Latest / The In-Between Tech and Trust Podcast / Systemic Trust: How We Test, Regulate & Translate AI Compliance - EP09

Transcript

Systemic Trust: How We Test, Regulate & Translate AI Compliance - EP09 · The In-Between Tech and Trust Podcast

• 0:01
  Everything we do, we do to establish and to strengthen safe
• 0:06
  and trustworthy AI in Europe. What does it take to trust a
• 0:11
  system we can see and don't fully understand?
• 0:15
  Welcome to the In Between Trust Podcast.
• 0:17
  In the podcast at the intersection of tech and trust.
• 0:21
  In this episode, I speak with Andreas Luta, Head of Strategy,
• 0:25
  Operations and Partnerships at the TUF AI Lab.
• 0:28
  We explore how trust in AI is tested, regulated, and
• 0:32
  translated, and why trustworthiness is not just a
• 0:35
  feature but a design principle. In this conversation, we talk
• 0:39
  about accountability, certification, and the culture
• 0:43
  behind code. Hi Andrea, welcome to the In
• 0:47
  Between Trust podcast. I'm so happy to dive into the
• 0:50
  intersection of tech and trust with you today.
• 0:53
  Hi Irfa, thank you for having me.
• 0:55
  I'm also looking forward to it. Nice.
• 0:58
  And let's start with like a question that I really wanted to
• 1:02
  ask you in regards to your personal but also your work
• 1:07
  relation with trust. Do you want to like it?
• 1:10
  Trade a bit on what trust means to you and like also what your
• 1:14
  perception experience is on it? And yeah, how you perceive the.
• 1:19
  Topic. For me, trust is a lot about
• 1:22
  consistency, ongoing correlation between a claim that someone
• 1:26
  makes and the outcome. So from my point of view, trust
• 1:30
  always needs a solid base so you can give trust in advance, which
• 1:35
  we usually do with people and rightly so.
• 1:37
  But if that trust is not met with trustworthy behavior, it
• 1:41
  erodes quickly. And another aspect, the level of
• 1:45
  trust we require depends on the value that is at stake.
• 1:49
  So you might just trust a stranger with $0.50, but not
• 1:53
  with a medical operation. I at least would like to see
• 1:56
  some basis for this first. And I think with AI, it's the
• 1:59
  same. There are so enormous
• 2:01
  opportunities in AI. So AI can already detect cancer
• 2:05
  cells or read traffic signs when driving autonomous.
• 2:09
  But of course it could also lack trustworthiness.
• 2:12
  So it could give you a wrong diagnosis because it hasn't been
• 2:16
  trained on your skin color, or it could misread one traffic
• 2:20
  sign for another because it's raining.
• 2:22
  And in such cases, if the stakes of a failure are that high
• 2:26
  because it could hurt or like even kill people, we need trust
• 2:30
  that's verifiable and not just assumed.
• 2:33
  And that's why testing and certification by independent
• 2:37
  third parties is so powerful because it can create the
• 2:42
  foundation for trust. And I think this is also true
• 2:45
  for low risk AI. So see and Germany, the adoption
• 2:49
  rates for AI are still relatively low and trust from my
• 2:54
  point of view can be the catalyst that change that so we
• 2:58
  can unlock the I men's opportunities that AI holds for
• 3:02
  us. Yeah, with Truth's legacy as one
• 3:05
  of the most trusted brands in Germany, we have the opportunity
• 3:08
  to earn that trust in AI as well.
• 3:11
  Amazing. And you've already like
• 3:13
  mentioned some of the examples that where distrust was created
• 3:17
  also what harmful effects this trust can have.
• 3:20
  And for you personally, was there like a moment recently
• 3:23
  where you felt trust being formed also where was fractured
• 3:27
  like on intent or maybe on also unintentionally and that you
• 3:32
  experienced and also how you then handled?
• 3:34
  It yes, there was a moment where I felt trust was fractured.
• 3:41
  I think unfortunately, right now we can see in the US how fragile
• 3:46
  trust in democratic structures is.
• 3:49
  So one moment that really struck me recently was reading about
• 3:54
  Donald Trump firing the head of the US Bureau of Labor
• 3:57
  Statistics, accusing her of faking figures after he didn't
• 4:01
  like the latest numbers on new jobs created.
• 4:04
  And you know, that's the agency that is responsible for
• 4:07
  producing the countries core economic data and firing here.
• 4:11
  And also the way it was communicated risks eroding trust
• 4:16
  in something very fundamental. From my point of view, that is
• 4:20
  the integrity of national data. And if people stop believing
• 4:25
  official statistics, we lose one of the shared reference points
• 4:28
  that help societies make informed decisions.
• 4:31
  And for me, this is a powerful reminder that trust can be
• 4:36
  fractured in a single moment, and it takes years or even
• 4:39
  decades to build. And I know there's not so much
• 4:43
  sexiness and statistics or in bureaucracy or regulation, but I
• 4:47
  personally value very much that we in Europe still have
• 4:51
  independent democratic structures that got credibility
• 4:55
  and trustworthiness fiercely. And I mean also within like your
• 4:59
  role in the tooth and specifically with the eyelet.
• 5:02
  This is something that you like strive for, but this is also
• 5:05
  something that you like enable and collaborate with, right?
• 5:08
  And I want to dive a bit more into the creation of the lab
• 5:11
  because I do think it's also quite special living between the
• 5:15
  different kind of TIFF structures.
• 5:17
  And so maybe you can share like the background on the creation
• 5:20
  path and also what you and your teams are exactly trying to
• 5:24
  shift within the ecosystem itself.
• 5:26
  Yes, thank you for the question. Yes, the Turfway I Love was
• 5:30
  founded in 2023, so relatively young, as an independent joint
• 5:36
  venture by TUF companies, TUFSU to Fineland, TUF N to Fasen and
• 5:41
  TUF Turing. And many people in Germany think
• 5:45
  that there's only one TUF, but they are separate companies that
• 5:49
  compete with each other. And nonetheless, it became
• 5:52
  evident that AI certification as a cent for a future topic that
• 5:57
  is becoming highly relevant for all of them.
• 5:59
  And therefore they decided to found the TIF AI lab together.
• 6:04
  You know, the TIF companies are already certifying many products
• 6:08
  also with AI components such as AI based medical products or
• 6:13
  industrial machinery using AI. And with the AIA coming in,
• 6:18
  there was a necessity to have certification for AI and
• 6:22
  compliance. And therefore we were founded.
• 6:25
  And the Tough AI Lab aims to translate these regulatory
• 6:28
  requirements for AI into practice.
• 6:31
  And our main goal is to make Europe a hotspot for safe and
• 6:36
  trustworthy AI. Yeah, the AI is a high level
• 6:41
  regulatory text. And you know, like even the
• 6:43
  upcoming standards would just give you a direction, not
• 6:47
  exactly specify how AI can be assessed in the testing and
• 6:51
  certification practice. And what we do in the Tof AI lab
• 6:54
  is to develop conformity criteria that are actually
• 6:58
  quantifiable and suitable test methods for AI assessment.
• 7:02
  And we're building the two AI assessment framework that is
• 7:07
  consisting of modular components and that can be used by our
• 7:11
  shareholders. And some parts of it are also
• 7:14
  openly published, such as our AI at Risk Navigator, which is
• 7:18
  published as a free online tool to assess one's risk according
• 7:21
  to the AI Act. You already mentioned, we are
• 7:24
  contributing also in the ecosystem.
• 7:26
  We have several research projects on AI assessments.
• 7:30
  So we're part of the, we are part of the German Lighthouse
• 7:33
  project, Miss Yong Kai, and part of the EU project TAFLS that is
• 7:39
  working on agile certification processes.
• 7:41
  And then we have projects on automotive perception and one on
• 7:46
  regulatory sandboxes. And then we are supporting the
• 7:49
  development of standards and norms for AI systems.
• 7:52
  And everything we do, we do to establish and to strengthen safe
• 7:57
  and trustworthy AI in Europe. And I do think like within your
• 8:01
  role and what you just described, it's like a role that
• 8:04
  you see across sectors, across cultures, and also across silos.
• 8:09
  And then again, it's also like a friction point being not only in
• 8:14
  your role within your organization, but being at the
• 8:16
  centre part of exactly this ecosystem is something that is
• 8:20
  super interesting, but definitely also super like
• 8:24
  ambitious to dive into. How do you see it from your
• 8:28
  person and also from your organization being part of that
• 8:32
  ecosystem? Yeah, that's true.
• 8:34
  Like we're often wondering how ambitious it really is.
• 8:38
  I personally believe we are working on one of the most
• 8:41
  exciting but also most complex fields at present, because the
• 8:47
  AI Act is is the first horizontal regulation of AI and
• 8:51
  it regulates AI systems across several industries.
• 8:54
  Yeah, let's just look at what I just said.
• 8:57
  So we have a fresh new regulation with a very high
• 9:00
  dynamic because as I already mentioned, standards and norms
• 9:04
  in the field are still evolving and the harmonized standards
• 9:07
  have still to be published. And at the same time, we're
• 9:10
  already experiencing a strong push back because the
• 9:14
  geopolitical mood has shifted. And then we have the rapid
• 9:17
  technological progress of the subject.
• 9:21
  So AI has such fast and frequent iteration cycles and these bring
• 9:26
  possibly significant changes. And also AI systems may continue
• 9:30
  to learn even after deployment. And all of this has somehow to
• 9:34
  be integrated into certification processes because no one wants
• 9:38
  to recertify their product every few weeks, of course.
• 9:41
  And then I said, yes, the AI regulates AI systems, but what
• 9:46
  exactly is the AI system? So even with the guidelines
• 9:50
  published by the EU, the term AI system covers completely
• 9:55
  different areas of application and of technical implementation.
• 9:59
  There's a world of a difference between a rule based expert
• 10:02
  system and a deep learning model.
• 10:04
  And and then we have two different industries.
• 10:06
  So a medical AI system for tumor detection has completely
• 10:10
  different requirements and it's testing and assessment than an
• 10:14
  assistant system in a car or a like cross sectoral language
• 10:17
  models such as TCPT. And very often the testing
• 10:21
  criteria such as robustness of bias demand the involvement of
• 10:26
  the main experts from the respective area in the testing
• 10:29
  procedure. So we are currently seeing and
• 10:32
  that's like my perspective and my role that even absolute
• 10:36
  experts in the field are still discovering how big the
• 10:39
  challenge ahead of us is. And for me, it's absolutely
• 10:42
  amazing to be able to contribute to such an important task for
• 10:47
  our society. And I guess it's not only
• 10:49
  important, but it's also shedding light on things that
• 10:52
  are potentially not even explored.
• 10:54
  And when you just like described in regards of like the different
• 10:59
  areas of also usage and so on, like the one hand side and in
• 11:03
  particular to AI, right? But then again, compared to
• 11:06
  technologies in general, I do think that there are like those
• 11:10
  differences as well. And if we now come back to like
• 11:14
  the common denominator trust, why is it different for AI
• 11:19
  systems than other technologies, in particular from your point of
• 11:23
  view? Yeah, I think that any
• 11:26
  technology is usually first met with skepticism and needs some
• 11:32
  time to establish trust. And, and I recently learned on
• 11:35
  our panel that it took about 10 years for ultrasound technology
• 11:41
  to gain the necessary acceptance for its widespread use and
• 11:44
  prenatal diagnostics because before that, the trust and the
• 11:48
  technology just wasn't there. And if we take these 10 years as
• 11:52
  a comparison, then trust and chat, CPT and the likes evolve
• 11:56
  very fast. But why is trust and AI
• 11:59
  different? Traditional technologies are
• 12:01
  usually deterministic. So if you push the same button
• 12:04
  under the same conditions, you will get the same result.
• 12:08
  And you also know why you are getting this result.
• 12:10
  And with a, in particular modern machine learning models, we're
• 12:14
  dealing with systems that are probabilistic, adaptive, and
• 12:18
  often a pack. And that means we often do not
• 12:21
  know why a certain input leads to a certain output.
• 12:26
  And the behavior of continuously learning systems even can change
• 12:30
  over time, sometimes in ways that we don't fully anticipate
• 12:34
  or understand. And you just can't test them
• 12:37
  once and declare them safe forever.
• 12:39
  So trust in AI has to be continuously earned through
• 12:43
  ongoing testing, ongoing monitoring and transparency.
• 12:47
  I can see that as like a quite the challenge for a lot of
• 12:51
  organizations of how to follow through on that, right.
• 12:54
  And I think that then again, on the one hand side, there is like
• 12:58
  this tension between exactly like having this capability of
• 13:02
  seeing AI incorporated into the processes and then again, in
• 13:08
  comparison to that, the societal comprehension on it of how they
• 13:13
  also perceive it. Do you also experience or from
• 13:16
  your point of view see this tension between the capability
• 13:21
  and the comprehension of it? How do you like experience it
• 13:24
  currently? Yeah.
• 13:26
  From my point of view, that is mainly a speed gap.
• 13:30
  So AI capability is advancing at an exponential speed, while the
• 13:35
  societal comprehension, so how well the general public
• 13:38
  understands AI in general and its capability and its risk
• 13:42
  moves much more slowly. So I can already today do things
• 13:47
  that most people don't fully grasp.
• 13:50
  And when that understanding of a technology lags too far behind,
• 13:54
  it can create anxiety and sometimes even mistrust against
• 14:00
  technology. And others totally underestimate
• 14:03
  the technology and think it will not become relevant for their
• 14:06
  field. And in this case, we're missing
• 14:09
  out on important innovation. And then at the same time, of
• 14:12
  course, there's a lot of hype and some people think that AI
• 14:15
  can magically do everything perfectly right now.
• 14:19
  And I don't see any need for regulation or control.
• 14:22
  And of course, reality falls short.
• 14:24
  And for me, a closing that gap means making AI not only
• 14:29
  technically more reliable, but also being transparent about its
• 14:32
  limits and its risks and balance, balancing
• 14:36
  trustworthiness and innovation. Overall, certifying for trust is
• 14:41
  something that's super relevant, but again, it can also be like
• 14:46
  quite complex. And what does it actually mean
• 14:50
  according to you and like your experience in your role?
• 14:53
  And how can testing be also more than just like being perceived
• 14:58
  as compliance? Yes, AI systems are very
• 15:01
  complex, but the good good thing is they can be assessed and
• 15:06
  certified by the two companies if they use the right
• 15:09
  requirements and tools and processes which we are
• 15:12
  developing. AI assessment is not based
• 15:16
  solely on the output of a model of system, but rather on a
• 15:20
  structured analysis of the entire development and life
• 15:23
  cycle. So how was the system trained,
• 15:26
  what data was used and also what control mechanisms are in place
• 15:31
  during its operation? And as I mentioned before, we
• 15:34
  have built that modular TIF AI assessment framework.
• 15:38
  It's a very comprehensive framework for AI assessment and
• 15:41
  our opinion so far unparalleled in terms of its completeness and
• 15:46
  systematic approach. And we are currently validating
• 15:49
  it in practical use cases. So we cannot certified trust in
• 15:55
  itself, but we can certify compliance with the law and if
• 15:59
  we have solid laws that should be enough to establish the trust
• 16:04
  needed. Also for the company side, it
• 16:06
  wants to mitigate regulatory and reputational risks.
• 16:09
  And you just also ask how testing can be more than
• 16:13
  compliance. And 1st, I have to say on that
• 16:16
  compliance is also much more than just testing because
• 16:19
  there's a lot of AI testing ongoing out there, but very few
• 16:24
  companies do the full link to the exact requirements of the
• 16:28
  regulation. So in particular of the AI Act
• 16:31
  and sector regulation and the respective norms and standards
• 16:34
  for all the different AI systems.
• 16:36
  Therefore, I, I personally think that not too many companies on
• 16:40
  the market can already offer full AIA compliance, at least
• 16:44
  for high risk AI systems. And yeah, back to the to the
• 16:49
  original question, how can testing be more than compliance?
• 16:55
  There is already very strong demand for certification from
• 16:59
  the industry because many companies have recognized that
• 17:03
  certification can strengthen trust in their products.
• 17:06
  And there are a lot of companies that see a that see a
• 17:09
  competitive advantage in developing or deploying
• 17:13
  trustworthy applications. And beside of that, I think that
• 17:19
  testing and certification are really amazing runways for
• 17:25
  higher quality. So from my point of view, any
• 17:28
  company should benefit from AI systems that are robust and
• 17:33
  cybersecure and perform their task the way intended.
• 17:37
  And for this station is key. And you just touched upon like
• 17:41
  the technological infrastructure and like the build up of the
• 17:45
  systems. And you also showcased some
• 17:48
  examples of how these are being built or perceived.
• 17:51
  But what I would be curious about is your point of view on
• 17:55
  how trust can also be built in culture and not just in code.
• 18:00
  What is your opinion on that? Yeah, I'm not a developer myself
• 18:05
  or and we are also not developing AI, but I of course
• 18:09
  have like a personal opinion on that because I think that trust
• 18:15
  in AI is not exactly a matter of clean code or accurate models.
• 18:20
  But it's like really about the culture in which AI systems are
• 18:24
  conceived, built and maintained. So the keyword here is
• 18:28
  trustworthiness by design. So you can have the most
• 18:31
  technically sound AI, but if you have a culture that ignores
• 18:36
  ethical concerns and transparency or inclusivity,
• 18:40
  then it will be highly unlikely that the eye turns up
• 18:43
  trustworthy. And for me, building trust and
• 18:46
  culture means that you make responsibility part of the
• 18:51
  everyday decision making so that you train your team to spot
• 18:57
  potential risks on their own, which is also true for any
• 19:01
  company, not only AI company very helpful and that leaders
• 19:05
  encourage questioning their assumptions.
• 19:08
  And also that organization welcome external audits instead
• 19:14
  of avoiding them because they might know that they can help
• 19:18
  them to increase their quality. And for me also trust in culture
• 19:23
  means bringing diverse perspectives into the room.
• 19:26
  So you don't have just engineers but also ethicists are the main
• 19:31
  experts and that you build everything with a human centric
• 19:34
  approach. And you've now just mentioned
• 19:36
  like diverse teams. And for a diverse team, you
• 19:39
  would need a shared language. And I assume that this is also
• 19:43
  one of the most complex things to establish.
• 19:47
  But what I would be interested in hearing is why you think a
• 19:51
  shared language or like a diverse setup is so vital when
• 19:56
  it comes to like regulations, ethics and the oversight, and
• 19:59
  what organizations will be benefiting from if they follow
• 20:03
  through on it, specifically for AI as well.
• 20:06
  Yeah. I think we already have a very
• 20:08
  diverse landscape when it comes to AI regulation and its
• 20:15
  implementation. So we have policymakers and we
• 20:19
  have the testing institutes and we of course have the
• 20:23
  manufacturers and deployers of the AI.
• 20:26
  So we already have like a super diverse landscape and I
• 20:31
  personally think it's super important that they have a
• 20:35
  common language that they aligned on because the term
• 20:40
  artificial intelligence is a umbrella term.
• 20:45
  So it covers extremely diverse areas of application and
• 20:50
  technical implementation. So artificial intelligence or AI
• 20:54
  can mean a large language model such as Cha chi PT, or it can
• 20:58
  mean the assistance system in the car or it can mean medical
• 21:02
  computer vision system. Everything is AI today.
• 21:06
  And you'll see that also the regulator of the AI Act subsumed
• 21:11
  them under the term AI systems. But there have very huge
• 21:15
  differences how these systems are assessed or or have to be
• 21:19
  assessed or have to be tested. And of course you can't start a
• 21:22
  new assessment process from the beginning for any single new AI
• 21:27
  system. And therefore we need clear
• 21:29
  categories that group similar application because otherwise
• 21:33
  without such categories we cannot systematically assess
• 21:37
  risks or select appropriate testing methods in a targeted
• 21:41
  manner. And therefore, I believe we need
• 21:43
  a common language among experts, experts in regulation and
• 21:47
  manufacturers and also in testing institutes in order to
• 21:51
  implement the AI effectively and scalable.
• 21:54
  And to fill this gap, we have developed the TIF AI system
• 21:59
  Taxonomy and the TIF AI Assessment metrics.
• 22:02
  Could you maybe briefly walk us through both just from a high
• 22:06
  level perspective of why you developed it and what people
• 22:11
  benefit from if they use it? Yes, I can.
• 22:14
  Do they both enable this common language of experts.
• 22:17
  So you could think of the TIF AI system taxonomy of a
• 22:23
  categorization system such as a closing sizes system.
• 22:27
  And then the TIF AI assessment matrix is the walk in wardrobe
• 22:32
  where you can find the fitting closest.
• 22:34
  So the assessment resources for your system.
• 22:38
  So I said you couldn't start from scratch anytime.
• 22:41
  So everybody is different, but the industry has developed some
• 22:46
  closing sizes so you know what to look for.
• 22:48
  And in in our case, if you know which exact AI system is
• 22:54
  assessed, then you will know which shelf of our AI assessment
• 22:58
  metrics you should look into. Yeah, this was like a metaphor.
• 23:03
  And now like I explained what what we exactly did.
• 23:07
  I already said that it established a common language by
• 23:11
  categorizing AI systems and for us the function of of the
• 23:16
  systems is paramount and that is why our taxonomy categorizes AI
• 23:21
  according to three clear criteria.
• 23:23
  So the first is the task, the second is the input and the
• 23:27
  third is implementation. And this allows us to
• 23:30
  distinguish between the task the system performs.
• 23:35
  What input format, So picture or text or voice format does it use
• 23:41
  and how is the task technically implemented?
• 23:44
  And out of these three categories or three criteria, we
• 23:50
  have a transparent and functional categorization system
• 23:54
  that places similar systems in the same category.
• 23:59
  In turn, this means that they can be tested in the same way.
• 24:02
  Let me give you an example for that.
• 24:05
  If you have an image classification, so you have an
• 24:08
  X-ray model that labels images either on tumor or no tumor, and
• 24:15
  you have on a social media AI system that gives you the
• 24:21
  sentiment of a social media post.
• 24:23
  So is it positive or is it negative or is it neutral, then
• 24:28
  they seem to be very different. But in fact, they both label the
• 24:33
  input given to them to one of certain labels.
• 24:38
  So both of them are labeling tasks.
• 24:40
  And with that, with that knowledge, we can also know that
• 24:45
  we can the same metrics. So in that case, amongst others,
• 24:49
  accuracy or precision and recall to judge their quality even
• 24:54
  across domains. And also our taxonomy helps us
• 24:58
  to make a difference between AI systems visible.
• 25:02
  So at the same time, their evaluation remains comparable
• 25:07
  across the industries, across the technologies and the areas
• 25:10
  of application. Combined with our TIF AI
• 25:14
  assessment metrics, which we already published in 2024, AI
• 25:18
  system taxonomy forms the basis for scalable AI assessment.
• 25:22
  That is fascinating in such that you've already been at the
• 25:27
  forefront for quite some time and labeling and also certifying
• 25:31
  it now. Could you it trade a bit on like
• 25:35
  the practical benefits that organisations have when they use
• 25:39
  it in terms of like the reflection aspects for
• 25:43
  themselves, but potentially also for like their external
• 25:47
  stakeholders, like clients or ecosystem partners?
• 25:50
  Yeah, I think we should build that common language because it
• 25:58
  makes it easier if you're a manufacturer or a deployer of an
• 26:02
  AI system to have it ready for compliance and have it ready for
• 26:08
  certification if that's what you wish for or you have to do.
• 26:13
  So I think for, for companies that develop or use AI, it's
• 26:19
  very helpful to use that taxonomy in their repository.
• 26:23
  And then they also know which category they have and can also
• 26:28
  hand it over to some testing institute or certification
• 26:33
  institute and then have the similar categories assessed in
• 26:39
  the way they need. So such as you would go to A to
• 26:43
  a shopping centre and only try the closest that you know could
• 26:47
  fit you. They already know what to hand
• 26:50
  over to have it assessed in the way that this specific AI system
• 26:56
  needs. Great.
• 26:57
  So it's really about like finding common ground and make
• 27:00
  information exchangeable. Amazing.
• 27:02
  I look forward to having you back in one of the future
• 27:05
  episodes to see your experiences with it.
• 27:08
  Definitely, that would be amazing.
• 27:10
  And as we come to close of our episode for today at least, I
• 27:15
  want to dive in some moments with you because in each like
• 27:17
  episode that I have with my guests, I have so-called in
• 27:21
  between moments. These are like moments when I
• 27:24
  have like a one word reflection that you would now answer like
• 27:27
  intuitive, personal and basically from your top of mind
• 27:31
  and in the moment and kind of response.
• 27:34
  And I ask in the first part of what trust feels to you like.
• 27:40
  I'd say reliability. And what does oversight feel to
• 27:44
  you like? Responsibility.
• 27:46
  And if you had to say what AI is to you, what would you say?
• 27:50
  A catalyst. What do you think we need more
• 27:52
  of? Humanity.
• 27:54
  Of that, what do you think we need less of?
• 27:56
  Excuses to innovate? OK, let's dive into that topic
• 27:59
  another time because I will be curious to hear your excuses
• 28:03
  that you received. And then lastly, one word you're
• 28:07
  currently in between of. Hopefully growth.
• 28:09
  And then I want to close with one reflective question because
• 28:13
  I really enjoyed our conversation.
• 28:15
  But I do think that our listeners would be curious to
• 28:18
  hear from you if you could leave them with one shift in how they
• 28:22
  think about AI and trust. What would that be?
• 28:25
  It is very important to me that innovation and trustworthiness
• 28:29
  are not contradictory, so the innovation that I want to see in
• 28:33
  Europe is trustworthy and puts people at the centre.
• 28:37
  What a great way to end a great episode.
• 28:39
  Thank you so much for all the insights that you shared, for
• 28:43
  the reflection part that we dove into, and for all the outlooks
• 28:47
  that we have to get back together and see and the
• 28:50
  progress that you made with your assessments, taxonomies, and
• 28:53
  also implementations with your team and beyond.
• 28:56
  Thank you so much for joining us.
• 28:58
  It was a pleasure having you. Yeah, thank you very much.
• 29:00
  I also enjoyed about it very much and also looking forward to
• 29:03
  the next interview or any other talk we'll have.