed to gauge the state of government-citizen relationships and perceptions of the balance between individual and collective benefits from public services, applying the different levels of the SCT framework as a guiding framework for inquiry. In the second stage, participants engaged in a moderated, scenario-guided user study involving a task-based interaction with the prototype, that prompted a decision on whether or not they would consent to the use of personal data for the sick-leave prediction system. During these interactions, participants were instructed to “think aloud” and share their observations, understandings, and reasonings. The final stage of data collection consisted of follow-up questions about their interactions with the prototype. Participants re-evaluated their views through a set of questions concerning the AI system’s competence, efficiency, potential negative impacts, and understandability. We also inquired about their comfort and trust in the system and gathered open-ended responses on the prototype's explanatory elements. The evaluation phase subsequently assesses the effectiveness of the artifact, and is crucial for informing the next BIE cycle, ensuring that the artifact not only meets the initial design criteria but also evolves in response to the insights identified during its evaluation. We conducted a detailed analysis of qualitative data from the think-aloud protocol and responses to open-ended questions. This involved multiple reviews of study recordings and interview transcripts to identify patterns, perceptions, and recurring themes, highlighting key incidents and quotes for in-depth group discussions and interpretations. Finally, we analyzed interactions with the prototype, noting which information elements were accessed, engagement levels, and the time participants spent on specific tasks. By integrating these findings with the findings from previous stages, we aimed to strengthen our analysis and identify emerging concepts and recurrent themes. This process of "guided emergence" embodies the dynamic interaction between the artifact and its stakeholders, capturing the underlying tensions and enabling a reciprocal shaping of the artifact as it evolves. Simultaneously during the BIE cycles, the evolution of the prototype also informed the formulation of the emerging design principles. Described as “Reciprocal Shaping”, it addresses the relationship between the development of the specific prototype and the emerging design principles influenced by the organizational context. The employee-focused research strand employed Foresight methods to facilitate employee participation and engagement in envisaging and shaping potential AI-driven futures. Due to the advanced internal design and development of the sick-leave prediction system and other AI applications, the research design focused on exploring perceptions, and future expectations of AI applications among employees. This approach aims to highlight the complexity and uncertainty of future scenarios and aims to understand and anticipate potential challenges and opportunities from the employees' perspectives. In the employee-focused study, a total of 19 caseworkers from three regional offices participated in interviews conducted between November 2022 and March 2023, varying from digital group to in-person individual formats. The interviews were structured into three parts: demographic and professional background, current work processes, and potential future AI applications. Initially, participants provided information on their age, job title, duration of employment, and AI knowledge, which set the context for subsequent data analysis. The interview then explored their daily tasks and interactions, and their views on AI's potential utility in these processes. Lastly, through a foresight approach, participants discussed hypothetical scenarios involving AI decision support systems, such as the sick leave duration prediction system. These scenarios prompted discussions on best-case, worst-case, probable, and preferred futures regarding AI in their roles. A thematic analysis helped identify and organize emerging themes from the interview data, providing detailed insights into the caseworkers' aspirations and concerns regarding AI. This analysis method allowed for a flexible yet systematic exploration of data, framing the narrative around the research aims. The SCT framework was utilized as a unifying framework for the classification of the findings from both research strands. The framework allowed us to assess power dynamics, the balance of collective versus individual interests, and participant perceptions of the employed agreement mechanisms from a citizen, as well as employee perspective 4.2.3. Reflection and Learning During the third stage. of the ADR methodology, conscious and deliberate reflection on the problem understanding, the applied theoretical frameworks, and the development of the emerging artifact take place. This is a critical process for assessing the potential generalizations and contributions to knowledge that can arise from the research. The goal of this stage is to transition conceptually from constructing a solution tailored to a specific instance to generalizing this learning to address a broader class of problems. The evaluated artifact of a sick-leave prediction system depicts a representative instance within the broader category of "Human-Centered AI Design in Public Services." This perspective is not limited to the singular application but extends to understanding how an HCAI approach can inform and improve public service delivery more generally. The process of broadening the learning involves rigorous analysis of how the artifact was shaped through its use within the organization, including the examination of the perspectives of citizens, the employees of the public service organization, as well as the input of the project stakeholders involved in the iterative refinement of the solution. 4.2.4. Formalization of Learning In the fourth and. final stage of ADR, the focus shifts to the systematization and articulation of the knowledge and insights gained throughout the project. This stage aims to elevate the situated learning derived from the specific case of sick-leave prediction into generalizable insights that can be applied to a broader class of field problems. This stage involves the documentation of achievements and advancements coming from the research findings of the studies with citizens and public service employees, alongside detailing the consequential organizational outcomes. These outcomes are not merely anecdotal but are systematically categorized and defined as design principles. Translating practical insights into broader design principles involves a rigorous methodological approach. It requires critical engagement in a detailed analysis of how the artifact addressed the specific challenges encountered during the intervention stages and reflecting on the scalability and applicability of the solution across different contexts. The formalization of the design principles represents the result of a dynamic collaboration with the public service organization. The findings from both research strands were systematically analyzed through the SCT framework and evaluated and refined in close, iterative coordination and collaboration. 6. Design Principles for Human-Centered. AI in Public Services This paper proposes design principles based on the situated learning achieved throughout the ADR project, bringing the perspectives of citizens as well as public sector employees together. We seek to transform the specific findings into broadly applicable guidelines that can address similar challenges across contexts. The defined design principles were not conceived in isolation but are the result of a rigorous and collaborative process involving engagement with the collaborating public organization. The current form and formulation of the principles emerged from iterative cycles of refinement and generalization, crafted through discussions and reconciliation with practitioners, while simultaneously adhering to the theoretical underpinnings of the SCT framework. 6.1. Citizen-focused design principles. 6.1.1. Balancing AI Benefits for the Public Good. and Individual Freedom. To address potential conflicts of interest between the public good and individual freedoms, AI system designers and developers must be transparent about different needs and benefits in a comprehensible and accessible manner. This will help stakeholders comprehend the implications for the other involved actors. Recommendation: Communicate clearly and briefly the expected impact for the public and individual citizens. 6.1.2. Comprehensive, Inclusive, and Necessary. Data. To ensure the responsible use of citizens’ data, AI system designers, and developers who decide on which data to source for AI systems in the public sector, need to align system requirements with citizens’ rights. This will help to critically assess data needs and evaluate the availability of existing data, or the need for new data collection to adhere to the principle of data minimization. Recommendation: Explain which data is used, where it comes from, and the rationale behind its use. 6.1.3. Alignment with Governmental Mandates and. Process Integration for AI Systems. To ensure citizens are aware of the role of the AI model within the overall operational process and the link to governmental mandate this should be clearly defined. The link to legal or regulatory documents can be provided but the language needs to be kept simple. Recommendation: Present how AI fits in the overall process to fulfill the governmental mandate. 6.1.4. Gradual Information Provision. AI system designers and developers should ensure citizens understand the system logic and its data use by gradually providing clear information allowing people to view what is relevant for them on the level of detail they're comfortable with. Citizens must receive relevant and timely information about how the AI functions, its inputs and outputs, and its effects on them. To help citizens, information provision should consider the diverse technical literacy of the wider public. Recommendation: Present an overview first and allow users to access details and drill down if desired. Use familiar terminology and design patterns and provide contextual explanations. 6.1.5. Voluntary Feedback Mechanisms. AI Systems designers and developers should enable direct citizen participation to facilitate continuous learning about citizen needs, experiences, and system shortcomings. This can help to foster a participatory environment where citizens feel valued and heard, strengthen public trust and engagement with citizen input directly informing enhancements and adjustments. Recommendation: Integrate accessible feedback mechanisms and communicate how the feedback will be used to help further improve the AI system or the overall service provided. 6.2. Employee-focused design principles. 6.2.1. Consistency over Personalization. To ensure equal treatment, AI systems designers and developers should ensure consistent behavior in the AI system to prevent unjust and discriminatory outcomes through learned adaptation from individual employee behavior. Recommendation: Prevent employee personalization of AI system behavior. 6.2.2. Providing Legal Reassurance. AI systems designers and developers should provide employees with transparency on data sources, to ensure integrity, lawfulness, and traceability of the data used within AI systems. Recommendation: Explain which data is used, the statutory basis, and the rationale behind its use. 6.2.3. Enabling Human Decision-Making Power. To facilitate cooperation between employees and AI systems in the public sector, AI systems designers and developers should create AI systems that are perceived as helpers, with humans staying in control of the decision-making process. Recommendation: Provide humans with the ultimate power of decision when teaming with AI-enabled systems. 6.2.4. Automating Repetitive Administrative Tasks. To support employees within public service organizations in using their resources most efficiently, AI systems designers and developers should enable them to delegate repetitive administrative tasks to automated systems. Recommendation: Enable employees to hand over repetitive tasks to automated tools. 6.2.5. Allocating Resources sensibly. To support employees of public service organizations focusing on people who need the most support, AI systems designers and developers should focus on features and functionalities that support the prioritization of work. Recommendation: Enable employees to focus more on selected subgroups of people that need extra support. 6.2.6. Voluntary Feedback Mechanisms. AI Systems designers and developers should enable direct employee participation to facilitate continuous learning about employee needs, experiences, and system shortcomings. This can foster a participatory environment where employees feel valued and heard, strengthen trust and engagement with employee input directly informing enhancements and adjustments. Recommendation: Integrate accessible feedback mechanisms and communicate how the feedback will be used to help further improve the AI system or the overall service provided. 7. Discussion & Conclusion. According to Gregor et al. (2020), for design principles to be impactful and practical, they must be clear, actionable, and applicable to real-world scenarios. This ensures that the design principles are not only grounded in robust theoretical frameworks but also hold practical relevance and significance. This compels design theorists to carefully craft principles that are not only conceptually sound but also practically valuable and applicable in the context of Human-Centered AI in the public sector. One of the primary values of applying design principles in AI development is their capacity to operationalize considerations for existing power dynamics. As highlighted by Zuiderwijk, Chen, & Salem (2021), this becomes particularly important within the specific structures of the public sector. This involves balancing public good and individual benefits, recognizing and mitigating risks of data misuse or inadvertently introducing and reinforcing forms of bias. Effective design principles for the public sector help to ensure that AI systems do not exploit or disadvantage any group, but instead work to empower and elevate the collective well-being. Furthermore, these principles are instrumental in balancing the benefits between individual needs and societal welfare. This balance is crucial in public sector applications, where the impact of AI can be profound and far-reaching. Design principles guide the development process in considering the micro, meso, and macro aspects of AI impacts, ensuring that the systems serve not just individual users but also contribute positively to the community and society at large. Another critical area where design principles prove invaluable is in the creation of agreement modeling mechanisms within AI systems. These mechanisms are designed to facilitate consensus among diverse stakeholders, reflecting a range of perspectives and interests. By incorporating such mechanisms, AI systems can better adapt to and reflect the values and norms of the communities they serve. This is particularly important in the public sector, where AI systems must often navigate complex regulatory and ethical landscapes. To facilitate organizational changes, we must engage in visionary and strategic thinking as well as constant reflection on the future societies we are about to create. Design principles that incorporate different stakeholders’ stances, not only instruct on the 'how-tos', but also highlight the need for a clear understanding of objectives, application contexts, and rationales underpinning AI design. Engaging more stakeholders including citizens and employees, in shaping the future will lead to better policymaking, more effective socialization of emerging technologies, and greater legitimacy for hard decisions, as well as creating constituencies for long-term change. 8. Contribution and Future Research. The proposed design principles contribute to the information systems discipline by providing insights into citizen as well as employee stances on a human-centered design, development, and implementation of AI systems in public organizations. They aim to advance the discourse on how to merge theoretical knowledge with actionable guidelines, creating new knowledge for scholars and practitioners and providing helpful guidance for introducing AI in public services. The design principles draw from established and proven guidelines but go beyond the commercial interests of profit-oriented companies by incorporating critical considerations of citizens and public servants. Applying these design principles within the context of public organizations will not ensure a completely beneficial and risk-free application of AI systems, but can contribute to more responsible AI strategies and implementations. Future research should emphasize evaluation, a cornerstone of Design Science Research. An evaluative framework that considers the roles of key stakeholders in the development of design principles will enable scholars to assess these principles' usability across different user groups. Critical inquiries should focus on the clarity and practicality of these principles for implementers, as well as their effectiveness in meeting the objectives of users who apply these principles in real-world scenarios. While a detailed evaluation of design principles offers greater control and internal validity, it must be balanced with the need to consider the contextual dynamics underscoring the complex interplay between theoretical rigor and practical applicability in HCAI design for the public sector. 4. London: Palgrave. Hevner, A. R., March, S. T., Park, J., & Ram, S. (2008). Design science in information systems research. Management Information Systems Quarterly, 28, 6. Hildebrandt, M. (2018). Algorithmic regulation and the rule of law. Philosophical Transactions of the Royal Society Inayatullah, S. (2012). Futures studies: theories and methods. In: Al-Rodhan, N. R. F., & BBVA (Firm). There’s a future: visions for a better world (Z. Hallington & A. Uldall, Eds.; Carma & F. González, Trans.) (pp. 35–64). BBVA. Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1, 389-399. Junginger, S. (2016). Transforming Public Services by Design: Re-Orienting Policies, Organizations and Services around People. London, UK: Routledge. Lopez, P. (2021) Bias does not equal bias: A socio-technical typology of bias in data-based algorithmic systems. Internet Policy Review 10, 1–29. Meske, C., Bunde, E., Schneider, J., & Gersch, M. (2022). Explainable artificial intelligence: objectives, stakeholders, and future research opportunities. Information Systems Management, 39, 53-63. Nagitta, P. O., Mugurusi, G., Obicci, P. A., & Awuor, E. (2022). Human-centered artificial intelligence for the public sector: The gate keeping role of the public procurement professional. Procedia Computer Science, 200, 1084-1092. Nash, J. F. (1950). The bargaining problem. Econometrica, 18, 155-162. Neuhoff, R., Simeone, L., & Laursen, L. H. (2021). Exploring how design-driven foresight can support strategic thinking in relation to sustainability and circular policy making. European Academy of Design Conference: Safe Harbours. Pappas, I. O., Mikalef, P., Dwivedi, Y. K., Jaccheri, L., & Krogstie, J. (2023). Responsible digital transformation for a sustainable society. Information Systems Frontiers, 25, 945-953. Ramos, J., Sweeney, J. A., Peach, K., & Smith, L. (2019). Our futures: by the people, for the people. Rawls, J. (1996). Political Liberalism. New York: Columbia University Press. Riedl, M. O. (2019). Human‐centered artificial intelligence and machine learning. Human behavior and emerging technologies, 1, 33-36. Rogers, Y. (2022). Commentary: human-centred AI: the new zeitgeist. Human–computer interaction, 37, 254-255. Rousseau J-J. (1964). The Social Contract (1762). London: Penguin. Schmager, S. (2022). From Commercial Agreements to the Social Contract: Human-Centered AI Guidelines for Public Services. Proceedings of the 14th Mediterranean Conference on Information Systems (MCIS), Italy. Schmager, S., Pappas, I., & Vassilakopoulou, P. (2023). “Defining Human-centered AI: a comprehensive review of HCAI literature”, Proceedings of the 15th Mediterranean Conference on Information Systems, Spain (MCIS). Schmager, S., Grøder, C. H., Parmiggiani, E., Pappas, I., & Vassilakopoulou, P. (2024). Exploring citizens’ stances on AI in public services: A social contract perspective. Data & Policy, 6, e19. Seidel, S., & Watson, R. T. (2014). Improving the societal effectiveness of IS research: the pursuit of prescriptive accuracy. Available at SSRN 2477917. Seidel, S., Chandra Kruse, L., Székely, N., Gau, M., & Stieger, D. (2018). Design principles for sensemaking support systems in environmental sustainability transformations. European Journal of Information Systems, 27, 221-247. Sein, M. K., Henfridsson, O., Purao, S., Rossi, M., & Lindgren, R. (2011). Action design research. MIS Quarterly, 37-56. Shneiderman, B. (2020a). “Human-centered artificial intelligence: Reliable, safe & trustworthy”, International Journal of Human-Computer Interaction, 36, 495-504. Statistisk sentralbyrå (2022). Sickness absence. the linked source fravaer-og-arbeidskonflikter/statistikk/sykefravaer. Stix, C. (2021). Actionable principles for artificial intelligence policy: three pathways. Science and Engineering Ethics, 27, 15. Van Aken, J. E. V. (2004). Management research based on the paradigm of the design sciences: the quest for field‐tested and grounded technological rules. Journal of Management Studies, 41, 219-246. Vassilakopoulou, P., Parmiggiani, E., Shollo, A., & Grisot, M. (2022). Responsible AI: Concepts, critical perspectives and an Information Systems research agenda. Scandinavian Journal of Information Systems: Vol. 34: Iss. 2, Article 3. Velsberg, O., Westergren, U. H., & Jonsson, K. (2020). Exploring smartness in public sector innovation-creating smart public services with the Internet of Things. European Journal of Information Systems, 29, 350-368. Venable, J., Pries-Heje, J., & Baskerville, R. (2016). FEDS: a framework for evaluation in design science research. European Journal of Information Systems, 25, 77-89. Walz, A. (2017). A Holistic Approach to Developing an Innovation-Friendly and Human-Centric AI Society. IIC-International Review of Intellectual Property and Competition Law, 48, 757-759. Wang, D., Churchill, E., Maes, P., Fan, X., Shneiderman, B., Shi, Y., & Wang, Q. (2020). From human-human collaboration to Human-AI collaboration: Designing AI systems that can work together with people. In Extended abstracts of the 2020 CHI conference on human factors in computing systems (pp. 1-6). Wirtz, B. W., Weyerer, J. C., & Geyer, C. (2019). Artificial intelligence and the public sector—applications and challenges. International Journal of Public Administration, 42, 596-615. Wright, A. P., Wang, Z. J., Park, H., Guo, G., Sperrle, F., El-Assady, M.... & Chau, D. H. (2020). A comparative analysis of industry human-AI interaction guidelines. arXiv preprint arXiv:2010.117. Zuiderwijk, A., Chen, Y. C., & Salem, F. (2021). Implications of the use of artificial intelligence in public governance: A systematic literature review and a research agenda. Government Information Quarterly, 38, 101577.