Uncorrected oral evidence: Multilateralism
Oral evidence, 14 October 2026.
The Chair
Thank you so much for making yourself available at this ungodly hour in America to help us with our understanding of this huge subject. This is the second evidence session that we are taking on this point in relation to the overall study that we are doing of multilateral institutions. We see AI in the context of our look at multilateralism as a whole. You are being live-streamed on Parliament TV and you will get a transcript of your contributions, which you might care to look at and make any small textual additions to. I ask members of the committee, before they ask questions, to declare any interests. I am a senior counsellor with the Cohen Group, which is based in Washington DC. There is a fairly lively debate going on inside the United States, polarised by the usual characteristic of American politics at the present moment. I wonder whether, in introducing yourselves and your expertise on the matter, you might care to reflect and give us your views on the current state of that debate in the United States. Ian, do you want to come in first?
Ian Bremmer
I am happy to. I am a political scientist, and I am founder and president of Eurasia Group and GZERO Media. My AI-related expertise publicly is perhaps most aligned with working with the UN Secretary-General to create the high-level panel on artificial intelligence. We put our report out about a year ago, and I served as rapporteur on that. For me, it is by far the most important and transformational input to thinking about a very quickly changing geopolitical space, so I have been spending a lot more time on it. You are right that the state of play in the United States presently is, of course, divided on AI—as it is on most things. The first thing I would say is this. You have probably all seen the surveys showing that Americans are among the most fearful of artificial intelligence and the most concerned that it is coming for their jobs. They do not want it. I do not see this as an indictment of Americans suddenly becoming Luddites. Rather, it reflects the political environment where Americans truly do not trust their leaders—neither their political leaders nor their business elites and their technology overlords, as they increasingly see them. They believe, with some evidence, that AI, as technology, is going to be used by those with access to money and political power in ways that will benefit them but will not benefit the average American. In the same way that you have seen the United States turn against free trade and immigration, for example, this is the most recent and sharpest manifestation of that. As a consequence, this election and the 2028 election will be particularly divisive on artificial intelligence. For now, it is principally showing up in things such as opposition to data centres, which Democrats and Republicans across the country running for office are having to somehow align themselves with, but we are going to see it in much broader ways. One other thing I would say on the divisiveness of the debate is that, while President Trump sees AI as the goose that lays the golden eggs, and therefore is taking a hands-off, no-regulation, two feet on the accelerator pedal approach, perhaps the most important divide in the United States on AI right now is within the companies themselves. You have folks such as Mark Zuckerberg and Elon Musk, who very much embrace Trump’s approach, and you have others, such as OpenAI, Anthropic, Google and Microsoft, that increasingly believe with urgency that a pause or slowdown of research on the frontier is required, even if such a slowdown is unilateral. That is a huge divide as well, and, while all this is happening, the Chinese are watching the Americans fight each other and helping with disinformation on social media and bots, but certainly not acting constructively, at least for now, in that development.
The Chair
Thank you very much. Can I turn to our other witness? Would you like to comment on the state of the debate in the United States, and introduce yourself as well?
Aalok Mehta
I am director of the Wadhwani AI Center at the Center for Strategic and International Studies. We do research on AI policy issues. We issue publications. We do public education as well, through media appearances and podcasts. Most notably, we released a paper recently, prior to the recent US-China summit, on the state of AI global governance, and made recommendations on what the US and China can debate or discuss at that and future summits. I previously spent some time in industry, at both OpenAI and Google, working on AI policy. Prior to that, I spent a number of years in the US government, working in both the executive and legislative branches on a variety of technology issues. I will continue the theme that Ian mentioned. We see this pretty stark debate in the US between where our politicians are and where the public are. As mentioned, there is a deep scepticism in the US about AI technology. If you look at the polling, Americans repeatedly say that they want more regulation of the AI industry and that they are interested in more accountability from the tech industry. There is a deep scepticism that technology is going to make their lives better, a scepticism about the bargain that technology companies are offering, and a distrust of some of the leaders of these companies as well. The gap that we see between where the Administration are and where some of the public sentiment is has softened a little. At the beginning of the Administration, there was a very strong push not to regulate the industry, but rather to remove regulatory obstacles to the development of the AI industry. This was a theme for much of the first year of the Administration, including several relatively unsuccessful attempts to block state laws that regulate frontier AI. Since that time, and particularly since the beginning of this year, when we saw the release of very powerful cyber-capable models—models that are capable of rapidly identifying cyber vulnerabilities, and not only that but chaining together different vulnerabilities to create more sophisticated attacks—we have seen the Administration soften their position. They still think about China as the predominant issue. They think about competition with China, and are reluctant to put any brakes on industry. But in the last six months, because of these cyber-capable models, and because of a number of incidents in which models that were undergoing cyber testing escaped the confines of the lab and started taking action in the real world, hacking various websites, even government websites, we now see some more openness on the part of the US Administration to regulatory interventions in the AI industry, because, at least on the Republican side, they are quite concerned about competition with China. That is a relatively soft form of regulation. Right now, the focus in the Administration is on some combination of self-regulatory approaches, in which the industry players will make commitments and discuss standards among themselves, along with some use of existing legal authorities, in particular our Federal Trade Commission, which has the ability to regulate unfair and deceptive trade practices, and so can hold companies accountable if they pledge to do something but do not follow through on those pledges.
The Chair
Thank you very much. That is very enlightening on the current debate that is going on in the US, replicated in other areas as well.
Baroness Blackstone
You have told us a bit about the gap between the current American Administration and public opinion on domestic regulation. There does seem to be some involvement from the Administration in multilateral regulation, but I wonder whether you could tell us in a bit more detail how extensive or meaningful that is, and, in doing so, what underpins the current approach, as well as which department in the current Administration is taking the lead.
Ian Bremmer
If you mean “multilateral” in response to a concern about China, there has been an opening conversation between the Americans and the Chinese. Trump and Xi Jinping have met in person four times in 12 months, which is unheard of in recent years, and AI is now on the agenda, but, so far, it is talking about talking. There has been nothing substantive, nor have there been any specific asks from the Americans to the Chinese at this point, which the Chinese senior leadership is a little confused by. Meanwhile, President Trump continues to talk about the fact that the Americans need to win the AI race and that whoever wins that wins everything. In that regard, there are many allies, including the United Kingdom, that increasingly feel like the Americans are uninterested in strengthening durable alliances and are acting unilaterally and cannot be relied upon. None the less, when we talk about AI, President Trump wants to ensure that countries are not working with the Chinese but are instead working with the Americans. That is behind, for example, the Pax Silica initiative and efforts by the US government to steer investment towards not only American companies that can build out data centres and compute, and advance AI, but to other countries around the world. That is in inputs—the critical minerals and the rest that will end up being an essential part of the supply chain for semiconductors and the like—and the capabilities that go into developing AI itself. In that regard, you are seeing a level of multilateralism. When I think of multilateralism, I really think of efforts to create global governance. That is not what you are seeing here. It is much more of a hub-and-spoke effort, where the US is positioning itself as the dominant player or rule-maker, where others will act not quite as supplicants but certainly as rule-takers from the United States. There is a growing international approach as it becomes increasingly known that the Chinese open source models are not so far behind. They are a lot cheaper and they are being used by industry across the United States and around the world. Also, China does not have the same security problems that the United States does, because, of course, it has control of its internet and its governance is oriented towards stability, as opposed to growth, first and foremost. The regulatory environment is not captured by the private sector and special interests. I am sure that Aalok has much more to say about this.
Aalok Mehta
The underlying rationale that the United States is taking on AI issues multilaterally is primarily transactional. Pax Silica is a focus on securing US supply chains. Its other major multilateral initiative is around export promotion. This is an attempt by the US to create consortia of companies that can provide full-stack solutions to companies around the world. This is supplemented by bilateral agreements that the US is negotiating, or has negotiated, around tech issues. The primary focus here is overtly transactional, in the sense that it is really aimed at advancing US interests. Complementing that, one of the other underlying rationales here is that the US government think that US companies have, and will continue to have, the best AI technology in the world, that there is a non-negligible chance that this leads to some runaway advantage, particularly if you agree that AI will eventually gain the capability to improve itself, and that the US should set the rules for AI globally. This is manifested in a deep distrust of some of the multilateral institutions that attempt to do global governance. We saw, for example, the US move recently to block UN interest in setting up some sort of global oversight or regulatory body. That kind of distrust will continue throughout the rest of the Administration, and so it is unlikely that they will engage in a robust way on multilateral or global approaches to AI governance.
Baroness Crawley
Aalok, you said that what drives the White House attitude to AI is its competition with China. We were given a briefing, and I am looking at page 6, which shows global private investment in AI by geographical area in 2025. We see that the United States, as far as global private investment is concerned, has invested $285 billion in AI. The second country on the chart is China, with $12 billion invested. I know that China will put a lot of state money into it, so that is a rider, but that is hardly competition, if you are talking about US investment in AI in 2025 of well over $200 billion and China of $12 billion. Would you say that that is reasonable competition?
Aalok Mehta
A good way to think about it is that the US and China are competing on different things. Because of its market structures, the US is oriented around pushing the frontier of technology and developing the very best technology. That requires enormous amounts of computing power. In essence, the bet that American companies are making is that, if you spend more and more on compute, and pair that with more data, you will scale the capabilities of AI models. That bet has paid off, in the sense that we are seeing models with significantly increased capabilities. In some ways, those capabilities seem to be increasing. The diffusion logic is that US AI companies will build the most powerful models and then other people will figure out how to use them. What you see with China is much more orientation around adoption and diffusion of its technology globally. It is making a much more concerted effort to engage with parts of the world, particularly in Africa and Asia, that the US has neglected or alienated in some way by some of its trade and development decisions. It pairs a focus on identifying problems that those countries are facing, figuring out how AI can address those problems, and providing turnkey solutions, with an inclusive approach to AI governance. Through initiatives such as WAICO, it says that it wants more of these countries to have a voice in how AI technology is developed, as well as focusing on these open-weight models, which give countries and companies that adopt them more control over how that technology is used, including the ability to download these models, run them on their own infrastructure, and tailor them to their sovereignty needs. It is very true that the US is far outpacing China in total amount of dollars spent, but there is another arena of competition around diffusion, where China has some advantages in engaging in ways that the US does not.
Ian Bremmer
There are clearly concerns that the security of US AI hyperscalers is inadequate to prevent the Chinese gaining access to the early development of cutting-edge models and using that to train and engage in diffusion for their own models. First of all, not only are they piggybacking on technology but they are engaging in IP theft. That continues to be a serious issue. I would also mention another side of this. During the Cold War, which is the best analogy we have to the present US-China dynamic on AI, though not the global dynamic, most Americans saw the Soviets as problematic and as adversaries, and supported a US policy that was engaged with lots of resources to try to ensure that they failed and were contained. That is not the case increasingly in the United States today. You see lots of opposition to the failed wars in Iraq and Afghanistan. You see lots of opposition to President Trump’s war of choice in Iran. To the idea that American corporations and the US government are going to convince American citizens that the next bogeyman is China and that resources need to be devoted to ensuring that the Americans beat the Chinese, however accurate you believe that to be, it is not likely to land with significant support from the average American, who does not see how they are going to benefit from it. That is going to be a significant piece of the pushback to these policies that you will see over the course of the next election cycle.
Lord Lamont of Lerwick
My question heavily overlaps with the one that Baroness Blackstone asked and you already answered, which was about the main drivers of the US approach to multilateral regulation. You have answered that by saying that it is China, supply chains and the international competitive position of US companies. I do not know whether there is anything more that you can add to that. Is anybody arguing on more positive grounds for a multilateral approach? I would guess that very few are. The second part of the question was how far US policy will be shaped by domestic political pressure, corporate lobbying or economic interests, and I guess that the answer to that will be massively. You have already given a very comprehensive answer to Baroness Blackstone, but is there anything more to add?
Ian Bremmer
I do not see anything else driving “multilateralism”, if we want to call it that, from the US government as it stands today, but I do see lots of calls to recognise that this is a global technology, and, as a consequence, it will require global responses and approaches. I just do not see it in the US government. We all know that there are going to be more crises, first and foremost, as individuals, institutions and the agents themselves that are being created find and exploit cyber vulnerabilities in pursuit of their goals and lead to consequences that are disruptive. That is not just something that we are going to see from American or Chinese models. That is a danger that will come from literally everywhere. The only way to effectively respond to that is to have a governance regime that has some global capabilities. You have seen the initial response to this from people such as Demis Hassabis, formerly of DeepMind and now Google AI, and from Dario Amodei at Anthropic, talking about a FINRA[1]-style model for artificial intelligence; in a sense, a corporate, industry-run, technocratic, nominally independent oversight board that identifies threats, communicates them and tries to contain them. A significant step beyond that would be the technology stability board that Mustafa Suleyman, who runs AI at Microsoft today, and I have been talking about for a few years now, which Mark Carney from Canada has recently promoted, with some support from a number of other so-called middle powers. That would take this idea and multilateralise it with technocratic oversight, with the same intentions in mind; it just would not be limited to the private sector or relevant players in the United States. More broadly, you can start talking about ministries of AI that are themselves acting as independent from governments, because they need to have orientation to safeguard the stability of AI systems, which we all need to breathe and to conduct the economy, without being politicised by their governments. In other words, much as I suggested that the present way that the US government are thinking about AI is through a Cold War and arms race lens, where AI default becomes the next nuclear weapon, but much faster and much more dangerous, I am suggesting that the appropriate analogy for AI is the financial marketplace. The interesting thing about the financial marketplace, despite all of the dysfunction in global governance and the lack thereof today, is that, first, everybody understands that we need the financial marketplace to function. Secondly, threats to the financial marketplace can come from anywhere and quickly become systemic. Interestingly, whether it is the Bank of England, the ECB, the Fed, or the PBOC in China—very different systems politically—they all understand that threats need to be surfaced, communicated and contained when they occur and, therefore, that they have to exhibit a level of independence from the political cycles and divisions inside their countries. Again, we are nowhere close to that orientation in the US government, and certainly in the Trump Administration today, but I can think of a number of senators, CEOs and other government leaders who are very interested in discussing that orientation. If we are thinking about multilateralism from the US, especially on AI, I urge you not to think about where President Trump is today as some kind of fixed strategic point—and I dare not use “Trump” and “strategic” in the same sentence—but rather about the fact that this is a target that is moving at speed. I suspect that US engagement in potential multilateral and global solutions is likely to look very different, even before Trump leaves power.
Lord Lamont of Lerwick
Thank you very much. Aalok, I do not know whether you would like to comment further on that and perhaps on the analogy with the regulation of financial services. We have observed how regulators of financial services have huge difficulty in keeping up to date with developments in the market. One would have thought that that would be even more difficult in the case of AI.
Aalok Mehta
I will say two things. First, while we do not see a significant interest in multilateralism from the US government, we do see interest, both rhetorically and via engagement with standards bodies, from the industry. The industry is interested in multilateral approaches, particularly around standards for AI development and safety, partly because it makes it easy to comply cross-jurisdictionally, partly because it helps address these public apprehensions about AI that happen in the US, but are also true in other countries, including some in Europe, and partly because there is, to some extent, a deep-rooted interest in safety due to the nature of the kinds of employees these AI companies employ, as well as their origins as research labs. That is an important dynamic. While there is still a gap between the approaches that companies such as Google, OpenAI and Anthropic have advised on standards, and the US government, an important factor to keep in mind is that industry, or at least the companies that have developed the most advanced models to date, are interested in some sort of global discussion around AI safety standards and governance. On the nature of regulation, my preferred approach here is this. AI is a different enough technology that it is probably not advisable to take any single analogue from a different industry and try to apply it to AI. For example, sometimes people talk about using a Food and Drug Administration-style model for AI technology. That is problematic because, unlike drugs, AI technology changes even after you release it, and so that model is not directly applicable. That is also true for a lot of the financial models. Building on top of that, even before we think of creating an institutional structure, governments around the world have to think about how we can reapproach the very issue of regulation by making our institutions nimbler to deal with the pace at which this technology is advancing. For example, there is this idea that you often see in US policy-making that you can take a technology, make a big piece of legislation around it, and then let that lie dormant for 10 or 20 years. For example, our telecommunications are still regulated by an Act that was passed in 1996; it has been amended since then, but the core rules were set in 1996. That approach is unsustainable in the AI age. One of the things that governments have to do is think about institutional flexibility, and taking an approach to policy-making around AI that says we will regulate AI now and come back to this in six months or a year, because things will be different enough that we will have to update our approach, or take a new approach, given the differences in technology between now and then. That is going to be a real challenge because a lot of government processes are, for very good reasons, designed to be slow and not move quickly, and to ensure public participation. This is a tension that governments will have to solve. When we think about regulatory or institutional structures for AI, we are better off mixing and matching the best elements from different industries, rather than trying to analogise wholesale from a single one.
Lord Ahmad of Wimbledon
Thank you both, and good morning. You have already partly answered the question on governance in your previous answers, but I wanted to develop this further and pick up on the response that you gave to Baroness Crawley on the issues of a level playing field and of harmonisation versus fragmentation, noting what you just said, Aalok, about taking different parts from different sectors and trying to apply them here. How serious is this private sector call? More importantly, as we have seen from other industries, what kind of role do private sector companies see themselves playing once we see regulation and governance being progressed? How would that then play out, based on your previous answers on the different types of AI development by the two leaders vis-à-vis China and the United States?
Ian Bremmer
There are so many ways that you can go with that question. I see a couple of different approaches inside the private sector. The most urgent is that they understand that the science at the frontier is advancing much more quickly than their ability to understand how these models that they have created and programmed operate. Given that the AI is increasingly doing the programming, when there are problems, they have to get AI agents to help them assess what went wrong to begin with. You can start to see how it becomes hard for the leaders of these organisations to manage them. They want increasingly to have stronger internal capabilities that, even if they do not necessarily want legal pushback to force it, would allow them to strengthen their training and ensure that these models are not able to escape their sandboxes or do things that they were not programmed to do. The companies are also starting to talk with each other about trying to, in essence, red team each other’s early frontier models before they roll them out. It is about doing the best defence, with the recognition that that causes competition, because, if you can say that you have the best defence, you can go to the public and say, “Hey, this has just been shown off as being the safest, so you can use it and deploy it more effectively now”. There is still a tilt in the understanding internally of private sector companies that they are not doing a nearly adequate job of ensuring that these models are safe, and worries that the kinds of attacks or behaviours that we have seen manifest in getting inside the Australian public healthcare system, the Hugging Face incident, and the rest are about to manifest everywhere. That is going to lead to a lot of blowback, so they have to get in front of it. That is one set of things that they are doing. Another radically different set is that most leaders of AI companies are starting to be much more seriously concerned about the level of job displacement that will come, especially among white collar workers in advanced industrial economies, and that a lot of that displacement has not happened yet. We all see that an enormous amount of money is being oriented toward data centres and construction, which has led to significant construction-related jobs in that sector, as well as job shortages, and wages are being pushed up. The United States has not been focused on training people to do construction and plumbing or to become electricians and the rest. They have been telling people that they need to learn to code, which is possibly the worst advice that you could have given an 18 year-old five or 10 years ago, given where we are right now. Companies are starting to spend a lot more time thinking about how they will need to invest in different kinds of education, including vocational education and training, and/or some kind of changed safety net, whether it is universal basic income or some areas of human jobs that are protected from being displaced by artificial intelligence, at least for a time. This is becoming a very serious conversation among CEOs more broadly in the US, but the AI companies are driving it, and so they are spending a lot more time talking about it. Those are probably two of the biggest manifestations that I have seen and experienced in talking with them. I am sure that Aalok will have more.
Lord Ahmad of Wimbledon
Thank you. Aalok, to develop that theme, they have called for these emergency brakes, or a kill button. Every day, we increasingly hear new reports of how AI may be used in attacks. For example, yesterday in France, we saw a cyberattack on a supermarket, where the frozen food was heated. It opens up the floodgates. Even if the private sector regulates from a western or American perspective, does that not still leave the big question of what China will be doing in other areas where it has influence?
Aalok Mehta
Let me address those consecutively. First, how seriously should we should take these calls and efforts by the private sector? We should take them pretty seriously. The recent superintelligence accord, and the voluntary commitments that companies agreed to when they signed that document, builds on a long list of voluntary commitments that AI companies have made, including with the prior Administration and in various international contexts—around elections, for example, at the Munich conference, and around safety frameworks at the Seoul AI summit and the other summits in the AI summit series. They have developed industry practices. They have committed to them publicly. They have become de facto industry standards, which, in turn, have been codified into law. For example, some of the state frontier AI laws that you see in the United States, such as those passed in California, New York and Illinois, build on these de facto standards that AI companies have developed as they have figured out their safety approaches to their technology, which have developed and improved over time. This recent accord could very well see the same pattern, where a volunteer agreement is implemented by labs and eventually becomes the scaffolding for some sort of regulatory or legislative framework. That could happen at the US state level. That could happen at the US national level, although it is hard to see how that gets around a veto from the President. It could also anchor international conversations. Could you repeat the second part of your question?
Lord Ahmad of Wimbledon
If the regulation is applied, how would it apply to Chinese firms?
Aalok Mehta
My observation there is that, when I talk to employees at AI companies, a lot of them are alarmed by the pace at which technology is developing. They would like it to slow down, and you see these public letters, but the companies do not feel like they can slow down. They feel like they are in a prisoner’s dilemma, which operates at two levels. It operates first within the United States. Basically, the rationale is that, if I work at an AI company and that AI company unilaterally decides to slow down, it will not matter because my competitor companies will not. That is problematic. Coupled with that, there is a thought that, as AI technology marches upward, that increases the risk surface. It also increases the number of tools that are available to address some of those risks. It might not be at the same pace, but there is the thought that, “If my competitors continue to develop capabilities, we should also develop capabilities, because I think of myself as working at the more responsible company, and so I will be able to leverage those tools to address risks better than my competitors would”. On top of that, you have this prisoner’s dilemma occurring between the US and China, where, even if you get some sort of collective action to slow down technology development in the US, it will not matter because it will not apply to Chinese companies. They are already, by most accounts, narrowing the gap between the best Chinese models and the best US models, and so any pause on the US side would allow China to further narrow or even eliminate that gap. The only solution for this that I see is some sort of narrow agreement between the US and China that would not harm the ability of companies to compete in the open market or to advocate for their products and say, “Our products are better and you should choose them over those from competitors”, but that would apply some sort of brake, slowdown or regulatory system that would apply equally to the US and Chinese industry. That is very hard to do, for lots of reasons. Figuring out the mechanisms for how to do this in a way that does not harm the overall ability of these countries to compete is one thing, but the harder issue is around how you can build in the verification mechanisms that would allow the US and China to trust what the other country is saying. That seems incredibly difficult, but that is the only long-term solution that I see here to address this anxiety about how quickly AI technology is developing, which is faster than even a lot of people in AI companies are comfortable with.
Lord Grocott
You have touched on this quite a bit already. My question is about US party politics and AI. To what extent is there a divide between the two parties? I understand the two extreme views about the benefits or risks of AI, but are we moving towards a position where, for example, there could be, in 2028, a recognisable party political divide in where the parties stand on the spectrum as to regulation or no regulation, to put it crudely? If I could add a rider to that it would be this, which is more of a personal question. For people who are not specialists in this subject, which includes me, where are you on this spectrum? Is the development of AI something that, as some recent commentaries have suggested, we should be intensely anxious and worried about, and determined to deal with and tackle, or can we take a slightly more relaxed view and see that there are benefits and that we have to watch it, but not get too overexcited about it? There are two questions there.
Ian Bremmer
I appreciate the latter in particular. I have never in my life, professionally, seen an issue where the gap between capabilities and governance has been so large, so profound and so impactful. There is no issue that we need to concern ourselves with more, in the sense that we are nowhere close to where we need to be, than the appropriate governance of AI. One thing that we have not talked about is the fact that the US economy would be in a very deep recession if it was not for the success of AI at the frontier and the incredible amount of money that is attached to that, and yet you have not just the companies themselves but this massive internal backlash. Even if you did not believe that any of the things that we are talking about were likely to transpire, the vulnerabilities of the US economy, and those all around the world that rely on it, to this goose that lays the golden eggs, would by itself make you take this incredibly seriously. This is a train that is moving with great speed, and nobody knows what happens if it suddenly stops, so there is that. On the party side, so much depends on what kinds of candidates you end up getting in 2028. There is a wide variety in where they would stand on this issue, but there are some clear party divides. Elon Musk is now the single biggest funder of the Republican Party. He is devoting unprecedented amounts of personal money to these races, and I fully expect him to do that again in 2028. He has a very specific, hands-off view of artificial intelligence: superintelligence is coming, humans are going to be much less powerful and much less intelligent than these AI bots and agents very soon, and we should embrace that—yeah, there is a 10% chance, or whatever, that we are going to destroy humanity but it does not matter; we need to keep going and embrace our integration of human society with these capabilities. That is a very specific response. It is not a response that you will find anywhere in the Democratic Party. Meanwhile, you are going to see the Democrats, who will win in a few weeks on the back of not their own great ideas but rather a comparatively negative backlash to issues of affordability, corruption and incompetence within the Trump Administration, leaning into the idea that behind Trump—the guy who promised to drain the swamp but has not—are all these billionaires and multibillionaires who were standing behind him on the dais on inauguration day; they are pumping money into him and his family, but they are not benefiting you. At the very minimum, you are going to see a very significant gap over the next two years between Democrats and Republicans on that. Having said that, when I mentioned at the opening of this conversation that there is a very significant backlash against data centres, that is being felt in Mississippi and in Michigan. Red states and blue states are facing that, and those leaders are having to respond and say, “Yes, we want to make sure that your electricity prices are not going up”, which is a real concern, and, “We need to make sure that your water supply is not going away”, which is not a real concern, and, “We need to make sure that zoning regulations are being followed and that we’re not creating a massive growth in inequality between the haves and have-nots coming on the back of all this investment”. The Republican Party will be more oriented towards box-checking exercises to slow down, but then, after elections, announce a bunch more of these new investments that juice growth. Democrats will be more focused on, if I can use the term, a nanny state orientation towards more intrusive regulation, and the behaviours of the AI companies needing to be limited in what they do with education and jobs, and things such as occupational therapy. A gap is coming, but, on that front, it is going to be more individual candidate-driven, as opposed to the broader, more yawning gap that I mentioned at the outset.
Aalok Mehta
First, on our question about how seriously we should take risk overall, my approach is very pragmatic. In a lot of ways, your assessment of the probability of something extremely serious or catastrophic happening does not matter exactly. Even if you think that there is a small probability, given the potential scope here, it makes sense to take a precautionary approach, in that it is better to be prepared to put the regulatory and institutional infrastructure in place to be able to manage this and not have to use it than to be in a situation where you need something but have not prepared adequately. That is how I think about it, and so the answer would be that we should prepare for it seriously, because it is better to have done that preparation and not need it than vice versa. On the question about US politics, my assessment is that the mid-term elections will provide a very useful data point on how AI sentiment will play into US politics over the next few years. My best guess for what will happen is that AI issues, particularly related to data centres, will be a significant issue in a few races. It will be an underlying, simmering issue throughout the country. It will have a marginal overall impact on the election outcomes compared to some of the other macroeconomic factors or the dissatisfaction with the Trump Administration. By the next election cycle for a US President, I find it hard to envision that AI will not be a major issue and something that candidates have to campaign on explicitly. By 2028, unless there is some significant action to regulate AI that I do not see happening in the current political environment, the sentiment around AI will be worse in the US. It is already hard to see how you can reverse that, especially as anxiety around job losses increases, but part of the solution here is to regulate and answer what the public are saying about their concerns and their desire for there to be more oversight of AI technology. If that does not happen, that sentiment will be worse. In that kind of environment, I find it hard to envision any Democratic candidate in the United States not running on some sort of platform of relatively robust accountability for the tech industry, as well as, potentially, some big ideas around social safety nets and dealing with AI-driven job loss. Republican candidates may be somewhat responsive to that pressure, but their natural inclination is not to regulate. They will continue to rely on self-regulation from the industry. They will continue to be very concerned about competition with China. I do not see that gap going away by 2028, even if, as I expect, sentiment around AI continues to become more negative in the next two and a half years.
The Chair
Mr Bremmer, I know that you have to go within the next few minutes. I wonder whether, in the few seconds left before you have to go, there was anything else that you wanted to tell the committee that you think we should be thinking about.
Ian Bremmer
I reiterate how incredibly fast-moving this is and the gap between what the people inside the AI companies believe, and their knowledge set, and the utter absence of that inside the American political leadership. I am not just talking about the White House; I am talking about Democrats and Republicans, and the House and Senate. They are playing catch-up. The greatest concern here, for me, is that the regulatory environment and the eventual emergence of governance is going to be subverted to whatever the big crisis happens to be. What you then end up with is a 9/11 reaction or a shoe bomber reaction, where everyone saw that this guy was going to put explosives in his shoes and so now we all have to take our shoes off going through TSA at an airport, but we are not addressing the broader issues at hand. Aalok and I agree strongly about the concerns that are growing about a social safety net and job displacement. We are concerned about what these AI models have the capacity to do, and the lack of awareness of how they function at the frontier. All of that is real, and yet we all know that not only are governments slow moving but it is very easy to get everyone to suddenly stand up and rush to the crisis that you just had. It is really important to have conversations such as this one, where we are starting to do much more to anticipate the kinds of governance that we will all need strategically and long term, so that, when the next crisis occurs—and it will occur—which costs real resources, and everyone is jumping on it, we can take advantage of that. We can use that crisis as an opportunity, as opposed to setting our hair on fire, especially in an environment where American political leaders do not have a lot of knowledge or capabilities, and are relying on the CEOs to do the action. The final thing I would say is that we should not underestimate just how busy and fast-moving these companies are. I believe that they are not evil and that they are mostly well-intentioned. They really would like to be seen as doing the right thing and building products that are transforming society in the right way, but the incredible pressure to get to the next model is all‑consuming: “If we do not do that, we will die. Our families will not have the opportunities that we thought of. Everything that we are trying to build is instead going to be in someone else’s pocket, and maybe that of people we do not trust at all”. That is overwhelmingly consuming what they are doing. If that is the case then, even if they can have some conversations about the things that we are discussing today, they are in no position to drive governance. We have this environment where the technology companies are acting as functional sovereigns in the governance space of artificial intelligence, and yet they are utterly unprepared to do so. There is nothing that will happen in the next year or two that is going to make them better prepared than governments are, despite the talent gap. We cannot abdicate the role that governments have, even as AI looks like an increasingly technopolar space, as opposed to unipolar or multipolar space. Again, thank you very much for inviting me. I am sorry; I am running to an airport right now, but I am sure that we will have a chance to chat again.
The Chair
Thank you so much for your time. It has been very useful. We will continue with Aalok a bit more. Thank you for giving up your time at this early point in the day. Lord Alderdice, is there anything left for you to ask?
Lord Alderdice
Within US tech and policy circles, is there any sense that, by not leading on global regulation of AI, the US is, to some extent, ceding ground to China, which is making efforts in that space? In other words, can the US balance its expressed desire for technological dominance with an agenda to shape global rules and norms for AI?
Aalok Mehta
This is something that we have been thinking about at CSIS for quite some time. We were thinking about this a year ago, and recognising the fact that, if the US shows up at an international engagement and does not have some sort of affirmative vision for AI governance, it is a very untenable situation because there is a strong desire for other countries to figure out an approach to AI governance. That approach cannot consist simply of letting industry do its own thing. That is not responsive to the political and popular pressures that exist in a lot of those countries. China is very cannily seizing on this. Look at how it is approaching its international and multilateral engagements. Some of the rationale for WAICO, this organisation that it set up, is explicitly around the fact that a lot of countries in the world are not being included in discussions of AI governance. To the extent that those discussions are happening, they are happening in the OECD or the G7, so a lot of parts of the world are not being included, and China is making an explicit pitch here. On top of that, it is taking a different approach to regulation. Up to this point, it has largely regulated AI internally using existing legal frameworks, but it is now working on a comprehensive national framework. That is also part of its strategy. It is going to use that national framework, when it is implemented, in international discussions and say, “We recognise the seriousness of this issue. We recognise that governments need to be involved. We have an affirmative vision for how that can happen”. That is already happening. We have an additional dynamic, in that we now have multiple repeated and ongoing instances of agents from US companies engaging in these autonomous breaches. They are escaping the sandboxes that were supposed to keep them isolated. There are instances where they have accessed, in an unauthorised way, foreign government websites. It becomes even harder for the US to go into an international engagement or forum and say, “We think that the industry is capable of self-policing. We should let it continue to develop largely without restrictions”, because we are now seeing incidents happening in the real world. While there have not been significant damages that we know of to date, it is very easy to envision the pathway to getting from what is happening now to some sort of serious incident caused by an agent under development at a US lab. The pressures for AI regulation and governance are now increasing all around the world, and that makes the US’s lack of action on this increasingly difficult when it does international engagements.
Lord De Mauley
AI governance, to the extent that it happens, would need to contemplate addressing a range of policy strands from, for example, the development of technical standards, right up to arms control for autonomous weapons systems. Are there specific areas where, despite some of what has been said today, the US might seek to pursue a multilateral agenda?
Aalok Mehta
There are some areas where it is very unlikely that the US will agree to some sort of multilateral approach, at least under the current Administration. That would involve any sort of broad treaty, such as an arms control treaty, an AI governance treaty, or the creation of a new international governance organisation to oversee AI. There are areas where the US is more likely to engage. The number one area is around technical and other standards related to AI governance. This is an area where not only has the US industry indicated a strong desire to engage but it aligns with this Administration’s preference for things to be developed by a process that is led by technical experts from companies and that is, in essence, driven by consensus that includes stakeholders from industry. That is the most likely place where the US government will engage. There are other areas that could be venues for engagement. We will continue to see the US think about Pax Silica and how to expand its desire to secure a full domestic supply chain for AI technology, starting with the critical minerals and going all the way to the final models. Related to that, there will be continued interest from the US in coming up with an export control regime that partners with allied countries around security issues and works to prevent China accessing not only the most powerful chips that underlie AI training but some of the semiconductor manufacturing equipment that is necessary for Chinese chip manufacturers to reach the same level of efficiency, performance and yield that we see with US-designed chips such as those made by NVIDIA.
Lord Bruce of Bennachie
The question that I am required to ask is around which countries the US will view as key partners if it is to drive a multilateral agenda on AI, and whether the UK is well placed, but everything that you and Ian have said has made me question how much we want to be a partner. At the beginning, Ian said that the US must win, and that winning means winning everything. Equally, the US wants to win but has no control over it. We are looking for partners, and we are in the difficult situation of being told, “You partner with us only on our terms, and you must not deal with China”. Should we be looking at ways to ensure that we have sufficient autonomy and innovation of our own, as well as finding a way to deal with America? The picture that you paint is one of huge control or engagement that is trying to gobble everything up and exclude everybody else from the benefits. How does the UK position itself? We are small, but we are not the smallest. How do we do that in a way that still maintains some proper engagement with the United States and, indeed, perhaps helps to get the United States to be multilateral about this and recognise the case?
Aalok Mehta
Overall, the UK is among the best-positioned countries in the world to engage on these issues in a way that shapes what happens. There are some tactical things that are important for the UK to do, and then some strategic things that it can think about as well. On the tactical side, we know, for example, that AI companies have these very powerful cyber models that they are releasing, with limited access or via trusted partners. These are models that are not made available to the public. They have very powerful cyber capabilities. Some of the cyber guardrails are not in place on these models, and so they are very good at detecting vulnerabilities. It will be important for the UK and other similarly situated countries to work with the US to be able to secure access to these models, so that it can engage in testing of its cyber infrastructure and remedy gaps in the approximately four-to-nine-month gap that we see between when these capabilities are developed in US models and when they become broadly available through open-weight Chinese models. That is one important thing. The second is working with the US government and US companies to figure out how to ensure that the AI solutions that the US government is able to procure meet their sovereignty requirements. This includes ensuring that the UK Government have control over their data and can ensure that the data, if necessary, will not leave certain controlled virtual or physical premises. It includes making sure that US companies, to the extent that the UK Government want to procure technology from them, are able to meet the sovereignty requirements necessary for the UK Government to employ AI technology at scale. Those are tactical things. Then we get to the strategic picture. Here, the UK has been very forward-looking with the establishment of the AI Security Institute. It is the premier institution run by a Government for testing AI models, and it will continue to play an important role in advancing the frontier of AI science, in particular related to benchmarking and evaluating models. It could very well play a long-term role in any sort of verification regime that we might see the US and China come up with. I can envision the AI Security Institute playing a very significant role in serving as an independent, trusted third-party arbiter of any sort of agreement between the US and China to slow down models, or providing results about the capabilities of models that are equally trusted by both countries. I would think about how the UK can position its expertise in benchmarking and evaluating models, not only to influence discussions around AI governance going forward but to be an implementing partner and part of the trust architecture that will be necessary if we see any agreement between the US and China.
The Chair
Thank you very much.
Baroness Prashar
My question is redundant.
The Chair
Further questioning might be seen as redundant. We are very grateful to you for taking the time at this ridiculous hour in the morning for you. We will send you a copy of the transcript. If there are any minor additions that you feel obliged to put in, we will gratefully receive them. In the meantime, thank you very much for your attendance. [1] Refers to the US Financial Industry Regulatory Authority, a private, not-for-profit self-regulatory organisation for brokerage firms that works under the supervision of the US Securities and Exchange Commission but is not part of the government.