Launch of the preliminary report by the Independent International Scientific Panel on Artificial Intelligence - Press Conference

All right. Good afternoon. In just a short while, I will bring out the Secretary-General, and we're delighted to be joined also by the co-chairs of the Independent Scientific Panel on Artificial Intelligence, Maria Ressa of the Philippines and Yoshua Bengio of Canada. Who are joining us remotely. I think Joshua, you're in Canada. Maria, I understand you're in an airport in an undisclosed location, but when you speak, you could let us know where you actually are. They are here to present us the findings of their preliminary report ahead of the inaugural meeting of the Global Dialogue on AI Governance, which, as you well know, will be held in Geneva next week on the 6th and 7th of July. The Secretary-General will make some brief opening remarks and he will leave us and then we'll hear from the co-chairs.

A very good morning. We are focused today on the future. But before we begin, allow me to say a quick word about yesterday's decision by the General Assembly to strengthen the United Nations for the future. You were just briefed by the Comptroller on a new reform of our financial rules. For far too long, the UN budget process has faced a fundamental contradiction. As we know all too well, Member States do not always pay their assessed contributions in full and on time. Yet, we were retired to return funds that we had not spent because we had not actually received them due to unpaid assessed contributions by Member States. This means that we have been hamstrung by a double blow: on one side, unpaid contributions; on the other side, an obligation to return funds that never arrived in the first place. In other words, we were trapped in a kafkayesk cycle. Expected to give back cash that did not exist. Yesterday, Member States agreed to change that. Under a new methodology, on a trial basis, unspent funds will be returned only when they are backed by cash. I have been pushing for this reform for almost a decade. It will allow us to manage resources more responsibly, protect the continuity of our operations, and better support critical mandates, including peacekeeping. And I am grateful to Member States for taking this important action, which will greatly benefit the incoming Administration and contribute to the financial stability of the United Nations in the coming years. But today we are here for a different reason. We are here to present an extremely important UN initiative on artificial intelligence, that was decided by the General Assembly in the follow-up of the Fact for the Future: the Independent Scientific Panel on AI. I am pleased to be joined today by its co-chairs, Yoshua Bengio and Maria Ressa. A few months ago, I came before you to announce the formation of this panel, comprised of 40 distinguished individuals from around the world. It is an extraordinary and unique group—the first global, fully independent scientific body dedicated to helping close the AI knowledge gap and assess the real impacts of AI across economies and societies. As I said at the time, the Panel is intended to help the world separate facts from fakes. And science from slop. We are looking to them to provide an authoritative reference point at a moment when reliable, unbiased understanding of AI has never been more critical. And I am pleased to say that they have delivered a down payment on that commitment—in record time. Indeed, the Panel is releasing today their first assessment on the eve of the inaugural AI Dialogue next week in Geneva. My thanks to them, and to every one of the Panel's 40 members, for their hard work and insights. This morning, their assessment goes to every government. And it is open to the public for everyone to see. Their report is honest about the extraordinary promise of this technology. Used well, AI could be the most powerful engine for development. Speeding the world's progress on everything from health and hunger to learning and climate. But the Panel is just as clear-eyed about the harm artificial intelligence can cause. The co-chairs will go through their findings in detail. But first let me draw a single lesson: the more AI advances without shared rules, the less say governments and people will have in the outcome. So my message to governments is simple: do not wait. This is a preliminary report. The Panel will keep working as the technology evolves. And it does not work alone. Next week in Geneva, the first Global Dialogue on AI Governance will begin to turn science into shared action, with every nation at the same table. I look forward to joining Member States there to help carry this work forward. And soon, I will set out proposals to help countries build the capacity to adequately deal with this technology and share in its rewards. The Summit of the Future asked whether international cooperation could keep pace with the speed of technology. Today offers one answer: the science is here. We can no longer say we did not know. What we do with it is now up to all of us. Dear members of the media, in full respect for the panel's independence, I will now leave you in the capable hands of the two co-chairs. Joshua and Maria, the floor is yours.

Thank you. Thank you very much, sir.

Great. Thank you very much. I think, Maria, you are set to go first. So if you can, in the spirit of transparency, let us know where you're speaking from and we'll give you the floor and then we'll go to Joshua.

Please, Maria.

Sure. I'm in Amsterdam at Schiphol.

Excellent.

Well, first of all, I want to repeat, thank you so much for giving us this time. Thank you, Secretary-General. And also to just repeat something that you had said. This is a phrase that stood out for me: "We can no longer say we did not know." Right? For someone who feels like Sisyphus and Cassandra combined, "We can no longer say we did not know." I think that's what today is. This report is the moment the world stops being able to say it didn't know. Right? A little bit about what we did. You heard from the Secretary-General. There are 40 scientists from 37 countries created by the UN to be independent. Our commitment is to the facts, the evidence, and to humanity. Physicists, doctors, economists, computer scientists. We have a philosopher and yes, a journalist. People who had never worked together, appointed in February, asked to assess the most consequential technology in human history. And to do it by now. We wrote this together in about a month and a half. This is a preliminary report. It's the best available evidence at this moment in a field that changes faster than any of us can write about it. The numbers we report run through May, and in this field, that already feels like a while ago. It is honest about what we know. It is also honest about what we don't know. And it was built to work at two levels at once: for governments, who are our primary audience, and for you, the public, who are never far from our minds. The technology is extraordinary. We use it. The opportunities are real. And we name some of those in science, in education, in healthcare, in agriculture. AI is already changing what's possible. And we believe in what it could deliver for humanity. But the dangers are just as real and they are happening now. What you're receiving is the floor of our concern, not the ceiling. You know, when you get 40 scientists from 40 different contexts and you have to make them agree, you don't drift toward the most alarming claim. You build toward the center. The most contested findings demand the most evidence and the most negotiation. So everything in this report cleared that bar. It is the minimum we all agree on, and that is alarming enough. I'll give you 3 headlines and my co-chair, Joshua Ben-Joah, will take you deeper. 3 headlines: pace is not slowing, the power is concentrating, and control is not guaranteed. The pace is not slowing. It is accelerating. You take just one of the benchmarks in the report. It's called H.L. EHumanities last exam. It's a test of 2,500 questions deliberately designed by experts to be too hard for AI, at the very edge of human knowledge. In 16 months, 16 months, top scores went from 8% to 45%. Now, the power that's concentrating a handful of companies and a handful of countries are making the most consequential decisions about humanity's future. The United States alone accounts for 75% of the computing power in the world's largest AI clusters. And control is not guaranteed. No expert today can promise you that the most advanced systems will do what you instruct it to do. And Yoshua can tell you more. In laboratory settings, these systems have already been shown to deceive and to resist being shut down. I've spent my career on one subject: what happens to a society when the information it runs on is poisoned. So let me be direct about the part that keeps me up at night, what we have studied in the Philippines and around the world. A lie told a million times becomes a fact. I lived that. I watched it shred a reputation built over decades in a matter of days and make the illegal look legal. Now you add generative AI. It runs on the same playbook— faster, cheaper, more personalized to each of us. The three sentences I've said over and over in a decade: without facts, you can't have truth. Without truth, you can't have trust. Without these three, we don't have a shared reality. We have no democracy. That's the trend we're seeing, right? That is not speculation. It is in the report with the evidence attached. But we're not telling you the future is written because it isn't. The single most surprising thing I learned on this panel was this: the hardest part was not the differences between nations. Scientists from every region across every divide you can imagine looked at the same evidence and arrived at the same place. Science gave us a common language that politics often can't, right? If 40 of us strangers from February can hold that line together, Then I hope governments can too. The governments can. The window to act is open. This report is the common ground to act from. Joshua, as brilliant as he is, principal, will take you inside some of the other findings.

Thank you. Thank you, Maria. Joshua, please go ahead.

Thank you. Thank you, Maria. I will go in a little bit more, but not much. The same themes that Maria told you about. I want to thank Secretary-General Guterres as well as everyone joining us today. And of course, this is the beginning of a discussion that we hope to amplify in Geneva with the Global Dialogue and the ITU AI for Good Summit. As Myriam said, The work that we did together was surprisingly effective given the little time we had, and we're very proud of it. We came together to create an evidence-based independent assessment of the current state of AI science, and I'll come back to the importance of that. Because the goal is not to make recommendations, but to ensure policy decisions are informed by the highest standards of scientific integrity regardless of external pressures and preferences. So that's our mission. And I believe it is of the utmost importance at this point in time because AI and the time in which we are now represents a turning point. It's about the growing intelligence of machines. And you have to realize that intelligence gives power. As this power grows, it can unlock great benefits if we act wisely, but it can also lead to many perils. In particular, is some— in humanity are reckless or seek to abuse the power that AI could give them, especially if most of us remain passive, mainly, mainly because we might not understand what is going on. So the decisions we're making today about AI will have lasting consequences for individuals, businesses, institutions, democracy, and maybe more than that. Our goal here is to help understand AI's opportunities, risks, and impacts. Let me unpack a few elements that you'll find in the preliminary report. First, as Myriam said, Research shows very rapid and accelerating progress of AI capabilities across domains. At the same time, as there are significant gaps in the societal infrastructure and scientific methods to reliably evaluate, secure, and govern these systems. Technical progress has proceeded quickly, doubling every few months according to some metrics, for several years now. Meanwhile, there are still no known technical guarantees that AI will follow our instructions, norms, or laws, and instead, as she mentioned, growing evidence to the contrary, including AI acting deceptively. But things would be simple if, on the other hand, it wasn't also the case that AI was unlocking useful applications across domains like health, agriculture, science, um, while its adoption and its benefits are accelerating unevenly across countries and sectors. And its development entails risks to human rights, social systems, and the environment. Looking ahead, the gap between rapidly improving capabilities and effective risk management methods may lead to catastrophic outcomes. We've seen on our radar screen recently the issues with cybersecurity, but companies are worried about bio-risks, and many in civil society are showing that AI can lead to disinformation that, that can threaten our democracies and our individual rights. So I want to go back to this question of concentration of power. Because as AI capabilities continue to grow, if they continue, the gap between the haves and the have-nots is likely to grow. And I don't think we're taking the full measure of that. To realize the full benefits of AI while minimizing its risk, we need to tackle all of these issues. And it has to be at the international level because AI, and in particular its misuses like cyberattacks are not limited by borders. But it's— I understand it's tricky for policymakers because governments face what has been called an evidence dilemma. They need evidence to make informed governance decisions, but the evidence often lags behind the pace of AI development. So the next steps need to be taken together to shape and steer that development in a beneficial direction. And it's really important in that process that the panel sticks to its mission, not making recommendations, but laying out the facts that scientists can observe and help the global discussion about a coordinated international and democratic approach to make sure the decisions about our future are taken in our global interests. So I look forward to continuing this work with my esteemed colleagues, and I thank all of them for their exceptional contributions. Thank you.

Thank you, Maria and Joshua. Thank you very much. We'll take some questions. Valeria.

Thank you, Steph.

Uh, thank you so much for this briefing on behalf of the United Nations Correspondents Association. So, uh, it's Valeria Robeco from ANSA Newswire. Uh, US and China together control about roughly 90% of global AI computing power. So what concrete commitment, uh, will you push for in Geneva next week, uh, to stop this warning from becoming another unenforced framework. Thank you so much.

Great. Maria, do you want to take a crack at it first and then Joshua?

Sorry, let me just turn that. Yes, yes. Um, uh, well, first, our task is to give you the facts, to give the government the facts, right? That's why I loved what the Secretary General said, which is you can't say you don't know. So we give the facts, but we're not prescribing policy. That goes to the member states. But we've certainly— if you look at the floor of the danger ahead of us, you pointed that out. We identify them and then what the governments decide to do, then that is up to them. We're— it's— we can't prescribe it, but we can, we can tell you where the world is heading.

I completely agree. I would add that We need a global discussion involving everyone, not just the governments, so that we all understand the situation.

Thank you.

Gabriel, then Chris.

Thank you very much. This is Gabriel Elizondo from Al Jazeera English. Thanks for your briefing on the important report. My question is very simply, the panel did not really give a recommendation about any sort of analysis of new models that are coming out. And this is one of the most pressing issues right now in the space of AI. As you very well know, the US government took away Mythos and Fable 5. They just reinstated it last night. This is not a theoretical discussion. This is something that is taking place hour by hour, minute by minute. Why did the panel not give any sort of recommendation on if there should be some sort of international body and/or organization of some sort that should analyze models before they are released to the public? Thank you.

Joshua, do you want to go first?

Same answer as we gave just before. Our role is not to make recommendations, but I think the issues you're pointing out, we are very well aware of, and they have motivated the directions we have taken to investigate what needs attention. So thank you for asking, and let's hope this starts a real discussion.

And I'll just add to that one, right? The— I guess it, it may sound like a cop-out that we're not making the policy recommendation, but frankly, when you have the scientists together in the room, they tell you what it is, what they know, what they don't know yet. So I think you look at this report as policy relevant but not policy prescriptive by design. That's why it is usable in Washington and Beijing and Manila. Right. The, the prescribing happens next week in Geneva when we hand it to the Global Dialogue where the states sit at the table. But obviously you've seen that there is a box in the report that looks at Mythos for one. Right. So it's there. The facts are there.

Yeah, I just want to add that there are good reasons why the work has been split between the Scientific Panel and the Global Dialogue. If we were to make policy recommendations, our work would become very politicized and it would pollute our ability to provide the the scientific evidence. So it's really important.

Sorry, last again. See what happens when you ask. I would say our goal, you know, because I always say that the tech has torn us apart in different realities. What the report will hopefully do for member countries of the UN is to come bring us together to the same reality.

Thank you, Chris, then Namo, then Sherwin Noreen. Hi, guys.

Chris Reyes with Canadian Broadcasting. Thank you for this briefing. This is the most represented I've ever felt in a panel before. Happy Canada Day, Yoshua, before I start my question. Um, this builds on the first question, uh, which is, uh, the power, as you guys mentioned, driving AI is concentrated in private companies with little government regulation. Are you getting any sense that tech companies and governments are willing to work with or even listen to a coordinated international panel like yours? And if not, what do you do if governments and tech companies don't heed your advice as you continue to advance this report?

Maria, do you want to go first and then Joshua this time?

Yeah, I can go first. First, she's Filipino and Canadian. Hi, Chris. You know, in terms of of power concentration. I think what we can do is, again, give the facts, right? But we give the evidence and the options. We don't prescribe it. The honest answer is the report doesn't tell governments what they are going to do about power concentration, but it does document the problem sharply and it lays out the levers that exist. Right. And then if the governments don't act— sorry to move this. This is me speaking not as a panel chair, but that's why we're coming to you. You then have to bring it to the people because the governments are in charge of their own nations. Yashua pointed out multilateralism is still the way to go. But if the citizens in each of these countries know that there's a better way or feel there's a better way, that is also up to the citizens, which is why it's important that we give citizens around the world the facts.

Yeah, I just want to add that Although we don't make recommendations, there are others who do. And there is a very live discussion about, for example, concentration of power from scholars as well as people in government around the world right now. I would bring you back to my own Prime Minister Mark Carney in his speech at Davos, for example.

Thank you so much.

This is Namo Abdullah with Rudaw Media Network. One question for Maria. As a journalist, are you more optimistic or pessimistic about the future of AI for journalists? And what do you think is the single greatest threat that AI can pose to journalism or journalists? Thank you.

I think we're living through it right now, right? I mean, I— it's been a decade that the— that facts have literally been optional, right? Which world do we live in where it's not optional, where facts become optional? Um, look, I, I think journalism now, the distribution has gone to Big Tech, uh, and that there needs to be some accountability pulling us apart, polarizing our societies. I mean, there's a reason why 74% of the world today now is under autocratic rule. We are literally electing illiberal leaders democratically. But let me point you to what the report actually says about this, right? The impact on the people is clear. We have a working group. We have experts in our panel whose, whose literal jobs— and there was a big discussion and we had talked about pulling out a thematic brief on children, right? The impact on the next generation. But beyond that, the impact on us as people. When you're pumped full of toxic sludge, which is where the first generation of AI social media happens, there's an impact, right? If you're around people who are screaming and yelling, you wind up screaming and yelling. You're changed by that. So the last part of your question, the impact on, on journalism, we're certainly feeling it. I'll quote some numbers from— this is the Medill Year in the United States. The Madill Report from Northwestern last year said more than 2 news organizations closed every week. Right. So this is creative destruction and we journalists are going to have to create even as it's being destroyed.

Thank you. I just want to add something from the technical point of view. So There are many possible futures with AI that we can choose. So there are ways to develop AI that will be heavily focused on honest answers and factuality. It may not be the most profitable way to develop AI, but it is possible to to make those choices collectively. We can choose in which direction we can take the technology. The technology is not some God-given thing. It is developed by humans under societal rules and incentives. So I think the example of the harms that AI in its current form and how it is deployed, for example, in social media can do to the role of the media and journalism is not something that we need to take as, you know, static and given. If we don't like it, we can change the rules and steer the technology and how society is deploying it in ways that we choose.

Thank you. We'll go to Sherwin, Noreen, Amélie, and then we'll go on screen.

Thank you. It's Sherwin Briceby, South African Broadcasting. I'm just reading from the report, right? "The rapid pace of technological development and the breadth of potential applications presents policymakers with significant challenges. The rapid unchecked deployment of the technology at scale also presents considerable risks..." So you won't talk about recommendations because that's not the mandate. I wonder if you can infuse this AI dialogue in Geneva next week with some urgency, given what this report says? I mean, we often talk about AI in the context that the horse has bolted and now we're coming with policy persuasions. I mean, are we not doing this entire process backwards? What is the urgency of the moment and what sort of pressure needs to be brought to bear on policymakers gathering in Geneva next week?

All right, Maria, I saw you give a thumbs up, so you go first.

Yeah. Yeah. No, I mean, I feel like I feel that urgency. Look, I think what the report does is it tells you— it puts us all together in the same place, which is that there is urgency to act on it. Right. Let's just talk. How can you act on something that isn't transparent? And this goes back to power and concentration. You can't govern concentration you can't see. So independent third-party assessment, right, like that— the report's pharmaceutical, aeronautic analogy. National and regional AI safety institutes, technical secondments for regulators, incident reporting. And then there's a structural point. Over 40 governance instruments exist, but they're fragmented and concentrated at the corporate level. So the gap isn't a lack of tools, it's coordination and independent measurement. Right. So that's the first. On the frontier technology, on the frontier model itself, the report tells you that developers The developers themselves, we know this from some of the headlines, have begun restricting deployment above internal, internal risk thresholds. But those thresholds are company-defined with no standardized evaluation or external verification. So that's the account— accountability gap right now. Restraint on the most powerful systems depends on developer goodwill. So to mirror your urgency. What we're after is transparency, exactly what journalists do, right? But now the scientists are pushing that transparency.

Thank you. Joshua, please. A few words about urgency.

But first, a very important point that report really hammers is the uncertainty of the future development of the technology. And I know it's difficult for most people, right? We want to hear one story, this is how things are going to unfold, and maybe we can change it with policy. But that isn't the reality in which we are now. We are in a reality where there are many possible scenarios for the possible advances of the AI capabilities. Some researchers think that we've reached a plateau. There is no evidence for that, but it could be next month that the companies hit a wall. Other researchers are thinking, well, there's a good chance the trend will continue, and the trend is exponentially fast. And, you know, why this is important, because currently we're not— societies are not prepared for the fast trajectory. We don't have the policy, we don't have the regulation, we don't have the international coordination. And to add on the urgency question, what many AI researchers fear is an acceleration due to how AI is more and more already now being used to do AI research. So the companies are using their own AI. In particular the ones that are available to them but not to the rest of the world, right? They have internal versions, as Maryam was talking about, to write the code for the next version of AI. And already the AIs are writing the vast majority of that code. So why is that an issue? Well, as AI is doing more and more of the work that used to be done by engineers and, Researchers, humans, that accelerates the potential advances. And this was actually something computer scientists thought could be an issue in the '50s and '60s because this, you know, they could forecast that at some point when AI, if and when AI gets to be as strong as the best human researchers in a particular field, but in particular in the field of AI, then there could be a real acceleration where the next version of AI is faster and better because it's designed by AI that is better than us or could be duplicated over millions of GPUs, millions of instances. And then that version, which is even better than what the humans would have done without AI, is designing the next version. So you can see a potential, like, what's called recursive self-improvement. Of course, I started with uncertainty because we don't know if that scenario is going to happen. It's just one plausible scenario. And going back to policy, if you're a decision maker in a government or a company, you need to take decisions that are going to be robust to whatever, you know, of these scenarios is going to happen. And it's difficult because Some of these scenarios, like the one I talked about, sound like science fiction, but the data says if the trend continues, for example, how many lines of code are written by the AI itself for the next generation of AI, we might very well get there. And people in companies, some of them are saying this might be as close as 18 months, 2 years. I, as a researcher, I'm agnostic about this. I don't know, I don't have a crystal ball. But I hear all these. Some of these possibilities are plausible. And so my role is to raise a warning here that this could happen and we're not ready.

Thank you. Noreen, then Amelie, then we'll go on screen.

Thank you very much. My name is Noreen Hossein with IPS News. The report mentions that AI usage and adoption is varied globally and that the risks that AI poses is also, is also uneven across populations and countries. The report also mentions that the Global South is apparently lagging behind the Global North when it comes to the, the use and adoption of AI. And what is that? Does it actually look like in practice? And what are the factors behind that? Does this have to do with the with the existing digital infrastructures in both in Global South countries and Global North countries? And does this also apply to the debates or the ethics of AI in public spaces and civic spaces? I mean, thank you.

Maria, do you want to go ahead first?

Yeah, actually these are two sides of the same coin. The very first question that we had, which is, you know, the concentration, so it's concentration of power concentration geographically. I think the report establishes the problem, but the concentration runs across the whole stack. And it's not just one thing on the supply side. You have single choke points that are dominating. And this, Joshua will know, but I've learned a ton in doing this. ASML at 80% plus of lithography, TSMC and leading edge fabrication, NVIDIA in chip design on compute. Geographically, the US holds 75% of the world's top 500 cluster capacity, China 15%, and everyone else 10%. On development, 91% of the notable models in 2025 came from the private sector, and US institutions produced 59 notable models to China's 35 and 13 for the rest of the world combined. You know, the, the report's framing is that this concentration of capability is becoming concentration of political power that affects development. And that's the rollout, right? So we have a case study. The myth of lockdown is the signpost. The sovereignty of most nations, that's also brought up because they depend on systems they can't build, inspect or audit. But I think the key part— sorry, just to bring it to your question, because the flip side of that coin is if you don't have compute capability, if you don't have, then how do you come in from the Global South? Speaking from someone from the Global South, right? So, we give you some opportunities in the report to actually imagine it. And it's in agriculture. It's in medicine. It's in, you know, one of our scientists, actually a few of them work on health and agriculture came up very early on. I guess, I don't know, Joshua, did I answer the question or did I give you what I wanted to tell you? Because I'm thinking through.

No, you gave— like a lot of important facts. I would add one related to what I just said to the previous question about the scenarios for the AI capabilities. And so if in the scenario where AI capabilities flatten pretty soon, then that's kind of good from the point of view of AI inequity, I mean equity. In the sense that countries will rapidly reach that plateau across the planet. And that means, you know, AI is less powerful, but we have less of these power concentration issues. But under the other scenario where the AI capabilities grow at this exponential rate, then it's very likely that the gap between the haves and the have-nots that I was talking about in my speech, that gap is likely to grow. And the economic power, the military power, the ability to influence public opinion, all of that becomes a tool of domination if we're not careful about how we globally make sure it doesn't happen.

Thank you. Amélie, and then we'll go to the Economic Times of India online.

Thank you. Amélie Bautelier from AFP News Agency. You clearly said that the panel is not there to give policy recommendation, but you also talked about urgency. So let me make a comparison with another transformational issue, which is climate change. I mean, many compared your panel to the IPCC on climate. 30 years ago, the IPCC was also a lot about describing, explaining the science, and it didn't work. Since then, many of the IPCC experts joined the fight against climate change and gave the, the lines, the recommendations to the government to what to do and to not what to— what not to do. So isn't there a limit to your panel not to be able to give any kind of recommendation? And should you not be at least given some kind of, you know, if we go this direction, that could be the end of humanity, basically, like for climate change? Thank you.

Let me try something to complement the things we've already said. First, it's not just the independent scientific panel. As you heard from the Secretary-General, the UN created also the Global Dialogue body, which has a political mandate, which has a mandate to find convergence among the member states about what to what to do in terms of policy. So at the UN level, that exists. Outside of the UN, many scientists are very active in policy discussion, and Maria and I are among them, many of our colleagues as well, because we feel like the researchers doing climate science have— that scientists have to get involved their voice matters. So it's not like we're not taking those policy questions seriously. It's that here we are part of a process, an exercise in which, in order to make sure the policy discussion goes as well as it can, we separate the science, the evidence gathering, from the discussion about what people want. What are the preferences of different people around the world, right? From a philosophical point of view, there's truth and there's preferences, like values, wishes. These are two different kinds of objects. And when they mix, the truth gets distorted. We've seen that over and over in history, including, to some extent, if you think about the kind of political wrangling around the executive summary of the IPCC. And yeah, so we have been advised by people who have seen this to do it the way we are doing it now, with a clean separation between the scientific part and the policy part.

I'll pick up. This is what we've learned, Joshua and me. We actually work well together. So, look, I love the question, but two big things. Policy fails when it waits for evidence. The evidence is actually moving so fast. So, policy fails when it waits for evidence, and it fails when it acts without it. So, that I think is the report's central tension. It's, you know, the scientists call it the evidence dilemma. Policymakers need evidence to to make good decisions, but by the time the evidence is conclusive, the window to act may already be closed because the evidence lags behind the pace of development. It's moving so fast, light speed, right? The report, we can't resolve this, but it names, the report names it as the defining condition of AI governance. So that means that we need to have governments that actually have courage because that's a defining thing. The implication it does draw, though, is you govern under uncertainty, which is normal. But AI is distinct because mistakes may not be reversible. They aren't always reversible. We've certainly lived through the harms. I think the second thing is that better policy targets systems, not outputs. And so much of our, our debates have focused on You know, very small things like in journalism, you asked this earlier, it was about content moderation and everything got focused on content moderation while the system remained corrupted, right? So the strongest directional claim in our report is in the human rights and democracy section. And we have some incredible scientist women who have done this. By the way, the panel is almost 50/50 men and women, right? And also so Global South represented. The most urgent governance shift is from content moderation to system architecture. That's the main part. You regulate the persuasion and manipulation machinery itself. You don't police individual pieces of content. The report is actually very firm here because governance that only chases outputs will always fall behind systems that are engineered to generate content that's That's the stuff I know really well. But there's a lot more in the report. If you look at it, it will give you the facts to actually jump to your own conclusions.

Thank you. We'll go online to Suraksha from the Economic Times of India. Go ahead.

Hi. Am I audible?

Yes, go ahead.

My question is to Mr. Yoshua Bengio. The report says that countries without their own infrastructure or AI testing capacities risk losing opportunities to co-develop key technologies, shape governance frameworks, influence emerging global standards, and retain talent. So he was in India just 4 months back. I'm sure he gathered some insights. Sam Altman was there. Dario Amodi was there. Where does India fare in these parameters? Can the co-chairs elaborate, please?

Well, the report is global and is not like looking at a particular country, but I could comment on the pieces that you mentioned from the report. Again, we're not making particular recommendation, but we are seeing that there's an issue that— and India, you know, among the Global South countries, is kind of privileged because it's so large and it has a large pool of talent actually. But think about the vast number of other countries in the Global South. In fact, even in developed countries, the question is, can they even be part of the development of the most advanced AI systems in a way that they can shape a future that's aligned with their values and their interests. And so what the report talks about, for example, is the scale of investment and talent pool that has become required to build the most advanced models. And if those models are going to dominate economically and potentially in other ways, what happens to the countries that don't own these models? Well, individually, it's clear that they can't do anything. I mean, they can try, but there is no chance that they can really manage this challenge. So I'll let you jump to your own conclusions about what are the other options, but it seems very clear from the data that this is a real issue. And when Maria and I were in New York with the General Assembly on the 19th of June, clearly it was something very, very central in the minds of the member states who've been asking a lot of questions about this, about the AI divide or the concentration of power or the sovereignty issues for which it doesn't look like there's a trivial, simple solution that each country individually can, you know, implement. Thank you very much.

We'll go to Andrea from Reuters and then we'll go to Victor. Andrea.

Yeah, hey, thanks so much for doing this. I'm, I guess, digging in on these same very questions. So I've been a little surprised by the lack of concern being expressed by the IMF and the World Bank and others about these— the— about the gap, the AI gap that you're describing. And I'm— I wonder, um, if you think that, like, you know, there's just not enough attention to this economic fallout of what this means. We're already seeing that the Sustainable Development Goals are not going to be reached, um, but, but partly because of drawbacks, you know, put, you know, because of COVID Now, you know, with this grave imbalance in terms of power and concentration of power, I just wonder if, you know, what your thoughts are on that. But, and then also, I mean, I realize you're not making policy decisions, but your point, Joshua, is important, is that no individual country in the Global South is going to be able to address this. So it's going to require some kind of broader initiative, if I'm correct, right? Thanks.

I didn't say it, but yeah. I just want to make a comment about organizations like the World Bank and the IMF and economists in general. They have a very rigorous tradition of allowing enough time to pass for data to be collected and having strong evidence about a conclusion. So in the report, for example, we look at the studies that exist about labor, the effect of AI on jobs, right? And first, these studies tend to be old because it takes 1 or 2 years for any of these studies to happen. In the meantime, AI has continued its rise in capabilities.

We're—

there are also some studies which are more recent and we can see some effect, for example, on young people, it's more pronounced, but it's still a small effect which doesn't seem to, you know, make a big difference on the overall labor market. And it's happening in some countries and not others. So, yeah, the only way that governments can take that is deal with this evidence dilemma and consider the possibility that the early signs that we're seeing could be showing a real, in the case of labor, danger for the global economy, but at the same time, not necessarily bet all their choices that this is going to be the future. The institutions you named, they're very conservative. And that, you know, has given them a good reputation for taking, you know, making solid recommendations. But it might not be appropriate now with the speed of deployment and capability advance of AI that we're seeing. And by the way, the economists among themselves don't necessarily agree on these things. And when you ask them, you know, to try to figure out why they come to different conclusions, it's typically because they have a different mental picture of where AI will go in the coming years. So it boils down to the thing I was talking about at the beginning. What's going to be the capability of AI in 2 years from now, 5 years from now, even 6 months from now? Nobody knows, but the important scientific point is there are plausible scenarios, and we should all be prepared for, you know, those scenarios. Joshua, do you have—

like, is there someone that you, or some institution that you think is doing a good job from the economic standpoint?

I'm not going to select organizations and people, but yes, there are.

Maria, go ahead.

I was just going to echo this. I don't think that economists are silent. Our report actually, quote, you know, draws on the work of the IMF, the ILO and the World Bank. They are studying it. But I think what's true is something that's a little bit more uncomfortable than silence, right? The economics can actually, as Yasha pointed out, can go in many millions directions, the forecasts of AI's effect, they span an order of magnitude. I feel like we're living in a science fiction world, but we do flag macroeconomics and labor markets as areas where the science can't draw confident conclusions. Leslie Cho, who is in my part of the world, he's in Singapore. He's an economist, right? He's actually gone through a lot of the things with us. I learned a lot in that time period. I don't think it's the institutional failing to sound an alarm. It's honest uncertainty about this fast-moving system. I think what the report says clearly, though, is that the gains may not happen. It's not automatic. And the distribution— who captures the surplus? That's the open question. And that's the alarm of the report. And it's an evidence-based one. I think the other thing that goes with that is that this happened in my own field. The problem often isn't that no one sounds the alarm. Again, I told you I feel like Cassandra and Sisyphus combined. In the last decade, we've been sounding the alarm in terms of our public information ecosystem. Part of the other reason is that in a polluted information environment, these alarms don't travel the way facts should. So this is where we can— we need your help, right? Careful, hedged, evidence-based warnings are quiet by nature, and they're quiet— that quiet loses to loud in the system we now live in. That's part of why this report exists, to put that shared evidence in front of every government at once. So hopefully— and in front of you, so the signal doesn't get drowned. Okay. Thank you.

We have time for one more question. We'll go to Victor Barrero from Rappler. I'm not sure if it's ethical for him to ask a boss, his boss, a question, but we'll let you ask a question, Victor. Go ahead.

All right, thank you. Um, so the report mentions gaps in the evidence base that would allow for the panel to make confident scientific conclusions in areas like macroeconomics, labor market effects, and the effects on AI on the environment. What would be needed, researcher, for us as process-wise to help increase the evidence base and maybe steer policymakers towards making better AI policies where they are?

Oh, you're asking for a recommendation, but I can tell you the report very clearly highlights that if there were more investments in the question of the impact of AI in society, the safety considerations, the security considerations happening outside of the leading corporations building the AIs because you need independent studies and often you need different kinds of scientists, social scientists in particular. If you think about how it is affecting our information ecosystem as Mira has been talking about, but also psychologists, you know, we've seen the effects on children, on emotional attachment, of emotional attachment and all sorts of things like this. So it is something that requires a lot more societal attention, but not just in the media, also investing in the academic research and other independent organizations that can help us quickly figure out what is going on and provide the evidence that governments need. So, yeah, you can draw your conclusions, but right now most of the money in AI is in private hands and the interest, what is necessary, what is the information that's needed society needs to take better decisions, you know, that may be a different goal and it's something that public organizations like universities have a long tradition of investigating.

Great question. Okay, sorry, of course he's a reporter, but I would say 3 things. The gaps aren't mostly are really about what we pointed out, what the gaps are. They're about access and standards. And 3 things that would move the needle on that. The first is independent measurement access. Right now, the people who can see— Yoshua said this— right now, the people who can see the systems behave in the real world are the developers, and everyone else gets what they choose to share, right? Give official statisticians and independent evaluators privacy-preserving access, and you can finally measure the economic and labor effects instead of forecasting across an order of magnitude. I think second is your standardized reporting on the environment. We can't even compare energy and water footprints because there's no common standard across the lifecycle. And then the third is capacity, the ability to do this measurement. And again, we go back to that, right? It has to exist. Exist outside the handful of countries where AI is built, or the evidence base stays as concentrated as the technology. I think, I hope we're closing gaps through these. We're coming up with thematic briefs and engagement with a wider scientific community. That's what we're going to be doing in Geneva. This report, a month and a half, right? That's what we— it's a baseline. It isn't the last word.

Maria, Yoshua, thank you both Thank you very much for taking the time for eye-opening briefing. We wish you safe travels to Geneva, a productive meeting in Geneva, and we hope to welcome you back here whenever you want. Thank you both and take care.

Thank you.

Um, for those of you who still want to report on a lot of events today, we'll start the briefing at 12:20. To give you a mental health break of 20 minutes. And then you have the President's Security Council at 1.