- 2 days ago
"As AI systems increasingly shape decisions, behaviors, and access to opportunity, responsibility is no longer an afterthought but embedded in how these systems are designed, trained, and deployed.
But how are values translated into models in practice? Who defines what these systems optimize for and how do those choices distribute power, risk, and accountability? This session explores what it truly means to build responsible AI, from technical design to governance, and what it takes to ensure fairness, accountability, and long-term trust at scale."
But how are values translated into models in practice? Who defines what these systems optimize for and how do those choices distribute power, risk, and accountability? This session explores what it truly means to build responsible AI, from technical design to governance, and what it takes to ensure fairness, accountability, and long-term trust at scale."
Category
🤖
TechTranscript
00:08Hello, welcome everyone to this panel about when AI systems decide, where does responsibility
00:16reside? I'm very honored to be moderating this panel, which is the second time on this stage,
00:22so very excited about to see how this conversation will move forward.
00:27I would like to introduce myself, for those of you who weren't here before, I'm Mujana Skari,
00:33I'm the co-founder and CEO of Thousand Faces, which is a funding platform for women building
00:39impact-driven companies. I am also co-founder of Women in AI, which is a global non-profit,
00:47started in France, but actually we have today 20,000 members across 150 countries, purely
00:53working on eliminating the gap in artificial intelligence in terms of gender diversity.
01:00And I am very delighted today to mother this panel, because this is one of the most important
01:07topics where we are heading with the development of AI, which is accelerating, and who is responsible
01:15when something goes wrong, and it is going wrong. So today, AI systems are making decisions,
01:23not assisting humans in making decisions anymore, but actually themselves they are making decisions.
01:31And some of the decisions are happening in milliseconds, and before even there is a human review happening.
01:43When something goes wrong, the question we haven't asked is, who is responsible?
01:49We have ethic councils, we have roadmaps, we have the EU AI Act, and the regulations are updating,
02:01and our dear policymakers are working so hard to make that happen.
02:04But the reality is that the regulations are lagging behind. There is a huge gap between where we are
02:12standing with what we know about the regulations and frameworks, and where the technology is actually
02:18right now. So first of all, I want to introduce our panelists. Today we have a very unique panelist of
02:26four amazing people from infrastructure, enterprise, research, and deployment of AI systems.
02:34I would like to introduce Arianna Legovini, Director of the Development Impact Department at the World Bank.
02:44She has founded over 20 years ago to embed rigorous evidence and experimentation into development policy.
02:54She leads Impact AI, which is an AI-powered system helping policymakers apply global evidence
03:02to design better policies. Welcome, Arianna.
03:07We have Bruno Zerbil. Bruno is Chief Technology and Innovation Officer at Orange.
03:14Before Orange, he was CTDO at Schneider Electrics, and before that, he led the Yahoo's transformation
03:22into a cloud-native platform. He now sits at the heart of the Europe's largest digital infrastructures,
03:27and has built Orange AI ethics council and responsible AI roadmap. Welcome, Bruno.
03:35The next person is Mark. Mark Roberts. You, Mark, lead the Capgemini's AI Future Lab.
03:43You have a degree and PhD in AI, and you're one of those veterans. We were just talking at the
03:49backstage
03:49that 30 years ago, you created your first AI model.
03:55And so you have over two decades on the front end of technological innovation,
04:00and you have spent your career developing the world's most forward-thinking organizations,
04:05unlocking real business values from AI.
04:08And last but not least, Antoine Bosleau. Antoine is Assistant Professor at EPFL,
04:16where he leads the NLP group working on AI systems that can reason about human and world knowledge.
04:25He is on the steering committee of the Swiss AI Initiative, and call it the Apertise Project,
04:31which we're going to talk about. It's very interesting.
04:36It's Switzerland's first large-scale, open, multilingual language model.
04:41Welcome, everyone.
04:44So, when we design AI systems today, what does it actually mean to build them responsibly?
04:51This is the first question I would like to start with.
04:54So, Bruno, I would like to start with you.
04:58Orange has built a formal, ethical council and responsible AI roadmap,
05:03and you have been a big part of it.
05:05And it's been, as you mentioned, it's been a deliberate choice not to wait for regulations
05:11to tell us how to act, but we're actually being part of it.
05:14What does responsible deployment actually look like, in your perspective,
05:20from inside a company operating at scale?
05:24So, thanks for the question.
05:25I'm very happy to be here with this panel today.
05:29Orange is a very large company.
05:32We actually serve 340 million customers, and we operate in 26 countries.
05:39So, clearly, we don't have one reality to some extent.
05:42We have, like, different realities, clearly in France and Europe,
05:45but also in Middle East and Africa.
05:47And when we think about responsibility, the first thing that we did was to come up with a set of
05:54core values
05:55and to refine those values, not to change them, refine them with external contributions,
06:01so thought leaders like professors, philosophers, people like that.
06:05That's the reason why we created this AI Ethic Council that is very important.
06:09Once you have that, it gives you a starting point, but it doesn't give you all the answers.
06:16The second thing we did was we defined a set of objectives, and we defined what was responsible AI all
06:24about.
06:24We said it has to be frugal.
06:27So, for instance, it means we have to use the most efficient model, and we know we all struggle with
06:33that, myself included.
06:34I tend to always use the most powerful model even when I don't need it, but I'm always afraid of
06:39not having the best tool.
06:40But we decided to enforce that and create tools that help us make sure we use the best model.
06:45The second thing we did is make everybody at the company aware of CO2 emission and cost impact, so you
06:53created that kind of culture.
06:54And then we asked this question that we have to ask, do we need it everywhere, and how are we
07:01going to use it?
07:02So we made sure that it's not about replacing humans.
07:05And the third pillar was what we call humans in the loop.
07:11We made it very clear that as much as we believe in agentica and automation, what have you, we had
07:16to identify very precise control points and make sure that humans will have the final say.
07:25Thank you so much.
07:28Coming to you, Arianna.
07:30For you, responsible AI begins not with the model, but with the decision being supported and the outcome being thought.
07:39Governments are making decisions every day on health, water, education, crisis response.
07:46What does responsibility actually mean when the question is whether those decisions have an impact and, you know, that can
07:57improve people's lives?
08:01We take a developmental approach to thinking about the introduction of AI in economies and countries, thinking about not only
08:13whether the models are efficient and they're doing the right thing, they're transparent and biased and so forth and so
08:18on, but actually whether they deliver better outcomes for the people which they should be serving.
08:24And so we go in the field to understand what is the actual impact of introducing a health LLM on
08:33health outcomes in Nigeria or going to schools in Kenya and introducing AI tutors to understand how to operationalize the
08:43idea of greater technology and greater opportunities.
08:47Injecting a lot of cognition into systems that have poor cognition to start with, poorly trained teachers or poorly trained
08:57community workers.
08:59The idea of regulating is something that is slightly problematic because regulation often means that some experts come together and
09:13take decisions ex-ante and create some static rules.
09:16And the stricter the rules, and the stricter the rules, the better.
09:19But in reality, we know very little as we introduce technologies and innovations and especially AI moving so fast with
09:31the adoption rate of AI in two years equivalent to 15 to 30 years the adoption of other technologies.
09:37So this is picking up really quickly and the question is how do we learn how to regulate in a
09:45responsible way but in a dynamic way by learning, testing, adapting over time and getting much greater insights into what
09:57are the benefits or the dangers, the risks.
10:01How do we ensure safety in the deployment?
10:05So, you know, when I walked into Kenya many years ago, we looked at health regulation and when we collected
10:13data, we realized that if we were to implement the regulation as it was designed, we would have had to
10:20close 97% of facilities.
10:23So clearly, that would not have increased access to health, it would simply shut down the system.
10:30And now we instead start asking the right questions.
10:35What are incentive compatible regulations that help the system improve over time and deliver better results?
10:41And I think of AI just the same and even more so than other cases of regulation, where we know
10:46so little, the technology is moving so fast, and we can even use AI to help us guide our regulatory
10:55frameworks forward.
10:57Thank you so much.
10:58Yeah, it's very interesting.
11:00We were talking about that, the Kenya, and then we're going to talk about also the other experiences that you
11:06had.
11:07Mark, coming to you, we had a little chat, and you had talked about that ethics and safeguards shouldn't be
11:15seen as restrictions, but they are accelerators of trust, and you're a believer in that.
11:22So, you know, coming from the Capgemini's perspective, which is basically, you're working with many, many organizations, it's a network
11:35of organizations spreading this, the use of the technologies and building these technologies.
11:41What's actually missing right now?
11:44Is it the technical guardrails?
11:48Is it the legal clarity or something more fundamental about organizations understanding their own responsibility?
11:58What do you think is missing?
12:00Okay, so, I mean, I think there's good answers to all of those points.
12:04I'll pick one in particular, which is, I guess, a slightly more technical point.
12:08I think we often, I mean, everyone wants to control AI.
12:12Everyone wants to be able to put good guardrails, good safeguards on AI.
12:16But I think when we actually try and do that, we find it's much harder than just, it's much harder
12:21than just asking it nicely, to tell an LLM, please don't do this, or please don't do that.
12:27The thing that's actually missing underneath, though, is that those language models, for example, they are just language models.
12:34They don't understand concepts very deeply.
12:37They don't understand the real world that we operate in.
12:39So there's a big shift in the field at the moment to focus a lot more on world models and
12:45this idea that we have a more explicit representation of the concepts that matter in making a decision.
12:53And that doesn't just mean the physical world, although it might do sometimes.
12:57It also means legal issues, ethical issues, societal norms, technological issues, domain-specific issues.
13:06So, I mean, generally, we could argue that we can't trust a model unless it speaks the same language that
13:14we do, unless it explicitly sort of represents the same concepts that we do.
13:18And at the moment, there's a mismatch.
13:20LLMs represent something deeply mathematical and statistical, but we want to govern them using sort of high-level concepts up
13:27here.
13:28And that mismatch is causing problems at the moment.
13:31So it's no accident that there's now a big push into world models.
13:36I think upstairs right now, Jan LeCun is talking about world models on the main stage.
13:41And that doesn't mean that LLMs don't work.
13:44They're brilliant.
13:45They're amazing.
13:46They have amazing capabilities.
13:47But we need to ground them in our reality if we're going to be able to trust them.
13:52Thank you so much.
13:55Antoine, so let's talk a little bit about your experience with Apertus.
14:01Your whole approach with Apertus is built on transparency, like radical transparency.
14:07We had a discussion about that, that you believe that this is absolutely mandatory to have it, and there are
14:14benefits to that.
14:15So could you talk a little bit on that?
14:16Is transparency a prerequisite for responsibility?
14:23So thanks for that great question.
14:26I guess the short answer is no, I don't think that transparency is a prerequisite for responsibility.
14:31Perhaps a spicy take, but I think that there's many institutions out there that behave incredibly responsibly, but that aren't
14:38necessarily completely transparent about how they integrate responsibility and ethical design into the technologies they create.
14:45In many cases, it's because they don't want to show the full recipe.
14:48In other cases, it's to maintain competitive advantage.
14:51But at the end of the day, they can still behave quite responsibly, even without putting transparency at the forefront.
14:57The big caveat to that, though, is how users receive that responsibility.
15:03At the end of the day, users generally want, in a very simple way, three things out of technology.
15:09They want it to work.
15:11They want it to not not work and harm them potentially.
15:14And they don't want to feel uncomfortable using it, because they feel like it may harm them or they feel
15:20like it may not align with their values.
15:21I would say transparency is absolutely key for that third one, and that that is ultimately the only way in
15:28which you can give users the comfort to use a particular product.
15:31And in our case, you know, I think we were in kind of a privileged position that responsibility and ethical
15:37design were kind of the reason why we decided to develop our own LLM.
15:42There's amazing proprietary models out there you can talk to through an API.
15:47There's amazing open weight models that you can deploy, you know, for your for your own context.
15:52What was kind of missing, we thought, is, you know, an LLM that makes it very clear in how it
15:58was designed, what went into it, what type of data that it was actually trained on.
16:02And that was relevant for a global community as opposed to just a U.S.-centric one.
16:06And so we really wanted to go and fill that gap when we created, and that made transparency a requirement,
16:12because that was exactly the special sauce that we were trying to bring into it.
16:16And so for us, it was absolutely critical.
16:19But I think it's a pretty nuanced subject to say that, you know, you absolutely need it for all responsible
16:24design.
16:25Thank you so much.
16:27Arjen, you're looking at me.
16:28Do you want to share something on top of that?
16:30No, just a comment on Mark's statement.
16:36The same misalignment is with institutional development.
16:40So we see AI zooming through, and the institutions are actually lagging behind.
16:48Even Anthropic says that now the barrier are people reviewing Claude's code.
16:55I mean, even at the engineering level, forget about bureaucracies, development, institutions.
17:02So I think of the greatest opportunities in low-resource environments where the constraints are so severe,
17:11and the injection of cognition can be quite transformational.
17:14But how are we going to develop the institutions that support the introductions of AI and allows all these different
17:25professions to be upgraded?
17:29My example, my critical example, you mentioned that we developed Impact AI.
17:35The idea of Impact AI was actually to introduce cognition for policymakers.
17:41Because we have a vast amount of evidence out there of the interventions that are very effective,
17:48we can price different type of interventions to reduce poverty, reduce mortality, to improve people's lives.
17:56At the same time, policymakers continue making decisions based on priors, status quo bias, limited cognition.
18:07And so the idea is how do we transform the availability and the process through which we don't replace decision
18:15-making.
18:15I don't think we are replacing decision-making, but we're informing that decision-making in a dynamic way
18:22to improve the way decisions are made in the world.
18:26I remember we just talked about that there are cases where it's even better if the decision-makers are using
18:35the AI's recommendation
18:37rather than making the decision themselves.
18:39We just had that chat in the backstage.
18:41In the previous panel, one of the panels shared that there are two cases of lawsuits, basically, against two doctors.
18:54One has listened to AI recommendations, one has not, and both of them are real cases.
19:00So what do we do?
19:02How can we know?
19:04How can we decide how to go about it?
19:07How can we trust the AI?
19:08And different LLMs, different LLMs actually disagree on some of those cases.
19:14A famous case about somebody suffering from anorexia, and different LLMs had different valuation of protecting her life
19:25versus protecting her rights not to be treated.
19:29So just like doctors, LLMs also have their reasoning and their way of thinking, which we need to understand.
19:36You know, transparency is good, but we actually do not understand how LLMs think and come to conclusions.
19:47I'll just briefly add that there's a great quote, and I can't claim credit, it's not mine,
19:52but the quote says that when creation becomes cheap, judgment becomes expensive.
19:59And I think that's exactly what Ariana just described, that we now have the ability to create infinite amounts of
20:05content,
20:06infinite numbers of decisions.
20:08And it just highlights the fact that that was never the problem, right?
20:12The problem is always how do we choose the right decision given the options in front of us?
20:17What critical thinking skills?
20:19How do we insert expert judgment into that process at the right time to make the right decision?
20:25Thank you so much for that.
20:28I think that, you know, we're touching on the next topic I wanted to bring up,
20:32which is that how we ensure accountability.
20:34And when something goes wrong, who actually holds responsibility?
20:41So, Bruno, I would like to come to you.
20:46When an automated system deprioritizes a hospital connection, let's say, during a traffic spike,
20:54or it flags a legitimate transaction as fraud, basically when AI goes wrong,
21:01that decision, it happens in milliseconds without human intervention, without any review,
21:06as Ariana also was mentioning.
21:08If that decision is wrong, who answers for this?
21:13How does this shared responsibility work in a system that we have many, many actors involved?
21:21So, it's a tough one.
21:22So, I love that discussion that just took place.
21:25You know, obviously, I mentioned humans in the loop,
21:28and I could make it sound very politically correct
21:31that we're just going to put humans everywhere all the time whenever that's needed.
21:35But the reality, it's a million or billion dollar question.
21:38How does it work?
21:40When you have millisecond-based decision-making, you know, processes,
21:46you can't really insert yourself and have humans being inserted.
21:50The other question is, the AI we know today has nothing to do with the AI we'll get in two,
21:56three years.
21:57So, the question is, first of all, you need to have an evolving view on where do you need humans
22:03to be involved
22:03and things where we decide today we need human beings six months down the road that will be delegated to
22:09AI.
22:10Like, for instance, a very good example of that is code review.
22:14Code review was like, okay, I'm doing development, I'm using cloud, I'm developing, generating code,
22:18but the code review are done by humans.
22:21Now, if you talk to developers, they have been shifting code reviews to AI.
22:26So, I think at the end of the day, it goes back to the definition of risk.
22:30Risk is probability times impact.
22:33And we have to balance that out with speed of automation and scale.
22:36And we have to be very clear.
22:40The question that we have in the discussion we're having right now is, how do you frame risk?
22:46Should I start by delegating too much too early to AI or should I be very conservative
22:51and then say, no, no, I really want to slow down the whole thing and have humans in the loop?
22:56The only answer that I can think about is the mindset has to be rooted in agility.
23:04If I'm trying to give you an answer that is very generic and long-term defined,
23:09then for sure I'm going to fail.
23:11But if I'm capable of incrementally say, I'm going to try that,
23:15I'm going to delegate that small amount of increment to AI.
23:19And I'm going to see how it flows, you know, how it works.
23:22And if it doesn't work well, I can revert back to humans being in the loop.
23:26And if it works well, 99.999% of the time,
23:29then I check and say, you know what, let's move on and let's figure out the next increment.
23:33And what we need to do is to be very much intellectually honest
23:37and validate if that increment was a success and or a failure,
23:41have that kind of ethics committee and discussion in a very transparent matter.
23:45Make sure that we have all the right data sets that allow us to tell us,
23:48did you know that those transactions were very important
23:51and you flagged them the wrong way and that was the impact?
23:53If the impact is small and 99.99% of the time it was okay, then that's fine.
23:58So I think my answer to you is really agility, being intellectually honest
24:03and moving forward as fast as you can, but incrementally.
24:07I remember you mentioned the word collective learning when we were talking.
24:11Yes, exactly.
24:12Thank you so much.
24:16Antoine, let's talk about the topic we had a little chat about it,
24:22which was, I believe it's controversial,
24:25which is about open models versus closed models.
24:28And you have a very strong opinion on that.
24:32Yeah, I mean, I have very strong opinions about it,
24:35but very biased opinions as well as I develop open models and not closed ones.
24:40You know, I think when it comes to this question of accountability,
24:45what open models do is they provide the ability to be more granular
24:49in where accountability lies.
24:53There is, well, maybe starting from closed models,
24:56there's kind of full control over the pipeline of model development
24:59all the way to ingestion by a user,
25:02assuming somebody is not building a scaffold on top of it.
25:05With open models, you know, you can get far more granular takes
25:10from, you know, what went into the model that was used in the first place.
25:13You know, who is the person who takes that open model,
25:15adapts it for themselves to create some sort of product?
25:18You know, how is that product potentially used
25:20by some other downstream distributor?
25:22And that allows for, you know, responsibility, you know,
25:25for certain failures to be assigned at different points in that pipeline
25:29and really put the onus on those that create it
25:33and distribute it in different places.
25:34I think with closed models, it becomes a whole lot more opaque where that lies.
25:40And, you know, I guess to use my own kind of famous quote
25:42that people are aware of, there's this line that machine learning
25:44is money laundering for bias.
25:47That essentially, it allows you to kind of take this statistical,
25:54I guess, the statistical truth that is produced
25:56and just treat it as though there's no way that it could be wrong
26:00because it reflects the world on which it was trained on.
26:02And I think there's a bit too much of that
26:04when we can't actually break down how a model and an engine
26:08makes it to a user into smaller and more modular components.
26:11So I'm a big fan of open because I think it allows for more opportunities
26:14to assign responsibility in other points of the pipeline
26:17while I think things are very opaque in closed models.
26:20And, of course, when things are opaque,
26:22it's much more difficult to point to where failures actually come from.
26:25Yeah, and exactly.
26:27That was when my argument was like,
26:29if somebody takes this open model and, you know,
26:31does something bad with it, you know, making harm,
26:34how can we go about it?
26:36And your response to that was,
26:37well, they could crack that anyways, you know,
26:40no matter, you know, if it's closed or open.
26:42So it's better it's open so we can democratize that, giving access.
26:45Absolutely, yeah.
26:50Mark, there is something in our prep call
26:53that I think is one of the, you know,
26:56important points that we can talk about,
26:59which is you mentioned you can govern an individual AI agent,
27:04but that tells nothing about the whole system.
27:08When we go to the system level,
27:09the behavior completely changes.
27:11So how can we go about this distributed, decentralized AI governance?
27:17Yeah, okay.
27:18And I think this is one of the big topics for the next few years,
27:22that the future of AI is not going to be one big model
27:25sitting on one big computer somewhere.
27:27It's going to be millions, billions of agents
27:30distributed and decentralized without any central command and control.
27:36And as hard as it is to govern one thing,
27:38we now need to think about how do we govern that complex,
27:41evolving system as well.
27:43And I think you're exactly right.
27:46The governance challenge there,
27:48we come from a world of software
27:49where everything is very neat and modular
27:52and you can test a piece of software
27:54and it's very predictable
27:55and then you can include that within other systems
27:57and those systems also become clean and predictable.
28:02And that's not what the future is going to be with AI.
28:05We're going to have autonomous agents,
28:08which by definition are unpredictable
28:11and stochastic and probabilistic.
28:14And we're going to combine them together
28:15and the emergent effect is going to be something
28:18dynamic and unpredictable.
28:20And that sounds like a crazy idea.
28:22That sounds like something we don't want.
28:24But actually, we really do want that
28:26because that emergent behavior
28:28is what makes them so powerful
28:29that we can give high-level tasks
28:33and this agentic system will break it down,
28:35divide and conquer,
28:36will communicate and collaborate internally
28:38to solve that problem
28:40and amplify the capability
28:43to give you something greater
28:45than the sum of the parts.
28:46And that's brilliant.
28:47We want that.
28:48But that works both ways as well.
28:50And it also amplifies the risks.
28:53It amplifies the risk of sort of unintended behavior
28:56and things like that.
28:58So I guess I would argue that we need to give up on the idea
29:02that we can micromanage individual components of that system
29:05and hope that the whole system will work.
29:08The reality is that we do need to still do that,
29:12but we also need to govern the system as a whole
29:14and the emergent behavior of that system.
29:18Otherwise, we're only doing half the work, right?
29:23And the outcome is not going to be as predictable as we think.
29:28So it's a balance.
29:30We want to unleash that capability,
29:33but we need to look at how we govern a system
29:37and a system of systems.
29:38And to be honest, that's not an AI problem.
29:41It's not a problem of models or parameters or anything like that.
29:44It's a systems engineering problem.
29:46We're going to need skills in optimization,
29:49in game theory, in network theory,
29:52all of these other things that are buried
29:54in computer science textbooks from the 1980s.
29:561980s, those things are going to come back.
29:59And it's not going to be about bigger, better models
30:03in that situation.
30:05Amazing words from an OG.
30:08Ariana, you touched this a little bit before
30:11about the fact that in many of the AI systems,
30:15actually, there is no human in the loop.
30:18Touch pass on that.
30:19And I really would like you to give the example
30:22you told me about the Syria or Lebanon case
30:26you work then to kind of go around then
30:30how can we go, how can we deal with situations
30:33when there is no real human involved?
30:37Well, I just came back from Jordan
30:39where we work with teams reconstructing Syria, Lebanon
30:45and different countries suffering great levels of destruction.
30:51And I thought something quite interesting
30:54about what you said regarding the human in the loop.
30:58So the assumption is that there is a human in the loop
31:01or there is somebody who is super skilled
31:06to actually check on what the AI is doing,
31:10but that's not necessarily the case.
31:13The reality is that the humans that are there
31:18are very constrained.
31:20They're overwhelmed.
31:21They can do not as much as we would optimally like them
31:28to be able to do.
31:29And so this idea of comparing AI with human
31:34is a false comparison
31:37because we need to compare the AI
31:39with what's actually available.
31:42And so when we fail to deliver health
31:45to a child in Syria,
31:47we cannot just assume
31:50that we should be comparing the AI
31:54with a perfect health system.
31:56That's not reasonable.
31:58So as we were working with this team,
32:01the team said, well, we have 1,800 primary healthcare units
32:05which have been destroyed
32:06and we have funding for 150
32:09and how should we prioritize?
32:11And there was an amazing team.
32:13They had collected all the geospatial data,
32:16population data.
32:17They actually had a platform already set up
32:20and wanted to understand the optimization rules
32:22that they could inject into this system.
32:27And this was all great and well,
32:30but my question was,
32:32are we doing the right thing?
32:34Are we going to invest the next five years
32:37in building 150 health units
32:40and will we reach a small portion of the population
32:43or should we think radically differently?
32:46Because today, the future is now.
32:50We could think of telemedicine,
32:54digital diagnosis, mobile health units,
32:58we could think of triage and risk assessments
33:01to prioritize the most at-risk populations
33:04and so forth and so on.
33:07This is what I like to compare AI
33:11to a status quo of reconstructing a system
33:15that is broken versus the opportunities ahead.
33:19So this requires huge investments in institutions.
33:25We had a lecture by Michael Kramer
33:27just the day before, just on Monday.
33:31Michael Kramer is a Nobel laureate in economics
33:35and he basically made the point
33:38that the value of the improvement
33:42in AI weather forecasting
33:47is much greater
33:50from a social point of view
33:52than it would ever be to the industry.
33:56A farmer has no information
33:59and plants at the wrong time.
34:02His crop gets destroyed.
34:04We need to localize the LLMs
34:09to local conditions.
34:10We need to understand how to communicate
34:13with farmers in a way that is understandable.
34:16We need to set up willingness from governments
34:20to communicate on a regular basis
34:22to all the farmer communities
34:24and so forth and so on.
34:25Build all the institutions
34:27that then create value
34:29from these investments
34:30which are amazing
34:32and really powerful
34:35but will not actually realize their value
34:38if we don't put a lot of institutional development,
34:43decision-making
34:44and coordination around it.
34:48So we have hygienic systems
34:50making consequential decisions
34:53distributed across different actors
34:56as we just discussed
34:58and Bruno mentioned.
34:59You mentioned that
35:00especially those people
35:01when they are struggling already
35:04with the current crisis
35:05and it's really hard to pin down
35:08who's responsible
35:09when something goes on.
35:12And on top of that
35:13we have the regulatory challenge,
35:17the fragmentation challenge
35:19of the regulations.
35:20So let's now talk a little bit about that,
35:24the regulations.
35:25And I want to start with you, Bruno.
35:28Orange is operating in,
35:29as you said,
35:3026 countries worldwide
35:32and you are facing
35:36many different regulations,
35:39different frameworks
35:40in many countries,
35:42I assume,
35:43that you are working.
35:44The regulation is non-existent
35:47or very different, basically.
35:49So how do you maintain
35:52a coherent responsibility framework
35:55across this whole range of countries?
35:58So that's a very important question.
36:02First of all,
36:02we have no problem with regulation.
36:04We have a problem
36:06when they start piling up
36:08and you have to deal
36:10with national regulation
36:11and European regulation.
36:13Then it becomes sometimes hard
36:16to figure out
36:17really how you're going to be able
36:19to work them out.
36:21Look, we live in a world
36:22that is complex.
36:23We operate in 26 countries.
36:25We don't have
36:26even the same environment.
36:27Like, for instance,
36:28we can use Chinese vendors
36:29very easily
36:31in Middle East and Africa.
36:33It's more challenging
36:34in France than in Europe.
36:35We can essentially
36:37be very open
36:38to using Chinese models
36:40in some areas.
36:41Maybe it's a bit more challenging
36:42elsewhere.
36:43Sometimes it's not clear.
36:44Like, for instance,
36:46is it a problem
36:47to use a Chinese LLM
36:50when it's on-premise
36:51under your control?
36:52You don't even have
36:53a regulation for that.
36:55What you know
36:55is that you want to make sure
36:56that you don't use
36:59maybe DeepSeq
37:00as a software,
37:01as a service,
37:01but maybe we can have
37:02access to their model
37:03and run their model
37:05on your premise.
37:06That's very, very different.
37:08It's something with Quen.
37:12So, to go back
37:13to your question,
37:14when you operate
37:15in 26 countries,
37:16first of all,
37:17you need to be agile.
37:18You need to empower
37:19the people on the ground
37:20to give you a perspective
37:22on really what are the needs.
37:23Because the worst thing
37:25we can do
37:25when you are operating
37:27at the scale of Orange
37:28is to believe that
37:29every decision
37:29needs to be centralized.
37:31So, what needs
37:32to be centralized?
37:33What needs
37:33to be decentralized?
37:35I think what needs
37:36to be centralized
37:37are the core values.
37:38We believe that
37:40AI needs to serve humanity,
37:43needs to be in favor,
37:45to go back to your point,
37:47of progress.
37:49We believe in privacy.
37:52We want to make sure
37:53that nobody is going
37:55to use AI in a way
37:56that could hurt our kids,
37:59elders, or what have you.
38:01We also want to make sure
38:02that we give a consistent framework
38:04to our employees,
38:05like, for instance,
38:06in terms of their own privacy
38:08using their, like, workstation.
38:10But at the end of the day,
38:13there is a value
38:14in operating in 26 countries
38:15where the variables
38:16are a little bit different.
38:18You get to experiment
38:19different things,
38:21and then it rolls back up
38:24into kind of a more
38:26centralized entity,
38:27and you see what works,
38:28what didn't work.
38:28I love the notion
38:30of trying different things.
38:31As long as we abide
38:32by the core value
38:34of the company,
38:36and right now,
38:38we are doing things
38:40in some countries
38:41that could not necessarily
38:42be done in others,
38:43and that's fine
38:44as long as it doesn't contradict
38:46the core values
38:47that I was talking about.
38:48The only charter
38:49that we have right now
38:49that is really applicable
38:50to all the 26 countries
38:52is what we call
38:53responsible AI,
38:54and that's already
38:55a very solid framework
38:56for us.
38:57When it comes down
38:58to implementation
38:59and very technical details,
39:01I think it's a good thing
39:03to have some amount
39:03of leeway,
39:04country by country.
39:05And for instance,
39:06when you work
39:07with the Chinese ecosystem,
39:08that gives you
39:09a lot of insight,
39:10a lot of data points,
39:12that even if you might
39:13not be able to work
39:14with them in France
39:15or in Europe,
39:16that's going to help you
39:17retrain your own brain
39:19of what's possible,
39:20not possible.
39:21We are seeing,
39:22like for instance,
39:23we go to Asia quite often.
39:25They are very advanced
39:26in the use of robots.
39:27They have small robots
39:29that they want to use.
39:30They essentially
39:31have even robots
39:32to help elders.
39:34Things that might look like
39:35makes no sense
39:36in Europe or what have you,
39:37but maybe in two or three years,
39:39it will happen.
39:40So I think it's a strength
39:41and I think the notion
39:43of being exposed
39:44to a variety of points of view
39:46at the end of the day
39:47is something that will help us
39:48in the long run.
39:51Mark,
39:53we had a,
39:54I remember you mentioned
39:56something very interesting.
39:58It stayed in my mind.
39:59You said that ethics
40:01are actually like brake pedals
40:04and it's not something
40:07that it slows you down,
40:09but it actually is there
40:11to help you
40:12to accelerate faster.
40:14Yeah.
40:15And again,
40:16I'm hijacking
40:17someone else's quote.
40:19I think this originally
40:20came from one of the
40:21Chinese ministers
40:21of science, right?
40:23And he made the point
40:24that the purpose
40:25of a brake pedal
40:26in a car
40:26is not to slow you down.
40:28It's to allow you
40:29to more confidently
40:29put your foot
40:30on the accelerator,
40:32which I'm not sure
40:33my teenage daughter
40:34would agree
40:34that she drives differently.
40:36But I think it's
40:37a nice analogy
40:37because it shows
40:38where we need
40:40to get to with AI
40:41because at the moment
40:42people do see
40:44governance,
40:45regulation and ethics
40:46as a restrictive force,
40:48something that slows them down.
40:49They see it as a tax
40:51on their projects
40:52and something
40:53that you do
40:55retrospectively
40:55if you have to,
40:56if someone forces you to,
40:57right?
40:58And we need to shift
41:00to frame it
41:01completely differently,
41:02to look at it
41:03more as an investment,
41:04to say that this is
41:05an investment
41:05in the long-term success
41:06of this project,
41:08that if we get
41:10the safeguards right,
41:11we can go further
41:12and faster
41:13than if we didn't
41:14have them
41:15in the first place.
41:16And just completely
41:17flip the conversation
41:18around from being
41:18something that costs
41:19us money and time
41:20to something
41:21that will improve
41:22success in the long term,
41:23that will improve
41:24the outcomes
41:25for all of the people
41:26and groups
41:26that Bruno just mentioned,
41:27right?
41:29And it's a commercial
41:30imperative,
41:31not just a risk mitigation.
41:32And I think
41:33we just need to
41:36just grow up
41:37a little bit,
41:37just get a bit
41:37more mature
41:38in our outlook
41:40on regulation
41:41and try and flip it
41:43into a positive
41:44rather than negative.
41:46Thank you so much,
41:47Antoine.
41:48I would like to
41:49ask you the next question
41:51and I'm also aware
41:52of our time.
41:55So,
41:55I want to know
41:56if you can share
41:57a little bit
41:58about the Switzerland
42:00perspective
42:01because it's,
42:02you know,
42:03it's in Europe
42:04but it's not really
42:07bound to the EU AI Act.
42:10At the same time,
42:11you know,
42:12there are things
42:14that there are,
42:14like there are strings
42:15that are connected
42:16and I would like
42:17to know
42:18what does that
42:19say about
42:21responsibility
42:21with your work
42:22with your company
42:24Apertus
42:24in compliance
42:26with,
42:27you know,
42:27the EU framework
42:29because you mentioned
42:29that you actually
42:31have decided
42:32to go
42:32steps further
42:34to be compliant
42:35even though
42:35it's not legally
42:36mandatory
42:37for your company
42:38being based
42:39in Switzerland.
42:42Yeah,
42:43there's a lot
42:44to unpack there.
42:45I guess as,
42:47I won't necessarily
42:48speak for Switzerland
42:49as a whole,
42:49just myself
42:50and my team.
42:52You know,
42:53what I would say
42:54is,
42:54you know,
42:54Switzerland,
42:55the minute
42:56that you want
42:56to scale
42:57some type
42:57of technology
42:58or product
42:58out of Switzerland
42:59which,
43:00you know,
43:00is a powerful
43:01market on its own
43:02but,
43:03you know,
43:04slightly small
43:04in the center
43:05of Europe,
43:05you need to abide
43:07by the rules
43:08of your neighbors
43:08and other markets
43:09that you're going to.
43:10So for us,
43:11it was kind of
43:12a no-brainer
43:13to already
43:14be thinking
43:15about how
43:15the regulations
43:16and other areas
43:17should be brought
43:17into the design
43:18of Apertus.
43:19What I'll say,
43:20though,
43:21is that based off,
43:22you know,
43:22these ethical principles
43:24that we wanted
43:24to put at the center
43:25of development,
43:26it didn't actually
43:27require us
43:28to change our roadmap
43:29all that much
43:30to be compliant
43:31in that setting.
43:32It was what we wanted
43:34to do in the first place
43:35based off the opportunity
43:36that we had identified
43:37for why our model
43:38could be valuable
43:39to certain groups
43:40of users.
43:41And I think,
43:42you know,
43:42that kind of teaches
43:43an important lesson
43:44as well here
43:45that I guess
43:45to support
43:46what I said before,
43:47regulation doesn't
43:48have to be
43:49a brake pedal necessarily.
43:50It can,
43:51in fact,
43:51be an opportunity.
43:53And it's also
43:54not as black and white
43:56as most people,
43:57I think,
43:57think it is.
43:58It's often,
43:59you know,
44:00folks that are developers
44:01don't have armies
44:02of lawyers
44:02to kind of help them
44:03interpret these things,
44:04so they naturally,
44:05you know,
44:06are, you know,
44:07perhaps afraid
44:07of what is said
44:08in these regulations.
44:09But the reality
44:10is that it's often
44:11kind of more flexible
44:12and it's meant
44:13to be flexible
44:13because regulations
44:15should survive
44:15kind of the changes
44:16in the current day-to-day,
44:18which are often more,
44:19you know,
44:19regulated by standards
44:20themselves
44:20as opposed to law.
44:22And so as a result,
44:23I think it can,
44:23it can absolutely
44:24be an opportunity
44:25to try to be legally compliant
44:27from first principles
44:28and shouldn't necessarily
44:29be a blocker.
44:30And that's something
44:30that we've put
44:31at the center
44:31of aperitist development.
44:32Thank you so much.
44:34Unfortunately,
44:35we're about to end
44:37this panel,
44:38so I want to wrap it up
44:39and thank you so much
44:41for your contribution
44:44to this panel.
44:45So what strikes me
44:46about this conversation
44:47is that responsibility
44:49in AI isn't a single question,
44:52it's a chain.
44:53It starts with data,
44:55runs through the design,
44:57the integration,
44:58the deployment,
44:59and the regulation,
45:00and it ends
45:02with the person
45:03who gets affected,
45:04gets impacted.
45:06And the hard truth
45:07is that this chain
45:09currently is broken
45:11in multiple places.
45:14And the pace of AI
45:16is beyond our capacity
45:19today to adapt
45:20in terms of the governance
45:21and regulation.
45:22But what I also want
45:24to mention
45:25and I heard from you
45:26is that there are tools
45:28that exist,
45:29radical transparency,
45:31ethical frameworks,
45:33living governance systems,
45:36evidence-based feedback loops,
45:38and what's missing
45:39is the collective will,
45:40in my opinion,
45:41to treat responsibility
45:43not as a constraint
45:44on progress,
45:45but as the condition for it.
45:47So with that,
45:48I would like to thank
45:49our dear panelists
45:50for their beautiful contributions.
45:52Thank you so much
45:53for the audience.
Comments