Latest / 1 More Paper / Exploring the Relationship Between Human-Centric AI and firm Idiosyncratic Risks — Liu et al. (2026)

Transcript

Exploring the Relationship Between Human-Centric AI and firm Idiosyncratic Risks — Liu et al. (2026) · 1 More Paper

• You’re listening to “Exploring the Relationship Between Human-Centric AI and firm Idiosyncratic Risks,” by Z.-Y.R. Liu and colleagues. Published in 2026. Abstract. Despite the extensive discussions of human-centric AI (HCAI) in Industry 5.0, its effects on firms’ idiosyncratic risks (IR) remains underexplored. This is an imperative issue for firms navigate financial risks during the current technological revolution, as IR reflects investor reactions to corporate heterogeneous AI strategies and implementations by isolating firm-level stock volatility from systematic factors. Integrating situated AI theory with social-technical systems theory, we conceptualise HCAI as a situated AI strategy that reduces AI-related ethical risks and fosters AI-Human synergies in firms’ business operations by aligning with stakeholders’ diverse expectations. Moreover, socio-technical factors, namely digitalisation, operational efficiency, executive shareholding, and CEOs with IT background, may moderate the HCAI-IR relationship. Using a multi-source panel dataset of Chinese listed firms from 2015 to 2023, we find that HCAI is associated with lower firm IR. Furthermore, digitalisation and executive shareholding strengthen this risk-reducing effect, whereas operational efficiency and CEOs with IT background surprisingly attenuate it. Our findings offer theoretical contributions and practical insights for both ethical AI governance and firm financial risk management in the AI era. Shivam Gupta the email address Zhen-yuan Ralph Liu the email address Yu-ting Wang the email address Jia-jia Yan the email address Mihalis Giannakis the email address 1 Department of Accounting, School of Economics and Management, China University of Mining and Technology, Tongshan District, Xuzhou, China 2 School of Economics and Management, Nanjing Forestry University, Xuanwu District, Nanjing, China 3 School of Economics and Management, Fuzhou University, Minhou District, Fuzhou, China 4 Department of Information Systems, Supply Chain Management & Decision Support, NEOMA Business School, Reims, France 5 Audencia Business School, Nantes, France 1 Introduction A wide range of businesses have adopted Artificial Intelligence (AI) in recent years, with many firms using it to support decision-making and enhance daily operations. AI is a broad term with many possible definitions depending on context, and it refers to giving machines the capacity to operate in domains that demand human judgment, e.g., tasks such as problem-solving, decision-making, and creative generation. Building on prior computer technologies, AI shows an unprecedented capacity to optimise operational processes, enhance strategic decision-making, and even drive business model innovations. Just as consumers have quickly embraced the possibilities of conversational AI in their daily lives, managers and employees have also incorporated AI into their corporate frameworks. However, the transformative potential of AI is still tempered by significant challenges, from generalisability, ethical issues and embedded bias to a tendency toward myopic optimisation. These issues can cause AI use to have significant repercussions, and together expose firms to financial, social, and ethical perils. The Commonwealth Bank of Australia (CBA) typifies the dangers involved, as the firm attempted to lay off 45 employees due to the integration of a new AI voice-bot system in July 2025, but reversed its decision a month later after admitting its initial assessments did not fully consider the business implications. The abortive layoffs led to decreased customer satisfaction, reputational damage, and increased operational costs1. Conversely, when IBM announced an AI partnership prominently featured with governance protocols, explainability standards, and human‐in‐the‐loop design, its stock price rose sharply2. These contrast cases suggest that investors reward firms’ AI strategies framed with human-centric orientation rather than workforce substitutions. More broadly, these contrasting reactions suggest that capital markets do not respond to AI adoption per se. Rather, investors appear to distinguish firms’ AI strategies that signal stronger governance, human oversight, and longer-term value creation from those intensifying stakeholder tensions and implementation uncertainty. This practical contrast therefore impels firms to consider how they can design and deploy AI in ways that reduce firm-specific risk. In response to growing concerns over AI risks, a new synergistic Human-Centric AI (HCAI) concept was defined as part of Industry 5.0, a proposed next phase of industrial development by the European Union in 2021. HCAI aims to enrich human-AI collaboration, moving the conversation from what is technically possible to ensuring responsible AI use. The key objective reframes human-AI interaction, positioning AI as complement rather than replacement. This shift minimises the typical risks of using AI for business management and industrial systems, shoring up trust and ensuring a base level of suitability. HCAI enables organisations to navigate complex internal structures and volatile market environments, while mitigating algorithmic bias and enhancing organisational resilience. In addition, by emphasising transparency, explainability, and ethical accountability, organisations adopting HCAI principles are more likely to generate social value through improved efficiency and fairness in decision-making and operational processes. The guiding principle is that AI-driven performance gains should not compromise fairness or social responsibility. In this way, HCAI has moved from a purely conceptual model to an increasingly crucial strategic framework, enabling firms to secure a sustainable competitive edge in a fast-moving, disruptive digital era. Against this backdrop, HCAI critically explains why firms’ AI initiatives may reassure investors instead of heightening concerns regarding AI execution, ethics, and stakeholder backlash. Despite most literature affirming the merits of HCAI across fields such as intelligent manufacturing optimisation, operational performance, and sustainability, there is little research examining whether and how investors respond to corporate HCAI initiatives. Addressing this research gap is crucial, as investor assessments significantly affect firm decision-making that aims to balance emerging technology development and industrial transformation. Investor reactions can also hinder a firm’s access to financial resources and exacerbate risks, thereby further obstructing its long-term prosperity. Reflecting investor responses to a firm’s future cash flows through stock price volatility, idiosyncratic risk (IR) accounts for over 80% of overall risk in the capital market. Further, IR is substantially distinct from systemic risk, assessing market-wide factors that affect all stocks, as it not only links more closely to specific firm-level factors but also serves as a robust measure of how investors react to firms’ strategic decisions. When shocks become correlated, for example through shared AI infrastructures, common vendors, or synchronized investors reactions, firm-level volatility can turn into systemic risk. Therefore, IR has been extensively used as the predictor of firm risks across the fields of information systems, accounting and finance, economics, and management studies. For instance, Li, Li and Sethi (2021) unveil that CSR practices exhibit the U-shaped impact on firms’ IR, the relationship of which could be flattened by firms’ AI innovations. Hudson and Morgan (2024) reveal that firms’ AI exposure significantly lowers IR, as both investors and financial analysts perceive AI as the core driving force behind future business, thereby bringing firms stable future cash flows. However, to the best of our knowledge, no prior study sheds light on the effect of HCAI on firms’ IR during the Industry 5.0 transition. This gap is crucial, given that HCAI innovation and implementation programmes require substantial resources, capital and organisational change, potentially exposing firms to high execution risks. Further, more recent research emphasises the importance of identifying contingent factors that curb IR from corporate characteristics and internal governance structures and then enriches managerial and policy implications for managing firms’ IR. To sum up, this study is guided by two research questions as follows: RQ 1 Whether and how does Human-centric AI affect firms’ idiosyncratic risks? RQ 2 If so, what are the contingent factors moderating the HCAI-IR relationship? To answer these questions, we integrate situated AI theory (SAIT) with the socio-technical system (STS) theoretical framework to view HCAI as firms’ situated AI strategy during the Industry 5.0 transition. SAIT argues that only by overcoming the generic nature and myopia of AI, and by addressing the needs of broader stakeholders, can AI strategies provide firms with a competitive advantage. In this vein, HCAI enhances employee skills, unlocks human-AI ensemble power, and fosters positive stakeholder relationships through human-centred design, ethical management, and responsible implications. Investors observe these AI choices through disclosures and innovation outputs. Attention shifts toward long-term value, and firm-specific volatility (idiosyncratic risk) tends to fall. Meanwhile, SAIT further indicates that organisational constraints significantly moderate the effects of firms’ AI strategies at both technological and socio levels. Therefore, we follow Bednar and Welch (2020) to explore contingent factors that potentially moderate the HCAI-IR relationship from both socio-technical perspectives. Specifically, we argue that more substantial risk reduction occurs when firms possess competitive advantages in both socio (i.e., executive shareholdings and CEO’s IT background) and technical (i.e., digital transformation and operational efficiency) systems. To test these propositions, we follow Liu et al. (2025)’s work to identify HCAI innovations from 2.873 million patent textual data of Chinese listed firms, as HCAI remains at its early stage and the relevant technological innovations reflect firms’ AI strategies more accurately. After econometrically analysing panel data containing 16,461 observations of Chinese listed firms from 2015 to 2023, we found that HCAI significantly reduces firms’ IR. This finding remains robust after several examinations, including alternative measures, time lagging model, propensity score matching (PSM), and sub-sample regression methods. More importantly, the moderating analysis results further demonstrate that the risk-curbing effect of HCAI is amplified when firms possess higher levels of both digital transformation and executive shareholding. However, operational efficiency and IT background CEO weaken the reducing impact of HCAI on IR. One plausible explanation pertains to shareholders’ pressure for throughput and a technology-first orientation of companies, which dilutes human oversight and cross-functional checks, leading investors to perceive higher execution risk. Regarding the contributions of this study, we first contribute to the ongoing discussion around HCAI by unveiling its curbing effect on firms’ IR. Prior work stresses the internal gains of HCAI in production systems, employee skill-upgrading, and sustainable performance. We extend this view and provide a novel perspective on how investors reflect upon and evaluate corporate HCAI innovations. Our evidence links design-led governance to this financial outcome and answers the special issue’s call for work that connects AI design to financial risk analysis. Secondly, we identify the boundary conditions of the risk-reducing effects of HCAI related to technical and social dimensions. Digitalisation and executive shareholding strengthen the negative association between HCAI and IR, while operational efficiency and an IT background CEO weaken it. These results give practical guidance for HCAI implementation. Finally, this paper expands the organisational constraints of SAIT by integrating the socio-technical perspective and by anchoring claims in a finance construct observed by investors, thereby offering a theoretical implication for future research in this field. 2 Literature Review and Theoretical Background 2.1 Human-Centric AI (HCAI) in Industry 5.0 Transition AI played a crucial role in achieving Industry 4.0 and led to exponential gains in production efficiency through automation technologies. When integrating AI into digital and smart transformation, firms obtain substantial operational advantages through automating routine tasks, boosting product innovation processes, and optimising resource allocations. However, AI also exposes essential risks to firms due to its generality, ethical concerns, functional myopia, and lack of algorithmic transparency. For instance, AI-based recruitment tools trained on historical data may systematically exclude certain groups (e.g., female technical candidates), as evidenced by the failure of Amazon’s AI recruitment system , which reveals fundamental conflicts of AI’s automation-based logic that overlook human values. The European Union launched its Industry 5.0 initiative to cope with these sorts of AI challenges, seeking to move humans back to the driver’s seat in terms of decision-making and advocating for the use of “human-AI ensembles”. HCAI serves as a core enabler here, with a more human-centred approach to AI technological development, innovation, and applications that address AI’s inherent technical and societal limitations. Specifically, HCAI is intended to assist in routine operations, freeing employees from repetitive tasks to focus on more complex work that demands greater situational awareness, decision oversight, or human-AI collaboration. This shift has the potential to revitalise the value and demand for such transformed roles. Core HCAI mechanisms also facilitate the capture, transformation, and utilisation of tacit knowledge that might otherwise be lost, thereby ensuring more sustainable organisational assets. Unlike Industry 4.0, this approach specifically emphasises algorithmic transparency and explainability, alongside the embedding of values like fairness and safety that might otherwise be undervalued by an AI agent. This is exemplified by frameworks like the EU’s “Trustworthy AI”, which aims to rebuild trust that has been eroded by early human-AI interaction. Beyond these broader HCAI concepts, more recent studies have specifically investigated the merits of HCAI in a management context. For instance, Psarommatis, May and Azamfirei (2023) propose the integration of human-centric principles into AI-enhanced Zero-Defect Manufacturing systems. Matthews et al. (2025) found that HCAI aligns firms’ resources desire with broader stakeholder groups through AI synergies along the value chain. At a more micro level, HCAI could grant employees with effective human-AI collaborations, which boosts corporate social performance and sustainable business operations. Liu et al. (2025) empirically examined the effect of HCAI on both social and operational performances of firms. However, to the best of our knowledge, the extant literature has not investigated whether and how investors react to firms’ HCAI. Addressing this gap is crucial for managing firms’ financial risks, particularly given that HCAI requires substantial resources, capital and organisational structures from investors. Accordingly, the primary objective of our study is to explore in depth the relationship between HCAI and investor reaction, the latter of which is commonly proxied by firm idiosyncratic risk in the extant literature. 2.2 Managing Idiosyncratic risk (IR) IR reflects investors’ evaluation of and reaction to a firm’s future cash flows via its stock price volatility. Distinct from systematic risk assessing market-wide factors that affect all firms, IR is tied more closely to firms’ specific characteristics and accounts for more than 80% total risks in the capital market. Also, IR serves as a robust indicator for assessing the efficacy of firms’ strategic decisions, influencing firms’ future cash flows and market values in either a favourable or a negative manner. Hence, effectively managing IR can safeguard investor and shareholder returns, ensure stable cash flows, and mitigate the risk of bankruptcy. Previous research has explored a wide array of IR antecedents, ranging from IT strategy and top management team characteristics to information disclosure and CSR practices. More recently, a burgeoning body of work has examined the link between AI and IR following three distinct perspectives in terms of AI exposure, AI design, and AI governance. In particular, studies adopting the exposure perspective treat AI primarily as the exogenous technological shock affecting firms’ business operations. Hudson and Morgan (2024) find that greater AI exposure reduces firm IR, as financial analysts attribute technological changes to systematic rather than idiosyncratic risk. Conversely, Boido and Aliano (2025) argue that AI exposure may increase IR via heightening business operations uncertainty, given that adopting an AI‐driven mindset consumes substantial resources and requires structural changes. While these exposure-oriented research offers valuable macro-level insights, they pay limited attention to how firms could strategically shape their AI deployments in ways that align with their characteristics and value-generating processes. Addressing this gap, Li, Li and Sethi (2021) introduced the AI governance perspective, showing that AI innovations can flatten the U-shaped relationship between firms’ CSR practices and IR. This finding highlights that the risk implications of AI are contingent upon how firms oversee, incentivise, and align AI with the interests of broader stakeholders. Building upon this pragmatic view, a nascent stream of research has begun to explore the design perspective, focusing on the intrinsic characteristics of AI systems that collectively constitute trustworthy and human-centric AI. HCAI on firms’ IR by integrating these three perspectives. Moreover, Morgeson et al. (2024) and Aldawsari, Choudhry and Luo (2025) indicate that one of the main reasons for the current fragmented findings in the AI-IR relationship pertains to the omission of contingent factors rooted in corporate characteristics and internal governance structures. They advocate for more fine-grained investigations that identify when and under what conditions AI curbs IR, as such knowledge offers richer managerial and policy implications in AI era. In this regard, the second objective of our study is to explore contingent factors that may moderate the HCAI-IR relationship. 2.3 Integrating Situated AI theory with the Socio-Technical Systems Framework Situated AI theory provides us with an essential theoretical lens to accomplish the main research objectives above. Moving beyond the mere technology adoption perspective, SAIT focuses on how organisations can use AI strategy to acquire multi-fold competitive advantages. These strategies operate using a three-stage process: grounding, bounding, and recasting. Grounding AI refers to addressing its generic nature by deeply integrating it into a firm’s operational processes, thereby enhancing resource allocation efficiency. Bounding AI seeks to undermine the possibility of AI performing unethical actions, such as data breaches, cybersecurity issues, and other unlawful encroachment. Finally, recasting AI is focused on reconfiguring stakeholder relationships to align the interests of broader stakeholders, thereby overcoming AI myopia. Following these theoretical guidelines, HCAI can be understood as a corporate AI strategy aligned with Industry 5.0 requirements, which can be further delineated into three core constructs: human-centred design, ethical management, and responsible implications. Human-centred design prioritises human needs in AI development via creating user‐friendly interaction interfaces as well as integrating ergonomic considerations. For instance, Covariant developed AI-driven robotic arm for collaborating with workers to process logistics parcel sorting and waste classification tasks in extreme environments3. Ethical management establishes base-level guardrails to address the challenges posed by AI’s inscrutable nature, including algorithmic black boxes, biases, and privacy violations, thereby cultivating the trust necessary for effective human-AI collaboration. Simultaneously, responsible implications are pivoting towards a broader perspective in AI development, integrating stakeholder interests to counteract short-sighted decisions driven purely by efficiency gains. It is important to distinguish HCAI from related frameworks such as “trustworthy AI” and “responsible AI”. While these frameworks articulate important ethical principles that AI systems should satisfy, they remain largely principle-based and context-agnostic. By contrast, HCAI is a more situated organisational approach, as it embeds those principles into firms’ operational processes, governance mechanisms, and stakeholder relationship management. In this sense, HCAI addresses not only what ethical AI should look like in principle, but also how such principles can be operationalised through firms’ design choices, ethical issue management, and deployment practices during the Industry 5.0 transition. Figure 1 illustrates how we develop the core concepts of HCAI via the SAIT lens. SAIT also indicates that organisational constraints moderate the effect of situated AI strategies. More importantly, Moser et al. (2024) pinpoint that SAIT needs to take the alignment between social and technical factors into account, especially when forming the comprehensive AI strategy for balancing stakeholders’ expectations, which resonates profoundly with socio-technical systems (STS) theory. STS 3 Source available at: ​h​t​t​p​s​:​/​/​c​o​v​a​r​i​a​n​t​.​a​i​/​r​e​s​o​u​r​c​e​s​/​i​n​t​r​o​d​u​c​i​n​g​-​t​h​e​-​n​ e​x​t​-​p​h​a​s​e​-​o​f​-​o​u​r​-​a​i​-​r​o​b​o​t​i​c​s​-​j​o​u​r​n​e​y​.​ theory holds that only through the congruence between technical and socio subsystems can firms achieve superior performance. The technical subsystem comprises technology and task factors, while the socio subsystem involves organisational structures and people. To operationalize STS theory in our moderation analysis, we screened one proxy from each of these four dimensions, ensuring balanced coverage of the potential boundary conditions that shape investor reactions to firms’ HCAI. Specifically, we capture the technology dimension through digitalisation, which reflects the foundational digital infrastructure wherein HCAI is implemented. The task dimension is represented by firms’ operational efficiency, as it reflects firms’ task-level capability via measuring how systematically and effectively firms can converts inputs into pursued outputs. On the socio subsystems, we capture the structure dimension through executive shareholding, which aligns managerial incentives with long-term value creation and risk mitigation. Finally, the people dimension is represented by the CEO’s IT background, which shapes leadership cognition and strategic orientations regarding technology deployment. Taken together, integrating SAIT with STS theory maps directly onto our hypotheses. SAIT provides the basis for explaining how HCAI, through grounding, bounding, and recasting, can be translated into human-centred design, ethical management, and responsible applications that reduce investor concerns about execution, ethics, and stakeholder conflict, thereby lowering firm idiosyncratic risk. STS theory, in turn, provides the basis for identifying the socio-technical conditions under which this risk-reducing effect is likely to vary. Specifically, the technical subsystem is captured through digitalisation and operational efficiency while the social subsystem is captured through executive shareholding and the CEO’s IT background. 3 Hypothesis Development Figure 2 presents the conceptual research model. In line with the integrated framework developed in Sect. 2.3, we derive H1 from SAIT and H2-H5 from the STS-based boundary conditions surrounding the implementation of HCAI. 3.1 The Effect of HCAI on IR Drawing upon SAIT, we argue that HCAI could decrease firms’ IR through at least three ways. First, investors particularly value HCAI when delving into target firms’ AI strategy. Given the growing importance of sustainable investment, investors frequently rely on firms’ sustainable performance for investment decisions. When adopting HCAI, investors are aware of firms’ efforts in balancing sustainable development requirements and AI technological pushes, improving the value of HCAI in hampering noise trading behaviour. Even though HCAI may impose additional processing costs on understanding firms’ motivations, the human-AI collaborative methodologies, data sources, and weighting systems could bring firms’ long-term returns. As a result, investors may judge a firm’s actual value in HCAI, thereby decreasing its idiosyncratic risk. Second, HCAI incorporates development approaches that help align a firm’s AI ambitions with the interests of internal stakeholders, such as human-centred design and ethical management. This is intended to ensure sustainable growth and minimise future cash flow concerns. Specifically, human-centred design aims to augment employees with AI, rather than replacing them entirely. Unsurprisingly, this attitude tends to enhance employee impressions of job security and even fosters innovation. Moreover, HCAI supports upskilling and Fig. 2 Conceptual research model reskilling initiatives, enabling employees to adapt to their evolving work demands in the AI era. This, in turn, mitigates operational risks associated with large-scale layoffs. For example, Haier, a large manufacturing firm in China, adopted HCAI principles in 2020 to transition employees into “autonomous entrepreneurs”. This transformation increased employee innovation participation by 61% and reduced staff turnover by 40%. Such HCAI-driven innovation further enabled the company to achieve a growth rate five times the industry average between 2021 and 20244. Third, HCAI has the potential to create synergistic benefits that extend beyond individual firms when understood from a multi-stakeholder viewpoint. In this way, value creation is a mutually reinforcing process where the success of each participant depends on the contributions of others. This interdependence is clearly reflected in HCAI principles, where intelligent systems require human collaboration and continuous data input to achieve optimal performance. The HCAI synergy effect therefore enhances value for broader stakeholders, while simultaneously increasing customer satisfaction and employee welfare, often without detriment to other parties. Moreover, adopting these principles may enhance coordination between upstream and downstream supply chain partners, thereby generating social value across the vertical business network. As AI investments can produce positive externalities, HCAI may gradually spread throughout the supply chain, contributing to improvements in collective social performance. Taken together, HCAI can serve as an important driver of more sustainable development paradigms. In this context, IR may be reduced when the focal firm adopts HCAI. Based on the above reasoning, we propose the following hypothesis: H1 Human-centric AI decreases firm idiosyncratic risk. 3.2 The Moderating Factors Despite the situating AI implementations for achieving competitive advantages, HCAI also needs to align with the social and technical subsystems a firm occupies. STS theory provides four components that should be analysed together when evaluating the organisational performance that emerging technologies could bring, i.e., technological foundation, operational processes, roles of people, and organisational structure. Based on the extant studies on IR antecedents, we propose to investigate the contingent factors that potentially moderate the HCAI-IR relationship under the STS theoretical framework, which includes organisational digitalisation, operational efficiency, executive shareholding and CEO with IT background. Digitalisation refers to the extent to which a firm undertakes technological digital transformation. Specifically, it signifies the deep integration of digital technologies with business processes, driving business model innovation and organisational change. Logically, firms with high levels of digitalisation possess a better foundation for integrating HCAI principles into their operational management and practices, as adopting HCAI enhances a range of digital technologies, to say nothing of its potential positive effects on the digital mindsets of employees. Therefore, we propose the following hypothesis: H2 A firm’s digitalisation level moderates the relationship between HCAI and IR by amplifying the risk-reducing effect. Operational efficiency (OE) measures a firm’s operational capability in terms of converting resource inputs into expected outputs. It is an imperative indicator of long-term financial performance, as firms with higher OE signal strong process discipline and resource orchestration. High OE supports integration of HCAI strategies across operations, while minimising the traditional risks. Investors can interpret HCAI in high OE firms as more likely to translate into stable cash flows. Therefore, we propose the third hypothesis as follows: H3 A firm’s operational efficiency moderates the relationship between HCAI and IR via amplifying the risk-curbing effect. Executive shareholding refers to the extent to which a firms’ shares are held by C-level executives as well as directors and board supervisors, attenuating the myopic behaviors of managers and motivating managers to implement digital technologies in more ethical ways. In this regard, HCAI aligns with top managers’ expectations when they hold more shares in firms and focus on sustainable development, which mitigates the risk-countering effects on firms’ IR. Thereby, we propose that: H4 The more top managers hold a firm’s share, the stronger risk-curbing effect HCAI generate on IR. CEOs’ previous personal and professional experiences and knowledge fundamentally influence a firm’s subsequent strategic orientations and related risks. When a firm’s CEO has an IT background, she may lay more emphasis on digital technologies, R&D investment and technology-driven operations. This professional background could navigate CEOs to obtain more profound insights regarding HCAI benefits, strengthening the risk-reducing effect of HCAI on IR. Interestingly, CEOs with backgrounds in IT are more likely to trust technological approaches like HCAI to solve obstacles facing the Industry 5.0 transition. Yet, the technology-centric leadership style may also introduce countervailing dynamics. CEOs with strong IT backgrounds may instinctively prioritize automation, algorithmic optimization, and efficiency gains over the human‐centric values that distinguish HCAI from generic AI. This also has a knock-on effect on how well a leadership team can understand HCAI, thereby allowing more informed decisions that eventually result in more traditional positive indicators such as profits and firm growth. Following this logic, we propose: H5 IT background CEOs moderate the relationship between HCAI and IR through enhancing the risk-reducing effect. 4 Methodology 4.1 Data and Samples As a rapidly developing economic power, China provides a unique opportunity to empirically test our hypotheses. The Chinese government launched a strategic push to integrate AI into intelligent manufacturing in 2015, aligning with the national “Made in China 2025” initiative. Under this policy, China experienced exponential growth in AI innovation. One way to visualise this is depicted in Fig. 3, which shows that the number of AI patents held by Chinese listed companies increased by nearly an order of magnitude over an eight-year span, from 676 in 2015 to 6,577 in 2023. One year later, President Xi announced China’s Human-Centric AI Initiative, which mirrors the human-centred principles of the EU’s Industry 5.0 and the UN’s Governing AI for Humanity agenda. While AI adoption in China continues at pace, specific characteristics of Chinese stock market may pose challenges to adopting HCAI principles, such as its high proportion of retail investors, prevalent speculative trading, and insider trading issues. These unique institutional factors have the potential to influence how effectively HCAI principles align with the transition to Industry 5.0. With both these headwinds and tailwinds in mind, these unique conditions make the study of Chinese listed firms a compelling sample for testing the hypotheses above. In addition, the Chinese capital market is characterised by relatively stringent disclosure requirements for technological innovation activities, including patent filings, digital transformation initiatives, and corporate governance practices. Listed firms are required to provide detailed disclosures in annual reports and regulatory filings, which enhances the visibility of AI-related strategies to investors. This institutional environment facilitates investors’ assessment of firms’ human-centric AI efforts and strengthens the Fig. 3 AI innovation patent in China informational link between HCAI innovation and market-based risk perceptions. Our sample includes all firms listed on the Shanghai and Shenzhen A-share markets between 2015 and 2023. These firms were evaluated in terms of HCAI innovation based on their patent records from the China Patent Database (CPD), encompassing some 2,837,000 numbered entries with application dates and descriptions. Next, stock trading data, corporate financials, and governance variables were sourced from the China Stock Market & Accounting Research (CSMAR) database. Finally, each firm’s annual reports were collected from the Chinese Research Data Services (CNRDS) platform. We further refined the sample firms according to the following screening criteria: (i) exclusion of firms under Special Treatment (ST) status to avoid the influence of financially distressed performers; (ii) removal of entities with significant missing or anomalous data; and (iii) winsorisation of all continuous variables at the 1st and 99th percentiles to avoid the impact of outliers. After integrating these databases and applying the above criteria, a final sample of 16,461 firm-year observations was obtained. 4.2 Measures 4.2.1 Dependent Variable: Idiosyncratic risk (IR) Following Aggarwal, Dai and Walden (2011), Gu, Jiang and Xu (2019) and Aldawsari, Choudhry and Luo (2025), we employ Fama and French’s three-factor model to calculate individual firms’ idiosyncratic risk. First, we calculate the abnormal returns of firm i in month m by the following equation: disp l a y - e q-Equ1  wherein inline-eq-IEq1 indicates stock returns of firm i that exceeds the risk-free rate on monthly basis. inline-eq-IEq2 is the market risk premium factor that captures the monthly difference between Shanghai-Shenzhen Index and risk-free rate, while inline-eq-IEq3 is book-to-market risk premium factor that derives from differential returns between portfolios with high and low book-to-market ratios. inline-eq-IEq4 represents the size-based factor, which is determined by the return differential between small and large-cap stock portfolios. While i nline-eq-IEq5 represents the abnormal returns, i nline-eq-IEq6, i nline-eq-IEq7 and i nline-eq-IEq8 demonstrate risk levels of each factor. Following Hudson and Morgan (2024) and Aldawsari, Choudhry and Luo (2025), we leverage the residual term of i nline-eq-IEq9 in Eq. to further compute firm i’s IR in year t by executing the following equation: y -displ a  eq-Equ 2 4.2.2 Independent Variable: Human-Centric AI (HCAI) Following Liu et al. (2025)’ s work, we leverage pre-trained large language models (LLMs) to identify and quantify HCAI in this study. LLMs provide us with the promising and theory-grounded classification tool to identify whether patents in a large corpus of firm innovation records are substantively related to human-centric AI. This approach is suitable for our study, given that HCAI is still an emerging and nuanced construct that is difficult to be captured through generic AI labels alone. Instead, processing patent narratives via LLMs provides rich information about firms underlying AI innovation strategies and then allows us to build a scalable and theory-informed firm-year measure of HCAI.Specifically, we design a four-step process to execute the task of HCAI identification, which consists of learning context setting, prompt structuring, results evaluation and validation, and measuring construct. Step 1: Setting the Learning Context We leverage the grounding, bounding, and recasting constructs deriving from SAIT to build the HER framework of HCAI. This framework required sample firms’ AI innovation to demonstrate the objectives of human-centred design, ethical management, and responsible implications. We then applied such the criteria and learning context to screen 2.873 million technological patents of sample firms. Step 2: Prompt Structuring To facilitate the identification process, we utilised a dynamic prompt engineering framework with specific inference parameters (temperature: 0.1, maxtokens: 150, topp: 0.9) to prioritise deterministic and concise outputs. Structured prompts defined the LLM’s role as a “Patent Innovation Identification Expert” incorporated classical AI definitions alongside our HCAI criteria, and enforced binary Yes/No outputs with brief explanations. Appendix A illustrate the specific structured prompt in this study. Step 3: Evaluating and Validating Results We induce our test data and structured prompts through APIs among LLMs, including Google’s Gemini 2.0 model, OpenAI’s ChatGPT 4o model, Deepseek R1 model, and Alibaba’s Qwen 2.5 model. As for the inference mode choice, we take Hillert, Niessen-Ruenzi and Ruenzi (2025)’ suggestion to adopt few-shot approaches, which could enable LLMs to understand our context with few examples better. After running the structured prompts, we randomly extract 10,000 results and evaluate the prediction accuracy of each model. Table 1 presents the accuracy among four LLMs, implying that Qwen 2.5 achieved the highest performance in both training and test sets. Therefore, the optimised LLM performed automated identification across the patent corpus, with results validated through manual verification and expert evaluation, achieving high reliability confirmed by a Kappa score of 0.91. While the LLM-based When model classifications differed, we assigned the final label using a majority vote across the three models. Figure 4 illustrates the detailed workflow of our HCAI identification process. Step 4: Measuring Construct. We operationalise inline-eq-IEq10 as the annual count of HCAI innovations from firm i in year t. To correct for substantial right-skewness in the distribution, we further took the logarithmic form of ln(1 + inline-eq-IEq11) in our empirical tests. While the LLM-based classification approach enables large-scale identification of HCAI-related innovations, it may inevitably involve a degree of classification noise due to ambiguity in patent language and overlapping technological concepts. However, such potential misclassification is unlikely to systematically bias our main results for several reasons. First, the identification framework is grounded in theoretically informed constructs derived from situated AI theory, ensuring conceptual consistency in screening criteria. Second, the high inter-model accuracy and the manually validated Kappa score of 0.91 indicate strong reliability of the classification outcomes. Third, any remaining noise is more likely to be random rather than directional, which would attenuate estimated coefficients toward zero rather than generate spurious significance. Therefore, the observed negative relationship between Fig. 4 The procedures of applying LLM in HCAI identification HCAI and idiosyncratic risk can be viewed as a conservative estimate of the true effect. 4.2.3 Moderators Digitalisation (Digital) Following prior research, we measured the extent of firms’ digitalisation via using textual analysis approach. Specifically, we first obtained the annual report texts from the CNINFO for all sample firms during the observation period. We extracted text from the full annual report narrative, excluding tables, financial statements, and footnotes. Then, keywords matching process were executed based on a seed dictionary of terms related to digital technologies, as these words represent the foundational elements of digital transformation strategies mentioned in annual reports. Finally, we identified only those texts that describe the application scenarios of these digital technologies within business operations. Through the above process, we aggregated the word frequencies from the digitalisation dictionary to obtain the total word frequency of digital transformation-related terms. A detailed illustration of this identification procedure is provided in Appendix B. This total frequency is then divided by the total number of words in the annual report and multiplied by 100 to calculate the final inline-eq-IEq12 of firm i in year t. Operational efficiency (OE) Drawing on the approach of, we adopted stochastic frontier analysis (SFA) to estimate firms’ operational efficiency. The first step involves specifying a conventional stochastic production function for firm i in industry j in year t, which incorporates both expected outputs and operational inputs as follows. s di play-eq-E qu3  wherein i and t represent the firm and year, respectively. inline-eq -IEq13 represents the outputs of firm i operating in industry j in year t, and the inputs for firms operations include, inline -e q-IEq14, and inline- eq-IEq16. inli ne -eq-I Eq15 While i nline-eq-IEq17 delineates the random and unobserved factors that influence firms operations, i nline-eq-IEq18 specifically refers to the inefficient items that ranges from 0 to 1. With value of 0, i nline-eq-IEq19 indicates the firms achieve at the maximised frontier with no efficiency loss. Therefore, i nline-eq-IEq20 is a relative measure of how inefficient a firm is when compared with the corresponding frontier in the same industry within the same year. In this sense, the operational efficiency of firm i in year t could be estimated as follows: q -Equ4  Executive Shareholding (Ehold) We follow Lyu et al. (2025) to identify directors, supervisors, board of supervisors, and senior managerial staff as the executives of sample firms. Eholdit is measured by the ratio of total shares held by these executives to the total number of shares issued by the sample firm i in year t. The higher the value of Eholdit, the more substantial equity incentives the sample firms implement. IT Background CEO (IT CEO) We first compiled sampled firms’ CEO profiles from CNRDS, including both formal position holders and acting/intern CEO (e.g. deputy general managers). Second, we adopted the constructs from prior studies to identify IT-related backgrounds from CEOs’ educations, professional certifications and work experience. Finally, all screening results underwent manual verification to ensure accuracy, and thereby we constructed IT CEOit. 4.2.4 Control Variables Following prior studies n firm risk and AI-related corporate outcomes, we control an extensive set of variables at both firm characteristics and corporate governance levels. Regarding firm characteristics, we control firm size (Size), firm age (Age), firm leverage (Lev), cash flow volatility (CashFlow) Rajgopal and Venkatachalam (2011), firm growth (Growth), firm financial performance (ROA), and the Book-to-market ratio (BM) on firms’ IR by affecting cash flows. As for corporate governance variances, we follow Hudson and Morgan (2024) and He et al. (2025) to control board size (Bsize), institutional ownership (INST), and top 1 ownership (Top1), constraining the governance structure and directors’ influences on firms’ IR. Table 2 illustrates all the variables’ symbols, measurements and data sources in this study. 4.3 Model Specification We employ a set of panel data regression models to test proposed hypotheses. Specifically, we specify the effect of HCAI on firms’ IR as well as the contingent factors in the HCAI-IR relationship through Eq. and Eq., respectively: disp l a y - e q-Equ 5  disp l a y - e q-Equ 6  wherein inline-eq-IEq21 represents the idiosyncratic risk of firm i in year t; inli ne-eq-IEq22 refers to the human-centric AI innovations of firm i in the observation year t; inline-eq-IEq23 is a vector containing all the contingent factors that may moderate the HCAI-IR relationship from both socio (i.e., inline-eq-IEq24 and inl ine-eq-IEq25) and technical (i.e., inline-eq-IEq26 and in line-eq-IEq27) subsystems. The vector of inline-eq-IEq28 includes all the variables for controlling firm characteristics and corporate governance that influence IR in this study. To address unobservable influences caused by time and industry, we also include the i nline-eq-IEq30, respectively. All regression analyses were performed in Stata 16.0 using the reghdfe function. 5 Results 5.1 Descriptive Statistics Table 3 reports the descriptive statistics for all variables before regression analysis. To evaluate the potential issue of multicollinearity, we computed the variance inflation factors (VIFs) for all predictors in the model. The maximum VIF observed was 1.34, which is substantially below the conventional threshold of 10, indicating no serious multicollinearity concerns among the independent variables. 5.2 Hypothesis Testing Table 4 reports the regression results of the effect of HCAI on IR as well as the moderating effects of digitalisation, operational efficiency, executive shareholdings and IT background CEO, respectively. Specifically, the regression results in Column 1 of Table 4 indicate that HCAI has a significant negative effect on IR (β = -0.024, p < 0.05). Such a result implies that each logarithmic increase in HCAI innovation output reduces firms’ IR by 2.4% points, supporting H1 proposed in our study that HCAI significantly curbs firms’ IR. The analysis results in Columns 2 and 3 demonstrate how technical factors moderate the HCAI-IR relationship. Specifically, the coefficient of interaction item between HCAI and digitalisation (β = -0.0766, p < 0.01) shows that firms’ digitalisation could strengthen the risk-mitigating effect of HCAI. Therefore, H2 was supported. However, the coefficient of interaction item between HCAI and operational efficiency (β = 0.1814, p < 0.05) in Column 3 suggests that HCAI efforts of firms with operational efficiency advantages significantly increase rather than reduce IR. Such a result profoundly rejects H3, implying that investors react negatively to HCAI innovations when firms are highly operationally efficient. This result diverges from prior findings that operational efficiency typically amplifies the benefits of technological innovation, pointing to the distinctive nature of HCAI. Further, the analysis results in Columns 4 and 5 illustrate how factors embedded in firms’ socio-subsystems moderate the effects of HCAI on IR. In particular, the interactive coefficient of HCAI and executive shareholding remains significantly negative (β = 0.0001, p < 0.01), suggesting that executive equity incentives could reinforce the risk-reducing Table 3 Descriptive statistics and correlations t statistics in parentheses. p < 0.1, p < 0.05, p < 0.01 effect of HCAI. On the other hand, IT background CEOs not only diminish but also convert the effect into risk-increasing ones (β = 0.0050, p < 0.10). This result indicates that investors react negatively to the HCAI innovation of firms with IT background CEOs. In this regard, our analysis results support H4 but reject H5. 5.3 Robustness Tests We undertook alternative measurements, time-lagged regression, propensity score matching (PSM) regression, and sub-sample regression to check the robustness of the analysis results above. First, we follow He et al. (2022) and Hudson and Morgan (2024) to apply the Fama-French four-factor (FF-4) model to IR estimation. Compared with FF-3 model, FF-4 took the momentum factor inline-eq-IEq31 into account when calculating stock returns of firm i in month m, capturing the difference between high stock performance firms and low ones. The results in Column 1 of Table 5 represent the analysis results of an alternative measure of IR (β = -0.0026, p < 0.05), which remains consistent with our main regression model. Second, we also lagged IR by one observation period, given that HCAI may exist with a time-lag from innovation to implementation. The coefficient in Column 2 of Table 5 is β = -0.0036, p < 0.01, indicating that HCAI also reduced firms’ IR lagged by one period, aligning with the main result. Third, we employed propensity score matching (PSM) to address potential endogeneity arising from sample selection bias. A dummy variable was constructed to classify firms into treatment and control groups. Specifically, firms with HCAI values above the sample mean of 0.031 were assigned to the treatment group (dummy variable value = 1). In contrast, those below the mean formed the control group (dummy variable value = 0). We then performed 1:1 nearest-neighbour matching without replacement, using the same control variables as in our main regression models. As reported in Table 6, balance tests confirm that all covariants in both treatment and control groups show no statistically significant differences in means after matching, indicating that the matching procedure achieved comparable groups. Finally, we re-estimated our main regressions using the matched sample. The results are presented in Column 3 of Table 5, which is consistent with our baseline regression models and then reinforces the robustness of our findings. Finally, we followed Tian et al. (2023) to test the robustness of our moderating analysis results via sub-sample regression. Specifically, we divided our sample firms into four pair groups according to above or below the mean values of digitalisation (i.e., High-Digital group and Low-Digital group), operational efficiency (i.e., High-OE group and Low-OE group), executive shareholdings (i.e., High-Ehold group and Low-Ehold group), and IT background CEO (i.e., High-IT CEO group and Low-IT CEO group), respectively. Table 7 shows the sub-sample regression t statistics in parentheses, p < 0.1, p < 0.05, p < 0.01 Table 6 Balance test and propensity score matching results results, which remain the same as our main moderating analysis parts. Collectively, these results of robustness checks support our main finding by mitigating concerns regarding sample selection bias, measurement error, reverse causality, and model sensitivity, thereby confirming the negative relationship between HCAI and IR. Moreover, the moderating results remain consistent in the sub-sample analyses, reinforcing the robustness of the identified boundary conditions in the HCAI-IR relationship. 6 Discussion and Conclusion This study examines whether and under what conditions HCAI shapes firm IR. Using data from Chinese A-share listed firms from 2015 to 2023, we find that HCAI is associated with lower firm-specific risk. This risk-reducing effect is strengthened by digitalisation and executive shareholding, but weakened by operational efficiency and CEOs with IT backgrounds. Taken together, these findings show that HCAI matters not only as an ethical or operational orientation, but also as a financially relevant AI strategy whose effects depend on the surrounding socio-technical context. Based on these findings, we generate several theoretical insights and practical implementations for both HCAI development and firm risk management during Industry 5.0 transition. 6.1 Theoretical Contributions First, we contribute to the HCAI literature by showing that HCAI is associated with lower firm idiosyncratic risk, thereby extending prior work from operational and social outcomes to financial outcomes. Prior studies on HCAI lay their emphasis on either its conceptualisation and design or investigating how to integrate HCAI into intelligent manufacturing processes, all of which identify HCAI as the core driver for improving operational and social performance in Industry 5.0 transition. However, the existing literature overlooks how investors respond to HCAI of their portfolio firms. As one of the primary stakeholders, investors play a vital role in acquiring financial resources and managing firm risks. In this regard, this study explores the effect of HCAI on firms’ IR, which has been extensively leveraged as an indicator of investor reactions across fields of information systems, accounting and finance, economics, and management studies. Even though prior studies also examine the same risk pattern at occupation, industry and firm exposure levels, we show that internal practices, design and governance sit closer to the risk outcome than exposure alone. Specifically, we find that HCAI situates firms’ AI strategy to respond to the requirements of Industry 5.0, as it addresses AI instinct limitations through human-centred design, ethical management, and responsible implications. In this sense, HCAI aligns firms’ AI desires with the broader interests of stakeholders, which lowers firms’ IR by ensuring sustainable and stable profitability and cash flows during technological changes. Second, we identify important socio-technical boundary conditions, showing that the risk implications of HCAI depend on firms’ digital, structural, and leadership contexts. In particular, our findings show that HCAI is not a panacea for lowering firm risks across all scenarios, given that firms’ characteristics embedded in their socio technical subsystems significantly moderate the HCAI IR relationship. Although digitalisation and executive shareholding strengthen the risk reducing effect of HCAI, we found that operational efficiency and IT background CEOs attenuate it. While prior research shows that operational efficiency amplifies performance benefits from technological innovation, we find that OE unexpectedly weakens the risk reducing effect of HCAI. This divergence suggests a boundary condition. Efficiency magnifies gains when innovation is mainly technical, but it can pose challenges when innovation involves socio-technical reconfiguration, as high OE firms often operate with lean buffers and tightly coupled routines, leaving little slack for unforeseen adjustments. Added human oversight and cross functional governance can increase transition frictions, leading investors to perceive HCAI as disruptive to operations rather than risk reducing measure. IT oriented leadership may prioritize automation and optimization, which can weaken the cross functional checks that are necessary for human centric deployment. Together, these boundary conditions support the external validity of our study and suggest that stronger socio technical alignment influences whether HCAI translates into stability as perceived by investors. Third, we contribute theoretically and methodologically by integrating SAIT with STS theory and by operationalising HCAI through a scalable patent-based measure. Existing studies explain performance through the fit between AI, routines, structures, and people, with emphasis on efficiency and ethics. We build on these studies and connect human-centric design and governance to a finance construct with established definitions, idiosyncratic risk estimated from asset-pricing residuals. Moreover, differing from related AI ethical frameworks, HCAI is firms’ internal situating approach, as HCAI embeds those principles into firms’ unique operational processes, governance mechanisms, and stakeholder relationship management. Thereby, HCAI answers how AI can be operationalised through firms’ proactive design, ethical issue management, and deployment choices during the Industry 5.0 transition. Even though our contribution is incremental, we believe it has the theoretical potential to open a direct path for IS scholars to integrate AI design, AI management, and AI governance perspectives when further delving into financial risk management in the AI era. Finally, we also provide an pragmatic measure of HCAI that is suitable for large-sample tests. The index is derived by classifying patent text with a large language model and a taxonomy of trustworthy and human-centric AI. It covers human-centred design, ethical management and responsible implications. We aggregate at the firm-year level. The approach is transparent and replicable, and more targeted than broad AI intensity or exposure counts. It supports more direct tests of socio-technical claims, and it complements text-based approaches that infer AI engagement from public filings and patenting activity in innovation research. 6.2 Managerial Implications This study offers distinct implications for corporate managers as well as investors and analysts. For corporate managers, our study indicates that HCAI could be incorporated into a firm’s situated AI strategy that not only copes with technological turbulence but manages financial risks under the ongoing AI-driven transformation. To achieve these strategic goals, we further demonstrate the specific road map for managers to develop effective and idiosyncratic HCAI through three pathways, i.e., human‐centric design, ethical management, and responsible implications. Second, we also inspire managers that the risk‐curbing effect of HCAI is contingent on organizational socio-technical contexts. As the moderation evidence points to the stronger association in firms with higher digitalisation and executive shareholdings, managers could take more advantages of HCAI for risk management through both technical (i.e., improving digital foundations) and corporate governance (i.e., optimising shareholding incentives) approaches. This aligns with the special issue’s focus on accountability and controls in organisational AI and with the socio-technical lens used in this study. For investors and financial analysts, our results provide implications for evaluating corporate AI strategies beyond mere industrial exposure or adoption announcements. Investors and analysts should look for portfolio firms’ HCAI clues through the evidence regarding human‐centred design, ethical management, and responsible deployment practices, incorporating these contextual factors into their risk assessments and earnings forecasts. Moreover, they should pay more attention to firms with operational efficiency merits as well as IT background CEOs, as these factors embedded in firms’ socio-technical systems hamper the bright side of HCAI. 6.3 Limitations and Future Directions The main limitation of our study is that it relies on China A-share listed firms from 2015 to 2023. However, institutions, disclosure norms and investor behaviours differ across capital markets. External validity is bounded. As a result, replication in other contexts is needed to test the generalisability of the findings. The second limitation is measurement. Our outcome variable focuses on idiosyncratic risk estimated from asset-pricing residuals. This captures firm volatility in returns, not operational incidents or systematic exposure. Monthly horizons and factor choices influence estimates. Future research could focus on four directions. First, a follow-up study should adopt event-based and quasi-experimental designs, for example, policy shocks, governance announcements or incident disclosures. Second, future research could be conducted to triangulate the HCAI construct by combining patent text with annual reports, governance disclosures, model cards, audit trails, and incident logs, along with a manual coding audit and inter-rater checks. Third, future studies could broaden outcomes and settings. They could pair idiosyncratic risk with operational indicators such as incident rates, customer complaints, product recalls, and cybersecurity events. Also, they can test weekly versus monthly risk windows and alternative factor models, and extend the analysis to markets with different regulatory stringency and disclosure rules to probe scope conditions. Another limitation relates to our patent-based measure of HCAI. Firms in different industries vary in their propensity to patent and in the extent to which patent filings reflect underlying innovation activity. Accordingly, the visibility of HCAI in patent text may not be fully comparable across sectors. Future research could address this issue by combining patent-based measures with industry-specific adjustments or alternative disclosure sources. Fourth, HCAI remains an evolving concept that has not yet attained universally accepted definitions. Future research could re-examine the HCAI-IR relationship by comparing our current and exploratory operationalisation with emerging connotations of HCAI as the construct matures over time. Appendix A Hyperparameter Set and Prompts Design in HCAI Identification 1. #LLM models Hyperparameter Setting Temperature=0, ensuring minimal variability in LLMs’ trail. responses. 2. Maxtokens=4,096 for ChatGPT, 8,000 for DeepSeek. and 16,384 for Qwen. 3. TopP=1, maximising diversity in the model outputs. for considering the entire distribution of token predictions without filter limitations. #Begin Task Patent Description: [Insert Patent Description Text Here] #Role Please clean your memory. The following textual data are sample firms’ technological patent description text. You are an expert patent analyst specializing in identifying cutting-edge technologies at the intersection of Artificial Intelligence (AI) and Human-Centricity within the context of Industry 5.0. Your task is to rigorously analyze patent descriptions strictly based on the provided text evidence and classify them according to the rules below. Avoid speculation or external knowledge. #Workflow For each patent description provided, execute the following steps sequentially and meticulously: ##Task 1. HCAI Determination (Execute ONLY if AI: Yes): Evaluate the AI technology described against the core tenets of Human-Centric AI (HER Framework) Judgment: If the text provides clear evidence supporting at least two of the three key HCAI features (Enhancement, Collaboration, Adaptability), conclude HCAI: Yes. If the text describes an AI technology but lacks sufficient evidence for at least two HCAI features (e.g., it focuses solely on automation replacing humans, pure algorithmic optimization without human interaction/benefit, or lacks adaptability/collaboration aspects), conclude HCAI: No. Output: HCAI: Yes or HCAI: No. Proceed to Step 2 ONLY if HCAI: Yes. ##Task 2. Rationale Elaboration and Relevant Phrase Extraction (Execute ONLY if HCAI: Yes) Carefully identify and extract key phrases or short sentences from the patent description text that directly support the HCAI: Yes classification. Output: A list of these extracted phrases/sentences. Format: Separate multiple phrases with a semicolon (;). Be concise and directly quote or closely paraphrase the most relevant text snippets. DO NOT invent phrases; base them strictly on the patent text. #Examples Example 1: Patent Text: "A novel chemical compound for enhancing battery electrolyte stability." Analysis: No mention of learning, interaction, or problem-solving cognitive functions. Core is chemistry/material science. Output: {"AI": "No", "HCAI": "N/A", "RelevantPhrases": "N/A"} Example 2: Patent Text: "An optimized deep learning algorithm for predicting stock market fluctuations using historical price data. The system operates autonomously to execute trades." Analysis: AI Yes (learning, prediction/problem-solving). HCAI: Lacks clear evidence of enhancinga specific human capability beyond financial gain (general automation focus), no described collaborationwith humans (autonomous), no described adaptabilityto individual users. Output: {"AI": "Yes", "HCAI": "No", "Relevant Phrases": "N/A"} Example 3: Patent Text: "A wearable AR system for industrial maintenance technicians. The system uses computer vision to identify equipment components and overlays context-aware repair instructions and safety warnings. It adapts the complexity of instructions based on the technician's experience level (detected via user profile and task performance) and allows voice commands for hands-free operation." Analysis: AI Yes (vision - perception/interaction, adaptation - learning). HCAI: Enhancement (augments technician's repair capability/safety), Collaboration (hands-free voice interaction, information overlay), Adaptability (adjusts instructions based on experience/performance). Evidence for all 3 pillars. Output: {"AI": "Yes", "HCAI": "Yes", "Relevant Phrases": "overlays context-aware repair instructions and safety warnings; adapts the complexity of instructions based on the technician's experience level; allows voice commands for hands-free operation"} #Output Format Mandatory: Provide the classification results strictly in the following JSON format for every patent: {"AI": "[Yes/No]", "HCAI": "[Yes/No/N/A]", "RelevantPhrases": "[Extracted Phrases; separated by semicolon OR 'N/A']"} Notes: HCAI must be "N/A" if AI is "No". Relevant Phrases must be "N/A" if HCAI is "No" or "N/A". Only extract phrases if HCAI is "Yes". Ensure JSON syntax is correct (quotes, commas, braces). 1. #Critical Reminders Evidence-Based: Your classification MUST be grounded exclusively in the text of the. provided patent description. Do not infer capabilities not explicitly described. If the text lacks detail, err on the side of No or N/A. Hallucination is unacceptable. 2. Strict HER Criteria: HCAI: Yes requires concrete evidence for at least TWO of the three pillars (Enhancement, Collaboration, Adaptability) within the HER framework. Automation alone does not constitute HCAI. 3. Step Enforcement: Do NOT skip steps. Step 2 (HCAI) is executed only if Step 1 (AI) results in Yes. Step 3 (Phrases) is executed only if Step 2 results in Yes. 4. JSON Integrity: Always output valid JSON in the exact. specified structure. Missing fields or incorrect syntax will invalidate the result. 5. Avoid Bias: Do not assume a patent is HCAI because it. mentions "AI" or "user". Scrutinize the details against the specific HER criteria. Appendix B The quantification process of firm digitalisation Digitalisation (Digital) Following prior research on digitalisation, we quantify digital transformation using text analysis to extract relevant content from corporate annual reports. Figure 5 outlines the identification process. Specifically, we first obtained the annual report texts from the CNINFO for all sample firms during the observation period6. We extracted text from the full annual report narrative, excluding tables, financial statements, and footnotes.Using the Jieba tokeniser in Python, the texts are cleaned and structured. Next, we conduct text matching based on a seed dictionary of terms related to digital technologies, as these words represent the foundational elements of digital transformation strategies mentioned in annual reports. Following prior research, we constructed and expanded the relevant dictionary using keywords associated with “artificial intelligence,” “big data,” “cloud computing,” and “internet technologies.” Finally, we identified only those texts that describe the application scenarios of these digital technologies within business operations. This is because digital transformation emphasises the integration of digital technologies with business operations rather than merely pursuing technological digitalisation. For example, a representative disclosure from Midea Group’s 2024 annual report states: “The company continuously deepens data-enabled business operations by leveraging ‘Data + AI’ technologies to deliver intelligent, data-driven solutions across core business scenarios including R&D, marketing, supply chain management, overseas markets, and ToB services” 2. Such statements explicitly reflect the integration of digital technologies into core operational processes rather than merely referencing technological adoption. Through the above process, we aggregated the word frequencies from the digitalisation dictionary to obtain the total word frequency of digital transformation-related terms. This total frequency is then divided by the total number of words in the annual report and multiplied by 100 to calculate the final of firm i in year t. Acknowledgements We are thankful to the reviewers and Editors for providing us with their valuable suggestions. Authors’ contributions All the authors have made equally substantial contributions to the conception, design of the work, the acquisition, analysis, and interpretation of data. Funding Open access funding provided by NEOMA Business School. This research is funded by the Jiangsu Provincial Social Science Foundation (No. 25GLB022) and the Fundamental Research Funds for the Central Universities (NO. 2025SK09, China University of Mining and Technology). Data Availability The datasets generated during and/or analyzed during this study are available from the corresponding author on reasonable request. Declarations Ethics Approval and Consent to Participate This research did not involve human participants and/or animals. Consent for Publication The author(s) give their explicit consent that this research was not conducted in any organization/ company. Competing interests The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.