Latest / The Road to Accountable AI / Richard Benjamins: A Multi-Year Journey Toward Responsible Enterprise AI

Transcript

Richard Benjamins: A Multi-Year Journey Toward Responsible Enterprise AI · The Road to Accountable AI

• 0:01
  Richard, thank you so much for being a part of the podcast.
• 0:06
  Yeah, thank you very much, Kevin.
• 0:08
  It's a pleasure to be here.
• 0:10
  Tell me how you first got into the area of, well, AI more generally initially, and
• 0:17
  then responsible AI.
• 0:19
  Well, AI more generally, I did my studies and my PhD already in artificial
• 0:24
  intelligence in the eighties, in the nineties.
• 0:27
  So that's a long time ago when nobody knew what artificial intelligence was.
• 0:32
  And then of course, since 2016, when this came back into society and later in
• 0:37
  business, that's what I focused on.
• 0:39
  So long history in artificial intelligence and in the responsible use of AI, I came
• 0:47
  across when in 2000...
• 0:49
  17, I was working at a large insurance company.
• 0:53
  And in my team, there was a small group who looked at bias and explainability.
• 0:58
  And I thought, before that, I just had a business look at it.
• 1:02
  Yeah.
• 1:02
  So how can you use this for businesses?
• 1:04
  But then I said, wow, this is actually important.
• 1:08
  And if you if you take this into account, and if this is going be applied
• 1:13
  everywhere.
• 1:14
  So when I moved back to Telefonica, I took that on board and I brought it into the
• 1:20
  company and said, well, if we want to be serious on AI, we should not only look at
• 1:25
  business opportunities, but we should also look at responsibility.
• 1:28
  And that's how I came in the field.
• 1:32
  What was the response in the company?
• 1:33
  Because obviously there's some effort that's required and there may be
• 1:37
  trade-offs in terms of designing responsible systems, dealing with issues
• 1:41
  like fairness and bias.
• 1:43
  So what was the response you got?
• 1:45
  Well, the response was surprisingly positive.
• 1:50
  Yeah.
• 1:50
  So for instance, when in 2000, we started this, this effort in 2018 in the
• 1:54
  beginning.
• 1:56
  And basically because the company was very vocal on big data and artificial
• 2:00
  intelligence and the opportunities.
• 2:02
  Uh, and at that time there were a lot of those scandals about discrimination, the
• 2:07
  compass things, uh, this company's hiring people and only hiring a male gender.
• 2:13
  So, um,
• 2:14
  I thought, well, it's important to make a statement as well.
• 2:17
  So we brought together a very, uh, a set of departments, yeah.
• 2:22
  Which you have to put together because this what about, it was an aspirational
• 2:25
  statement that we wanted to make.
• 2:27
  So we wanted to publish our ethical AI principles.
• 2:31
  Uh, and it took us a month of four months or five months to agree on the set of
• 2:37
  principles and then how do you formulate them?
• 2:39
  Because of course, many
• 2:41
  commercial people think, well, if we do this, then we will lose speed, we will
• 2:45
  lose revenues, but it was surprisingly fast.
• 2:51
  So I think we published them in October 2018.
• 2:56
  And a very important lesson, because I've also spoken to many other companies, and
• 3:03
  there are two approaches to...
• 3:06
  make these statements, publish your ethical principles.
• 3:09
  The one is, is a kind of aspirational one.
• 3:11
  So this is what we want to do.
• 3:13
  And we acknowledge that we're not yet there, but we're gonna work hard to be
• 3:17
  there.
• 3:19
  And then there are companies that before publishing anything, they want to be
• 3:23
  absolutely sure that they are already complying with what they're saying.
• 3:28
  Those companies, they are still discussing their ethical principles, because
• 3:32
  obviously it's not so...
• 3:34
  easy to make sure that you comply with all of them in all aspects.
• 3:38
  So we are on a journey as a sector, private sector, public sector, as a
• 3:43
  society, and this is going to be incorporated increasingly more, but
• 3:47
  gradually.
• 3:49
  Absolutely.
• 3:50
  Well, this is an important point, though.
• 3:52
  How does an organization ensure if they actually get principles out the door, that
• 3:57
  they are more than just aspirational, that they can actually influence the
• 4:01
  development that goes on?
• 4:03
  Yeah, so now you are alluding to ethics washing.
• 4:08
  And so any company, any organization that starts on this journey will first start
• 4:14
  with principles and there is really nothing behind it.
• 4:19
  So you can, you can accuse them as being ethics washing.
• 4:26
  But in the end, this is the way how you start.
• 4:28
  And then the next year they have to, you have to do something more.
• 4:30
  And the next year you have to do something more.
• 4:32
  So.
• 4:32
  It's a journey.
• 4:33
  So we started in, let's say 2018, we published into end of 2018.
• 4:39
  And we only approved a formal internal regulation last December.
• 4:45
  So it took us like, uh, like, like five years to do that.
• 4:51
  And that's not because, uh, we had to convince everybody because there is no
• 4:56
  experience whatsoever in the world, how to do that.
• 4:59
  So you have to find it out.
• 5:01
  yourself and then talk to some other companies who do the same, but there are
• 5:05
  very few, especially in the beginning.
• 5:07
  So what we did is actually the first step was we did a training course where we
• 5:14
  explain our employees and most people most affected.
• 5:17
  So we're not speaking about only developing artificial intelligence, but
• 5:21
  also about procuring it on the market and also about using it internally and also
• 5:26
  about selling it.
• 5:27
  So all those people.
• 5:28
  And that's a big, that's a, that are thousands of people in our company.
• 5:33
  And they need to understand something about this.
• 5:35
  Yeah.
• 5:35
  So that's what we started with.
• 5:37
  Then next step we started about a kind of thinking about the governance.
• 5:41
  So these are the principles, but so who is responsible for it?
• 5:45
  What are the functions?
• 5:47
  Do we need a new role in the organization to be able to deal with this?
• 5:53
  We, of course, like everybody, we started to make a questionnaire where we do this
• 5:58
  ethics impact assessment.
• 6:01
  And I think we went through like 15 different versions based on experiences
• 6:06
  with people.
• 6:07
  Then we, we brought it out in the wild and we tested it with real, uh, real use cases
• 6:14
  and people who are not so involved.
• 6:17
  And then finally we decided to run some pilots with a governance model.
• 6:20
  So do we need an AI ethics committee or do we need.
• 6:23
  a special role, like we call it the responsible AI champion.
• 6:28
  And how do you organize this completely distributed, decentralized?
• 6:32
  And what are the things?
• 6:34
  How, what is the workflow?
• 6:36
  How is compliance involved?
• 6:38
  How is ESG involved?
• 6:39
  So we did really a pilot, uh, in four different business units.
• 6:43
  And based on that experience, we, we thought we were ready to write really
• 6:47
  something up that can be then, uh, complied with, and that is what is
• 6:51
  currently the case.
• 6:53
  So it's a journey that took us many years.
• 6:56
  Of course, today, if you start a journey, there is much more experience.
• 6:59
  So it will take you maybe three years.
• 7:01
  And anyway, if you're in Europe, you have to do something anyway, because of the
• 7:05
  upcoming AI Act, which will be, which will enter in force in, in two years from now.
• 7:14
  So in that journey, how do you evaluate what's success?
• 7:18
  You talked about a number of choices, a number of options.
• 7:22
  How does the organization decide what the right things are to implement?
• 7:26
  Well, I think success is that you move forward in a possible way.
• 7:32
  And most of my experience is sometimes the process is more important than the result.
• 7:38
  So of course, the result is that all the systems that you launch and actually, so
• 7:43
  we have more than 500 systems every day running in our in our corporation
• 7:49
  globally.
• 7:50
  So it's not like 10 systems or 15 or 40.
• 7:54
  It's really a huge amount of systems.
• 7:59
  So you really have to do something seriously in order to understand that.
• 8:05
  Of course, what we now have to do, we start with the new systems, where by
• 8:10
  design we take this into account, but then slowly we have to do retrospective work on
• 8:17
  looking at all the systems that we have.
• 8:22
  But if you have a question like, for example, do we have a centralized AI
• 8:26
  review board for ethics issues versus do we decentralize it out?
• 8:31
  How do you get to answer some of those questions at an organization?
• 8:37
  Yeah, that's a good point.
• 8:41
  It depends of course, whether I think one of the first things that is important is,
• 8:45
  are you a provider of artificial intelligence, mainly, or are you a user of
• 8:51
  artificial intelligence?
• 8:52
  A company like Telefonica or energy companies or financial companies, they are
• 8:57
  all users, pharmaceuticals, and companies like Microsoft or Google, et cetera, they
• 9:03
  are providers.
• 9:05
  of this technology.
• 9:06
  So I think that makes a very different thing.
• 9:09
  If you are a provider, and you're a big tech serving thousands or millions of
• 9:15
  customers across the globe, I think you have to be very centralized.
• 9:19
  You can't leave this locally because then many things will happen.
• 9:23
  If you are a user of this technology, it depends a bit on how
• 9:29
  unified you are as a corporation.
• 9:30
  I mean, if you have acquired, if you exist because of an emerging acquisition of 30
• 9:35
  different companies, it's not so easy to do centralized.
• 9:39
  But if you are a centralized company to some extent, then it's always good to have
• 9:44
  a mixed approach.
• 9:45
  Yeah.
• 9:46
  So for instance, we have a, let's say at the corporation at headquarters where we
• 9:53
  run several of our global businesses and our support functions.
• 9:58
  We have a central model where we have an AI ethics committee, and we have certain
• 10:05
  persons in the central business units, because there are also central business
• 10:09
  units.
• 10:09
  There is, of course, finance, there is HR, but there is also digital transformation,
• 10:17
  let's say all our networks from a global perspective.
• 10:20
  We have all that centralized.
• 10:21
  And then, of course, we have operations in 15 companies across the world.
• 10:26
  And then what we tell them.
• 10:28
  is here you have a role model which you can use.
• 10:32
  And if you want to use it, we can help you.
• 10:34
  But given your situation locally, you are free to set up an alternative way of doing
• 10:40
  this as long as this is approved globally.
• 10:45
  So it's a kind of you cannot impose everything from the top because it's not
• 10:51
  one size fits all.
• 10:52
  But there are certain principles that you have to follow.
• 10:58
  that you have a figure called, actually a role, it's a function of another job
• 11:04
  called responsible AI champion.
• 11:06
  That is in every business unit that uses, develops, buys, or sells artificial
• 11:12
  intelligence needs to have such a function.
• 11:15
  And this person is the go-to person for AI ethical responsibility questions across
• 11:21
  that unit.
• 11:24
  Now, if you don't use any AI in a certain area, of course you don't need such a
• 11:29
  person.
• 11:30
  If you use a lot, maybe you need three and it's not even full-time.
• 11:34
  So it's just assignment of a new real refunction to an existing job position.
• 11:42
  And so that is completely distributed because it sits in the businesses.
• 11:49
  And then of course you need some...
• 11:51
  obstruct you to organize, to bring those people together, to train them.
• 11:55
  You need an AI ethics committee.
• 11:58
  And then of course you can have the same, you can have for each big area of
• 12:02
  business, you can have a separate AI ethics committee, or you can have one
• 12:06
  centralized for the company.
• 12:08
  What is the best situation?
• 12:10
  So actually there are two factors.
• 12:11
  One is, do you have enough experts to make such statements?
• 12:15
  Because it's not an easy.
• 12:18
  Of course, it sounds like anybody has an opinion about ethics and about AI, but
• 12:22
  that's not what you want.
• 12:23
  You want people with a mix of profiles, of backgrounds, but you also want to have
• 12:30
  some expertise in the area.
• 12:32
  So you don't have always the luxury to have like five committees in the different
• 12:36
  business because simply it's not yet there over time.
• 12:39
  Yeah.
• 12:39
  There will be much more experts, but today that's a difficult question.
• 12:43
  And, and the other part is.
• 12:49
  whether how much AI there is.
• 12:51
  Yeah.
• 12:52
  So if there is a hundred systems, maybe with one committee, because it's
• 12:56
  sufficient, because it's not a, so you have also two ways of using an AI, AI
• 13:02
  ethics committee.
• 13:03
  One is permission-based and the other is problem-based.
• 13:08
  So permission-based means anything that will be done by the company first goes to
• 13:13
  the committee.
• 13:14
  There is a discussion, yes, we can do that.
• 13:16
  No, we cannot do it or whatever.
• 13:18
  And that's permission based.
• 13:20
  And problem based is, look, we're working with this AI product.
• 13:25
  We are discussing about the impact, and we see there are some doubts.
• 13:28
  Should we use gender as a variable in the model?
• 13:31
  Yes or no?
• 13:32
  Now, that's a problem.
• 13:34
  And then you bring it to the AI ethics committee, and they discuss.
• 13:37
  Only about ethics, never about business.
• 13:39
  And then they highlight, and they give recommendations.
• 13:42
  And then they go back to the team.
• 13:44
  Now, if the team agrees, then they
• 13:46
  they go forward, if they don't agree, is escalated to an ad hoc committee where you
• 13:50
  bring together ethics and business.
• 13:52
  And in the end, the company takes a decision.
• 13:54
  The important thing is an informed decision, it's an explicit decision, and
• 13:58
  it's a documented decision.
• 14:00
  So that's the difference by just doing things.
• 14:05
  So that works quite well from the ethics perspective.
• 14:11
  Now, if you are a technology provider,
• 14:15
  where you have billions of customers, maybe is more a permission-based.
• 14:20
  So we want to build an AI system that analyzes all electronic communications in
• 14:27
  enterprises.
• 14:28
  And then we can find hubs of knowledge and isolated hubs of departments that are not
• 14:36
  in contact.
• 14:38
  So that is a product that has, well.
• 14:42
  potentially serious implications for workers and some ethical questions.
• 14:47
  So then it's good actually, so if you provide this product to billions of
• 14:52
  people, then it's good to have a permission base because the impact is
• 14:55
  huge.
• 14:56
  So it's again, it's not the same if you're mostly a user of this technology where
• 15:01
  you're a big tech provider.
• 15:04
  Startups is a different world because they are smaller.
• 15:09
  And they, of course, they have so many resources, but still, if it's high risk,
• 15:15
  it's important to think those things in advance.
• 15:19
  Does the fact that telephonica is in telecommunications, which is a regulated
• 15:23
  industry, make it easier to adopt these kinds of practices?
• 15:31
  Well, of course, but there are many regulated industries, yeah, and
• 15:36
  telecommunications is regulated, but finance and insurance is even more
• 15:39
  regulated and healthcare is enormously regulated.
• 15:42
  So, of course, the culture of compliance, yeah, of having to having processes and
• 15:52
  systems in place that check things, that is...
• 15:57
  already in place.
• 15:58
  So you can a lot of piggyback on that.
• 16:00
  If you are a company that is not affected by any regulation, of course, then it
• 16:04
  probably is harder to start with.
• 16:08
  But take into account that we started from the business.
• 16:13
  We didn't start from compliance because at that moment, this was not on the radar of
• 16:20
  compliance.
• 16:21
  Today, companies who are aware of upcoming regulation
• 16:25
  Of course, they will start from compliance.
• 16:29
  How does generative AI change the landscape for responsible AI practices in
• 16:34
  your view?
• 16:38
  Technically speaking, not too much.
• 16:43
  I mean, ethical AI, responsible use of AI is about estimating the ethical and social
• 16:52
  impact on human rights by design.
• 16:55
  I don't know if that is machine learning with structured data or whether that is
• 17:02
  generative AI with unstructured data.
• 17:06
  You still have to evaluate the impact on human rights.
• 17:13
  Now the difference is that, let's say traditional AI is narrow AI.
• 17:18
  Yeah.
• 17:19
  So you have a use case and you evaluate the use case.
• 17:22
  Now, generative AI is kind of general purpose to some extent.
• 17:25
  Yeah.
• 17:25
  You can use it for different tasks.
• 17:27
  You can use it for very innocent tasks and also for very serious tasks.
• 17:33
  And that means that.
• 17:36
  Um, it is different because for narrow AI, it makes sense to evaluate use cases and
• 17:42
  not the technology.
• 17:43
  You don't, uh, you don't evaluate a recommendation algorithm or a random
• 17:49
  forest algorithm in terms of ethical impact.
• 17:52
  Once you apply it to a situation, then you, uh, evaluate the impact.
• 17:58
  And with, with generator to vie is this, this border is not so clear cut.
• 18:04
  Because I can bring to the market, and this is what the companies do, they have
• 18:08
  the big tech do, a general purpose AI that I can use as a user or as a business for
• 18:14
  many different things.
• 18:16
  I can use it for generating marketing texts, which is probably not so impactful.
• 18:23
  But I can also use it to analyze natural language of patient dossiers.
• 18:31
  and then make a decision for certain patients in terms of healthcare decisions,
• 18:35
  which is very high impact.
• 18:37
  I can record interviews for HR automatically and then have them analyzed
• 18:46
  by Generative AI to see whether I should hire the people that has a lot of impact.
• 18:52
  So it's difficult to put an ethical impact assessment on chat GPT.
• 18:59
  as such.
• 19:00
  So you have to do it.
• 19:02
  And as a user, that doesn't change because we never use chat GPT or generative dive
• 19:08
  for whatever.
• 19:10
  So we use it to improve this or in our legal area or in our whatever area.
• 19:16
  So we analyze the use case of this technology.
• 19:21
  But if you are a provider,
• 19:22
  then it makes a difference.
• 19:23
  And that's also where the AI Act has found a lot of difficulties in coming to
• 19:29
  agreement.
• 19:31
  And because if you want to regulate it as one system, let's say as a high risk,
• 19:39
  impactful system, then you also regulate a lot of innocent use cases and the other
• 19:44
  way around.
• 19:46
  Absolutely.
• 19:47
  But let's say you're in a company like Telefonica that's not the one developing
• 19:50
  the large language models, but which has a lot of these use cases.
• 19:54
  And so a product group says we want to use a chat bot for customer service,
• 20:00
  generative AI.
• 20:02
  How do you think about the addressing those kinds of ethical issues?
• 20:08
  Well, there is beyond the ethical question.
• 20:12
  There is also just a performance question.
• 20:14
  So if I use a chat bot externally with customers and I based that large part on
• 20:23
  existing systems like Lama2 or like, like chat GPT, then I have a significant
• 20:29
  business risk, not even speaking about ethical because those things can answer
• 20:35
  you whatever.
• 20:37
  whatever they can go into a direction.
• 20:40
  So what companies have to do is kind of how do you tame or limit what those
• 20:47
  systems can say if they are talked to with your customers.
• 20:50
  And that's even only from a business perspective.
• 20:53
  Yeah.
• 20:54
  You don't want, so if we have a chat bot where we can, we can talk with our
• 20:57
  customers, you want them to ask questions about the, about the elections in the U S
• 21:02
  and that we make a statement that George chat bot makes a statement about.
• 21:06
  candidate or so.
• 21:08
  Now, how do you limit that?
• 21:12
  Of course, if you manage to do that, then there are technologies for that, rack
• 21:16
  technologies, whatever, where you take your own content, and then you use that to
• 21:23
  generate the information, and then you use ChetGBT to have the dialogue and the
• 21:28
  conversation to put everything together in a natural way.
• 21:31
  That reduces significantly the hallucination part.
• 21:36
  But still it can still have a lot of bias.
• 21:38
  Yeah, so like bias professional bias in terms of professions typical male or
• 21:46
  Female process professions or it can even have Yeah, let's say use language that we
• 21:53
  don't want them to use insinuating Things that are not inclusive.
• 21:59
  So there's still quite some challenges from an ethical perspective.
• 22:02
  But then what you have to do is
• 22:04
  This is my knowing what is your target audience and then say, okay, suppose that
• 22:10
  something is going wrong in this direction.
• 22:12
  So what is the impact on the people it's working with?
• 22:16
  Is that impact severe or light?
• 22:21
  What's the probability that it happens?
• 22:23
  Is it 0.1% or is it 20%?
• 22:26
  And then how many people are involved?
• 22:29
  Is it 10, 100, a million, a full nation?
• 22:32
  So all those things matter in estimating what is the impact.
• 22:36
  And if you do that by design and in a systematic way, then the risk that
• 22:42
  something comes out that you don't want is much less.
• 22:46
  Of course, you can never avoid it, yeah?
• 22:48
  Because some things you never know in advance.
• 22:51
  But still, I mean, it's much better to do it in such a way than just do it as quick
• 22:55
  as possible, put it in the market and then discover, oops, we didn't know this.
• 23:03
  What about cultural context, language, how specific do these kinds of mechanisms have
• 23:11
  to be?
• 23:14
  Well, it is true that because those machines, this is massive brute force
• 23:20
  calculations with statistical models.
• 23:23
  Of course, the dominant culture on the internet is English and US.
• 23:30
  So of course, many things that it answers, even though in a different language, still
• 23:35
  the culture that is encoded implicitly in the model is a specific culture.
• 23:42
  And that has as a disadvantage that smaller cultures can be, yeah, let's see,
• 23:48
  have no, no role.
• 23:49
  Uh, not so much.
• 23:51
  Of course, in the Spanish speaking world, there are still a 600 million Spanish
• 23:55
  speaking people in the world.
• 23:57
  So still a very large language.
• 23:59
  Yeah.
• 23:59
  Not as, of course, by far, not as big as English, but definitely bigger than the
• 24:03
  Dutch and from the Netherlands.
• 24:05
  So only 20 million people in the world who speak that.
• 24:09
  So, um,
• 24:11
  From a language perspective, in terms of words, that can be worked on, because
• 24:17
  there are initiatives, local initiatives, where people who love their language
• 24:23
  always put focus on that.
• 24:25
  So I'm not so worried about that.
• 24:28
  But the cultural aspect is something that has to be taken into account.
• 24:35
  As far as I know, that is not a solved problem, but it's a research problem.
• 24:42
  So you mentioned a couple of times the AI Act that was recently adopted in Europe.
• 24:47
  How do you see that or the broader move globally to adopt AI regulations impacting
• 24:54
  what companies are doing?
• 25:00
  I think it's a very good step, first step.
• 25:02
  It's a good step in the right, an important direction.
• 25:08
  I think global regulation or global agreements are very important.
• 25:13
  That's why the EU act is important, because it covers 27 countries.
• 25:19
  It's not for one country only.
• 25:21
  Of course, the world is much larger.
• 25:24
  than 27 companies and there of course there is the US and there is the UK, there
• 25:29
  is China, Russia.
• 25:32
  In that perspective I'm a fan of the UNESCO Ethical AI recommendation because
• 25:38
  it covers 193 plus one countries including the US, China, Saudi Arabia, Iran, so
• 25:49
  countries that are very hard to get together.
• 25:53
  and they are together and they have signed up to an ethical use of artificial
• 25:58
  intelligence.
• 25:59
  Of course, this is just a sign up, yeah, there's no sanction.
• 26:02
  If you still do many different things, nothing happens and that's the situation
• 26:07
  today.
• 26:08
  But that situation is also happening with human rights.
• 26:11
  There are countries who signed an international human rights declaration and
• 26:16
  are not respecting human rights as they should.
• 26:19
  So...
• 26:20
  This is a necessary step towards a more formal regulation.
• 26:25
  So I think the AI Act is a good thing.
• 26:30
  But of course, it also can be improved to some extent and is still on paper.
• 26:36
  There's very little experience with what it means in practice.
• 26:41
  I think a very important thing is this is risk-based and that the majority of the
• 26:47
  use cases of AI...
• 26:49
  are not high risk.
• 26:50
  They're limited list, so they have no regulation at all.
• 26:53
  But there are certain areas where there is potential risk and those areas are
• 27:00
  regulated.
• 27:01
  So from very light regulation in terms of if you make a deep fake, you have to say
• 27:07
  it's a deep fake.
• 27:10
  You can't claim that it is for real.
• 27:13
  There's transparency obligation.
• 27:15
  If it's high risk in terms of
• 27:19
  impacting human rights.
• 27:20
  So if I want to use AI to hire people or to fire people, or to accept students, or
• 27:26
  to give a loan, determine the interest rate, or whether I accept people for a
• 27:31
  loan, a mortgage, medical applications.
• 27:37
  So everything that is essential for people.
• 27:39
  If you use AI for those specific things, then you need to comply with certain
• 27:46
  requirements.
• 27:48
  before putting it in the market.
• 27:51
  So the data should not have bias, should have quality, should have human oversight,
• 27:57
  the adequate human oversight.
• 28:01
  You should log the system in the sense that if something happens later, you can
• 28:06
  investigate whether there was a problem.
• 28:08
  So it's a set of requirements that you have to take into account, but only for
• 28:14
  high risk use cases.
• 28:16
  And then there is unacceptable risk and those things are outright forbidden to do.
• 28:22
  If you do that, you have to pay 35 million or 17% of your global annual review.
• 28:27
  So, so, and then there is the part of general purpose AI.
• 28:31
  So generative AI, which they have put in later.
• 28:36
  Um, and they have struggled a lot with the fact that
• 28:42
  The AI Act was based on use cases and suddenly there is a system that you can
• 28:46
  use for many use cases.
• 28:49
  So I think that if you use generative AI for high risk, then you just do the same
• 28:57
  thing as you have to do for any other system.
• 29:00
  You have to assess the impact, you have to comply with certain things.
• 29:03
  Now, if you use a third party for generative AI, like OpenAI or Microsoft,
• 29:10
  for your use case,
• 29:12
  of little risk, nothing to do from the user perspective, from the company
• 29:19
  perspective.
• 29:20
  But if you are using it for high risk, then you have to comply.
• 29:23
  And that means there is a transparency obligation towards the provider that the
• 29:28
  provider has to say, okay, this system is trained on this data, I've done this
• 29:32
  checks on bias, I can monitor, etc.
• 29:36
  So that is a kind of tension that is written in the law.
• 29:40
  but we don't know how it spells out in practice.
• 29:45
  And then if you are a provider of such a model like OpenAI or like Microsoft, then
• 29:52
  they speak about systemic risk systems.
• 29:55
  And that means systems that are so big that they actually can create an impact in
• 30:02
  society at large.
• 30:04
  And there are some requirements on top of the system.
• 30:07
  So they have to...
• 30:08
  have a system of notification, they have to alert system, they have to check
• 30:12
  periodically the data, et cetera.
• 30:15
  And they have to report on those things to the European Commission.
• 30:17
  But how they check, what they check is left to the industry.
• 30:22
  So it's a kind of hybrid self-regulation, formal regulation, and that is definitely
• 30:28
  a compromise between different perspectives.
• 30:33
  And in my view, and that's my personal view, is a...
• 30:38
  incorrect discussion, because it speaks about either you innovate, or you are
• 30:45
  responsible.
• 30:46
  Yeah, so if you want to innovate, you cannot be responsible, because there are
• 30:51
  other countries, yeah, who are not responsible, and they can innovate faster.
• 30:56
  So that's the kind of discussion that's put up like the US innovates, China
• 31:02
  copies, and the EU regulates.
• 31:06
  Now,
• 31:07
  If you speak about responsible AI by design, so regulation by design for high
• 31:15
  risk things, that means that in all the way, in all the process that you build
• 31:20
  things like chat GPT or whatever, you start thinking already, or what could be
• 31:25
  the impact if things go wrong or what about hate speech?
• 31:27
  How do I train?
• 31:28
  So you think in advance about all those things.
• 31:31
  And then in the end, what comes out of it is probably a more responsible.
• 31:37
  project rather than correcting it once in the market.
• 31:40
  And also think, if you launch something to the market and you discover a problem,
• 31:45
  which has happened several times with Google, with the generative AI's they put
• 31:50
  in the market, actually impact and the cost of repairing that is much higher from
• 31:56
  all perspectives than if you do it while you are designing.
• 32:00
  Yeah.
• 32:02
  And then there are actually several business benefits from doing innovation in
• 32:07
  a responsible way.
• 32:08
  So it's not doing either you innovate or a responsible, no, you innovate in a
• 32:12
  responsible way.
• 32:13
  And if you do that, society is more and more requiring companies to be responsible
• 32:20
  beyond profit.
• 32:21
  We see that with climate change.
• 32:23
  Climate change is actually a business factor.
• 32:26
  And now at this moment, investors look at your climate rankings and ratings.
• 32:32
  and emissions before they invest in you.
• 32:34
  So there is a trend that even your customers value, not only your profits,
• 32:40
  but also value your responsibility.
• 32:43
  So if you're responsible and you show that, and actually you do it actually in
• 32:47
  practice, that's a positive point.
• 32:51
  The same for your employees.
• 32:54
  So employees become more and more aware of those things and they want companies that
• 33:00
  not only
• 33:00
  make money, but it also have a responsible role in society.
• 33:06
  So that also is actually a business benefit.
• 33:10
  If you are a responsible company and you've thought about the things, actually
• 33:13
  governments who are also your customers, they talk to you more.
• 33:20
  They ask you for support when this goes about regulation of AI.
• 33:24
  They want to have your opinion.
• 33:25
  So you create actually a better relationship.
• 33:31
  with the public administration, which always is a large buyer of many companies.
• 33:35
  So, of course, those are not so intangible that they appear on the balance sheet, but
• 33:40
  I think those are all recognized factors that are positive.
• 33:47
  You recently left Telefonica, although I know you're still very involved in this
• 33:50
  whole area of responsible AI in various contexts.
• 33:53
  Tell us what you see as the biggest issues going forward that companies and other
• 34:00
  organizations need to focus on.
• 34:04
  I think on the short time for companies, it is, I mean, depending of course, where
• 34:10
  they are, but in Europe is be prepared for the regulation.
• 34:15
  And that for some companies is going to be like the GDPR was a lot of work.
• 34:19
  So it will be a lot of work, but it is not something completely new.
• 34:25
  We kind of copying the process we did for GDPR with a different content, some maybe
• 34:32
  a bit more difficult.
• 34:34
  But so it will, that is, I think, one of the things that is coming from a company
• 34:41
  perspective.
• 34:43
  I think another thing that is, should be coming and is maybe not yet coming are
• 34:47
  more the indirect and more global effects of the use of artificial intelligence.
• 34:56
  So what happens to the future of work is in the end.
• 35:01
  Creating more work or is it creating less work if it creates less?
• 35:05
  So what does it mean for the financial sustainability of states?
• 35:08
  Because there will be less income taxes.
• 35:11
  Is this the time for Robo tax?
• 35:13
  Is it the time for UBI?
• 35:15
  So those discussions are not dealt with at the moment with regulation.
• 35:23
  But I think they are important.
• 35:24
  There is also an issue on equality and gaps.
• 35:29
  So this is.
• 35:31
  very powerful technologies.
• 35:33
  It helps increase global GDP, it helps increase national GDPs, and also market
• 35:40
  caps and profits of companies.
• 35:42
  But if you look at how this wealth is distributed, then it's ending up in a few
• 35:47
  countries and in a few companies.
• 35:51
  So there is a challenge on how to make that a fair system in the long run.
• 36:00
  Otherwise, if we keep going as we are, so the companies who are doing this now, they
• 36:04
  will become bigger and bigger and bigger and more powerful and more powerful, and
• 36:08
  they will become richer and richer and richer at the expense of other people.
• 36:14
  Yeah.
• 36:15
  So the question of leaving nobody behind is very important here and is not dealt
• 36:22
  with either.
• 36:23
  So regulation takes things into account of the direct impact of AI.
• 36:28
  within the scope of individual organizations.
• 36:33
  But the global issues that survive the use of this technology at large, they are
• 36:40
  still only part of discussions.
• 36:44
  Okay, much to think about.
• 36:46
  Richard, I want to thank you so much for joining me on the road to accountable AI.
• 36:50
  Okay, it was a pleasure to be here, Kevin.