Latest / The Road to Accountable AI / Karine Perset: Building Bridges for Global AI Governance

Transcript

Karine Perset: Building Bridges for Global AI Governance · The Road to Accountable AI

• 0:00
  Kareem, welcome.
• 0:01
  Thanks for joining me on the podcast.
• 0:05
  Thank you, Kevin.
• 0:06
  I'm delighted to join you today.
• 0:09
  First of all, for those who aren't familiar, what exactly is the OECD and how is it different from other larger intergovernmental organizations?
• 0:21
  So the OECD stands for the Organization for Economic Cooperation and Development.
• 0:28
  It's an intergovernmental organization that helps countries better design policies for economic and social well-being, or better policies for better lives, as we say.
• 0:41
  It was founded in 1961 after the Second World War II to basically help implement the Marshall Plan.
• 0:49
  uh The OECD now has 38 member countries, so they're market-based democracies like the US, the UK, Japan, Germany, and Australia, for example.
• 1:01
  uh We cover nearly all policy areas, basically mirroring the structure of a national government uh and focus on the areas in which governments need to cooperate and find
• 1:15
  common solutions to common challenges or
• 1:18
  learn from one another or actually need to cooperate because they're interdependent.
• 1:25
  So we cover all policy areas except for defense, which we leave to our colleagues at NATO.
• 1:32
  The OECD acts as a convening forum, ah as I said, for learning from one another, uh countries learning from one another in each branch of government.
• 1:46
  um
• 1:48
  So that's where they coordinate action on issues like uh education, innovation, inclusive growth.
• 1:57
  I think the OECD is best known for its data-driven reports, global policy recommendations, and benchmarking tool like things like the PISA education rankings that come out every two
• 2:08
  years that high school students and their cognitive abilities.
• 2:17
  abstract reasoning abilities and components like that.
• 2:20
  And that helps schools know where they compare on the international scale.
• 2:25
  uh Tax policy frameworks as well are quite important uh and trustworthy AI guidelines in our case, uh which we'll talk about more today.
• 2:40
  Great.
• 2:40
  So then tell me how did the OEC get involved in AI policy?
• 2:47
  So our work on AI policy has been ongoing for nearly a decade since 2016 when we convened policymakers around an OECD AI foresight forum.
• 3:03
  And since then, our work has culminated around two main areas.
• 3:08
  One is the OECD recommendation on AI, which was from 2019.
• 3:13
  And the other is the other
• 3:15
  big uh chunk of work, I'd say, is the OECD.ai Policy Observatory, which helps to bring those principles to life through data, live data, where feasible trends, tracking trends
• 3:30
  worldwide, uh and tools uh for finding trustworthy AI tools uh or metrics uh with partners and other resources.
• 3:43
  The OECD recommendation on AI was the first intergovernmental uh standard on AI.
• 3:51
  It was adopted in May 2019 and was then adopted by the G20 in the form of the G20 AI principles.
• 4:00
  It was uh further updated recently in last year in 2024 to consider uh technological developments.
• 4:09
  It was lightly updated.
• 4:11
  um
• 4:13
  I was going to ask you about that, about the updates.
• 4:15
  Obviously, things change very fast in this area.
• 4:18
  So what are some of the ways that those principles have
• 4:23
  um Well, the recommendation overall aims to foster trust and innovation in AI by promoting the responsible stewardship of AI uh and ensuring respect for democratic values and human
• 4:38
  rights um and complementing uh existing standards in areas such as digital security risk management.
• 4:49
  uh
• 4:50
  uh The recommendation focuses on AI specific issues and sets a standard that's implementable and flexible to stand the test of time.
• 4:59
  uh When the updates pertain to um requests by particularly the European Commission um to have uh a definition of an AI system that
• 5:17
  could use and that would be a universal definition so that we didn't have a definition in Europe, a definition in North America and so on.
• 5:33
  And so we're talking about the same thing when we're talking about artificial intelligence.
• 5:37
  So that was one of the updates where we made a few updates because at the time of writing in 2019, we had not
• 5:47
  considered AI systems that in some instances will set uh sub-objectives based on overall objectives provided by humans.
• 5:59
  So we had written in the definition of 2019 that were human specified.
• 6:05
  uh just to be technically accurate in some cases, uh the sub-objectives or the actions to achieve a goal are not
• 6:14
  going to be detailed by a human.
• 6:16
  They ah will um be ah figured out by the AI system.
• 6:25
  So that was one of the changes.
• 6:26
  The other changes that countries, which was not really uh needed because it was already encompassed, was really to add the word content generation, even though it was included in
• 6:38
  other types of actions.
• 6:40
  uh Countries felt at that point in time,
• 6:43
  In 2024, was the language model, large language model oh generative AI period that they wanted that reflected specifically.
• 6:59
  so that those, think, were the two main changes.
• 7:06
  I just suddenly said you stopped recording.
• 7:11
  I think we're okay.
• 7:13
  I think it's looking at the other window is still open on your other computer.
• 7:20
  I'm sorry.
• 7:23
  Perhaps I can continue after.
• 7:27
  So things did not stop there last year.
• 7:32
  The OECD, um which is, as I mentioned, a group of 44 uh democracies that are market economies, and the Global Partnership on AI, which was a somewhat related initiative.
• 7:49
  um GP, as we call it, decided to join forces as the number of bodies interested in AI multiplied.
• 8:04
  We tried to consolidate to avoid duplication, uh to create synergy.
• 8:11
  So we joined forces to advance an ambitious agenda on AI uh embodying these
• 8:18
  principles, the OECD principles.
• 8:20
  And so this newly integrated partnership initially brings together OECD members and GP countries.
• 8:27
  uh So 44 in total across six continents.
• 8:31
  um And it aims to welcome new members, including developing and emerging countries that are committed to implementing these principles.
• 8:41
  um it's OECD members plus
• 8:47
  large countries like India, Brazil, Argentina uh and others and hopefully we have a list of countries which are uh candidate accession countries so we will find out more later on
• 9:06
  this year.
• 9:07
  The group should grow.
• 9:12
  How does the OECD bring all of those parties together, given that there are so many different views and so many technical issues as well in terms of doing things like
• 9:22
  defining AI?
• 9:26
  I would, I have, I think the core value or one of the important components that the OECD brings to the equation is uh basing discussions on evidence.
• 9:41
  So we're not, we have data, statistics trends that are from national statistical institutes or from ah other sources that are
• 9:53
  that really help us make well-informed decisions and advise, you know, help countries discuss the best decisions.
• 10:07
  And I would say, and probably even more important than that is engaging with a variety of stakeholders and with partners from around the globe, uh which is at the heart of the
• 10:19
  OECDs.
• 10:21
  a world and this has been continues to be the case in the work on AI and from businesses to the technical community and standards bodies and civil society as well as trade unions.
• 10:39
  I think the value is in the exchange between governments and those that are in the labs doing the work or
• 10:50
  trying to protect citizens or that really is the value add.
• 10:58
  um We're anchored in the OECD principles, so those are very solid uh and um that we have a very large uh multi-stakeholder expert community and it's today about 800 experts.
• 11:17
  uh
• 11:19
  that feed into the discussions at the OECD and within the global partnership on AI and uh brings tremendous value to governments in terms of specific topics that they need to work
• 11:34
  on.
• 11:34
  This expert community aims to translate the principles into practical guidance.
• 11:42
  This uh expert community has 10 expert groups today.
• 11:47
  that respond to the priorities of member countries, help them on complicated topics, uh help them understand also evolving topics uh where AI systems are no longer typically one
• 12:03
  system, but multiple systems with consequences from that interaction.
• 12:13
  uh
• 12:16
  I won't go through all of them, but just to give you a few examples, one of the groups looks at AI incidents and hazards.
• 12:24
  so that is one of the evidence sources.
• 12:30
  it's not hypothetical incidents and hazards.
• 12:35
  It's actually what has gone wrong, where risks have materialized into actual incidents or could have materialized into actual incidents.
• 12:46
  um And that group has recently developed a common reporting framework so that incidents are reported in much the same way in all countries.
• 12:57
  Because if they're not reported in the same way, we can't compare them.
• 13:01
  We can't track trends globally.
• 13:04
  And that means that we're bound to reproduce errors made elsewhere.
• 13:10
  So that's an important component.
• 13:13
  uh
• 13:14
  We have an expert group on AI futures, which conducts foresight exercises to assess opportunities and risks of AI technologies in the medium and long term.
• 13:25
  And so this is more, know, each group will have different, a different balance of constituents.
• 13:32
  And this would be more research oriented lab, technical folks working in labs on the products that, you know, we will see tomorrow or.
• 13:41
  uh in a few years, maybe a few months, because they're going very fast.
• 13:48
  The uh expert group uh on AI risk and accountability is working on practical guidance for companies on AI value chains.
• 13:59
  uh And then other groups consider the measurement of AI compute, which is actually quite difficult to compute.
• 14:07
  It sounds like it shouldn't be, but it is.
• 14:11
  were the first ones that uh we will come up with measurements in September, uh or how AI affects workers and the future of work.
• 14:24
  And those are just a few examples.
• 14:26
  So all of these expert groups feed into the OECD.ai Policy Observatory, which uh listeners can visit at OECD.ai.
• 14:35
  It's got a wealth of resources and tools.
• 14:38
  uh
• 14:40
  and over 250 indicators on various dimensions of the AI ecosystem, as well as right today, I think 1,500 policy initiatives from 71 countries.
• 14:57
  And that's growing rather quickly because more and more countries are inputting their own national AI policies or strategies to be able to learn
• 15:09
  develop as a group best practices, but also learn from what other countries are ah doing right and have succeeded in, especially the leading countries in AI.
• 15:22
  But perhaps even more importantly, where they have gone wrong, what types of initiatives have gone wrong, depending on the way you look at it, both are important.
• 15:34
  this repository of government uh policies
• 15:38
  is the largest uh government-backed and up-to-date repository of AI policies in the world.
• 15:45
  You can, again, it's at oac.ai um and it's countries and you'll see all the countries covered there, all the policies of those countries uh up to date.
• 15:56
  I think the important component here is not to have a database of national AI policies, it's to have a national database of AI policies that is up to date and that's the hard
• 16:07
  part of this work.
• 16:08
  uh because policies are popping up every day, so we have new initiatives all the time.
• 16:17
  uh The live data also includes indicators on AI news, scientific publications, research publications, investments, AI jobs and skills, software development, uh search trends, and
• 16:32
  much more.
• 16:33
  So taken all together, they give you a
• 16:36
  picture of where a country stands and we're developing an index, uh an AI index that is very AI specific.
• 16:45
  Many AI indexes exist, but they're not necessarily AI specific.
• 16:49
  They take into consideration other components that are related, digital related, but not necessarily AI specific.
• 16:59
  And this one is uh all the indicators that compose it are AI specific.
• 17:06
  So uh the portal really leverages this live data to show timely trends about where, how, and at what rate AI is being developed and used and in which sectors.
• 17:22
  So it's a powerful tool to keep up to date.
• 17:27
  And I'd encourage listeners to check it out.
• 17:31
  Fantastic.
• 17:32
  And then that index would uh track countries' maturity in AI, or what exactly would it seek to measure?
• 17:42
  Yeah, it's maturity in AI based on uh its development in policy areas, uh in uh research and development, in uh investment in research and development.
• 17:57
  So investment is an important part, and investment in infrastructure, uh in infrastructure, in data, and the availability of data, of uh compute.
• 18:12
  of algorithms of knowledge about those things, availability of uh an enabling policy environment is obviously critical, uh as is the availability of talented workforce.
• 18:32
  And uh in our case, also, uh well, not in this scenario, but a talented workforce that has the right skills.
• 18:41
  for AI today because AI is evolving and if you're looking at the skills that were required in the AI field a couple of years ago, they're very different from those that are required
• 18:51
  today.
• 18:52
  um So it's really the real-time nature of that type of data is very important.
• 18:59
  It's also looking at news and at trends and at forecasts.
• 19:04
  So forecasts, you know, a few one, two, three years down the line, where can we expect...
• 19:11
  the needs to be and sometimes 10 years down the line, which is harder.
• 19:20
  critically important.
• 19:21
  I can absolutely see that.
• 19:22
  yeah.
• 19:23
  It also looks at international cooperation, international agreements uh on AI, which is an important component to have uh interoperable frameworks that can work together so that
• 19:41
  countries can benefit from what other countries are doing uh in specific sectors.
• 19:46
  For example, applications in agriculture that can be exported to other countries.
• 19:52
  useful elsewhere.
• 19:54
  What's the biggest challenge right now in terms of international coordination or even cooperation on the kind of data sharing that you're talking about?
• 20:07
  I think the biggest challenge for international coordination uh or harmonization rather because around AI governance is uh international wealth.
• 20:23
  I'll start by saying that international cooperation on AI policy is important because just like the internet, AI knows no borders.
• 20:32
  And so what's developed in one jurisdiction will be available to maybe in some places to advanced users and others to everyone.
• 20:46
  And this is especially true since the generative AI boom uh of the late 2022.
• 20:53
  uh But like, you know, as a general purpose technology like electricity, uh AI
• 21:01
  is already or will very soon be critical to pretty much, actually it might even dare say it is, to pretty much every sector from agriculture to healthcare in most countries.
• 21:21
  And the diffusion of AI is ramping up, the diffusion rates are ramping up, which has real impacts in a world where
• 21:29
  many firms do business internationally and that's been the default for a long time and that's something that's very hard to, I mean, that's very lengthy to develop and to change
• 21:43
  uh as the case may be.
• 21:46
  So having global standards for trustworthy AI uh is critical to provide businesses with stability uh and predictability.
• 21:56
  uh
• 21:57
  and to give users assurances that AI applications that they're uh using are trustworthy, robust, and secure.
• 22:06
  uh And in addition, many systems around the world are using similar core AI algorithms to train AI models.
• 22:14
  This means that many countries are actually vulnerable to the same AI-related risks like security issues or...
• 22:23
  uh
• 22:24
  or infringement of human rights.
• 22:26
  And so we need to take these some safeguards into international guidelines and standards.
• 22:32
  uh The challenges to achieving that are significant.
• 22:36
  uh Geopolitics today have become more complex than they were a little while back, but not insurmountable.
• 22:46
  uh And at the heart of the work that we do at the OECD, uh things are taking longer.
• 22:54
  the discussion is still taking place.
• 22:56
  ah It's not as oh rapid ah and perhaps not as rapid as it could or should be, but it's taking place and it needs to continue to take place.
• 23:09
  um And those challenges also stem from what could be seen as strengths at the same time.
• 23:18
  For example, diversity among member countries and particularly if we talk about
• 23:23
  the global partnership on AI, has diverse levels of, know, in diverse countries in terms of size, level of maturity and so on.
• 23:34
  Like India has different priorities uh than some European countries do or the United States.
• 23:43
  uh And so, and different levels of experience, uh different histories, cultural sensitivities.
• 23:50
  and economic characteristics, economic industrial structures that were in which AI is being embedded that lead to understandable differences in how they'll approach things like
• 24:02
  uh AI or uh AI in tax or AI in education systems.
• 24:09
  um But let me continue with the real opportunity here because it's...
• 24:17
  absolute harmonization uh between countries is impossible and it's not necessary most of the time.
• 24:24
  mean on most issues it's impossible with different backgrounds.
• 24:32
  But that doesn't mean that approaches should be disjointed patchwork and work against each other.
• 24:39
  this means that the timing now
• 24:45
  especially as geopolitics are uh becoming more sensitive, the timing is increasingly critical and that AI laws, regulations and standards currently being implemented uh or in
• 25:00
  varying levels of development are uh critical for the future.
• 25:06
  um So the opportunity for further alignment uh is what we strive for uh in our work at the OECD.
• 25:14
  uh from the work I mentioned on creating a common definition of what we're talking about or providing a definition for AI incidents so that we can accurately report and compare
• 25:28
  trends across country and avoid that uh type of uh recurrence of issues.
• 25:38
  The OECD- excuse me.
• 25:42
  Sorry, I was sick about a week ago in the coffin.
• 25:44
  I'll not go away, but we'll flip it.
• 25:48
  The OECD also has a role with the G7's Hiroshima process, which has a voluntary private sector reporting component to it.
• 25:57
  Can you talk about how that fits in with the other work that you're doing?
• 26:02
  Sure.
• 26:03
  um So the Hiroshima AI process was begun uh three years ago uh under the Japanese presidency of the G7 and uh developed a of conduct uh for responsible AI.
• 26:26
  um And the following year,
• 26:32
  Last year, 2024, the Italian G7 presidency followed up on that work by developing a reporting framework.
• 26:40
  Well, by asking, tasking actually the OECD to develop a reporting framework for companies to explain how they're implementing trust with the AI, what measures they're taking at
• 26:53
  different stages of the AI system lifecycle.
• 26:57
  And so the OECD, we've been working with
• 27:00
  uh big, ah small, ah largely, mostly big because the work originally targeted ah developers of advanced AI systems.
• 27:16
  ah And so we've been working with those companies to develop ah voluntary reporting transparency uh measures ah that has proven very helpful.
• 27:30
  uh to governments to understand what companies are doing and they're doing a lot.
• 27:36
  They're doing a lot in all of the areas where they should.
• 27:40
  And they're also, it's also helping the companies in terms, because large, very large companies like Microsoft, which is doing a tremendous amount, uh is so big that uh this
• 27:52
  kind of having one reporting uh framework and...
• 28:00
  Forces, forces, uh coordination, internal coordination, uh which, you know, you're always running and, and, and other, there are always priorities between that coordination helps
• 28:16
  everyone understand how they fit into the bigger picture.
• 28:19
  uh And so all of this, the, the, the framework was launched, as I said, earlier in this year in Paris.
• 28:29
  at the AI Summit.
• 28:30
  It's 20 companies reported, but many more are interested.
• 28:35
  uh And also not just uh developers of advanced AI systems, but also deployers and also uh analysis different types of organizations and of different sizes.
• 28:53
  we have SMEs versus uh
• 28:57
  uh very large companies versus ah scale ups versus ah and then we also have uh companies that are operating in business to consumer uh B2C and companies that are like Salesforce
• 29:17
  operating mostly in B2B and they'll have different practices will rely also on the
• 29:25
  the business models of those different companies.
• 29:27
  ah we have a task force that is uh exchanging good practices on what they're doing on different topics and what works and what works less well.
• 29:40
  And uh that's been very helpful.
• 29:46
  uh These results are facilitating transparency and comparability of
• 29:53
  risk mitigation measures and contribute to identifying and disseminating good practices.
• 30:01
  we expect to see that it's now broader than a G7 initiative in terms of the companies reporting.
• 30:14
  A number of countries are reporting and more countries wish to, particularly countries from the Global Partnership on the Eye like India, which is...
• 30:22
  many, many deployers of AI systems and a lot of developments like Infosys and others.
• 30:34
  And so we see this as a bit of a bridging the gap while there's some regulatory movement and uncertainty.
• 30:43
  This is really what the companies are doing and they're keen to share that because they know that
• 30:53
  the businesses do rely on trust and they are doing much, very much on all fronts from security to preventing other types of harms and testing and so on.
• 31:13
  all the results are public, I mean, everything is publicly available on OECD.ai, of course, needless to say, I'm not advertising.
• 31:21
  I'm just saying that anybody interested can see the detailed reports of the companies that have submitted.
• 31:32
  We've drafted a summary that should be online soon as well.
• 31:38
  Going forward, are there any particular either technological or market developments that you're looking at that might shift the uh environment for AI governance?
• 31:59
  There are some significant shifts where think everyone is aware of on how some countries approach AI regulation and view AI regulation, some as an enabler of trust and some as an
• 32:25
  impediment and as a...
• 32:28
  you know, as an impediment to competition, worldwide competition.
• 32:34
  ah How does the OECD address issues when its member states may change due to elections, new administrations that happens all the time?
• 32:47
  that's not, there's nothing special about it.
• 32:50
  ah We're a consensus driven body, so ah policy shifts.
• 32:56
  uh
• 32:57
  after elections are normal and expected.
• 33:01
  And our job is to create continuity uh by convening countries around the table and anchoring discussions.
• 33:10
  And that shared evidence that I was mentioning earlier, the long-term values that uh these countries uh cherish uh and peer-reviewed policy tools.
• 33:21
  So that enables the OECD and JPEG uh more and more.
• 33:27
  policy advice to stand the test of time even when the political oh winds do shift and they do shift often.
• 33:36
  Indeed.
• 33:38
  Anything else that's on your agenda for the coming year?
• 33:43
  For 2025, think maybe two lenses uh at an operational level with the new integrated partnership with with GPA that we're working hard to implement the annual work plan and to
• 34:00
  strengthen the merger with uh new countries working with these new countries that we hadn't worked so closely with before like
• 34:10
  like India, Senegal, and Africa, uh Brazil, and others, at a policy level, uh hand in hand with this expanded membership, thanks to GPay and in view of further expansion because we
• 34:25
  have uh more countries which uh we hope will join soon that are doing incredible things in AI.
• 34:34
  So we're focusing on building tools to help countries at varying levels of
• 34:39
  AI uh development to implement the AI principles.
• 34:50
  They also have, know, but in a localized manner.
• 34:55
  so that's not just, you know, here's a guidebook if you're a developing country.
• 35:01
  No, it's a lot of um co-creation workshops.
• 35:05
  So we've been doing work
• 35:07
  a lot of work with the uh African Union Commission, with the ASEAN countries, with Latin America.
• 35:16
  ah And so that's, think, an important direction that we're taking.
• 35:26
  We're particularly developing an AI policy toolkit, which countries can use to help them set up national frameworks.
• 35:34
  programs and tools to guide national AI policy that's aligned with the AI principles.
• 35:41
  We have numerous other work streams that are also priorities, but I'll just mention this one.
• 35:47
  uh And listeners can find out more if they're interested.
• 35:53
  Absolutely.
• 35:53
  That's it's a lot.
• 35:54
  ah But it's great to hear about all the different work that you're Kareem, thanks so much for being with me.
• 36:01
  Thank you.