19
Examination of witnesses
Kayla Blomquist and Matt Sheehan.
Q29 The Chair: Welcome to this meeting of the International Relations and Defence Committee of the House of Lords. You are very welcome. Thank you very much for giving up your time. The session is live on Parliament TV. I believe there is an audience out there. You will get a transcript of today’s proceedings so that, if there are any minor changes you want to make, you can do that. Members of the committee will be required to declare any interests they might have before they ask a question. I am a senior counsellor with the Cohen Group in Washington DC.
I will ask the first question and members of the committee will come in on other aspects of this issue. We are looking at the multilateral aspects of artificial intelligence, specifically how that relates to China as a whole. As each of you introduces yourself, can you give us a general sense of the Chinese approach to the development of artificial intelligence, the technology around it and how the Chinese state, with all its characteristics that we know about, approaches this kind of novel technology in terms of philosophy, economic development and regulation, and whether you think that what Xi Jinping said about China being the architect of true multilateralism squares with its view of AI generally as well?
Kayla Blomquist: Thank you so much for having me. I am the executive director of the Oxford China Policy Lab and a researcher at the Oxford Internet Institute. I am pleased to be here today, and it is excellent that these specific questions about multilateralism are being asked in this context, because while AI, or the development of it, is highly concentrated in two countries, the impacts of it will be absolutely global, without question, and will require global co-ordinated responses now and into the future.
When it comes to how China has approached AI over time in its development as well as its vision for its role in the world, there has been a lot of thought going into this, stretching back deep into the 2010s at least, and traces of it definitely before that. Some of this has been promoted and thought of as being one of the potential drivers of the great rejuvenation of Chinese civilisation and part of a broader vision to restore its place as a leading global power. There has been thought across the development side as well as the regulatory side, with many regulations coming out year after year, but again the aspect of multilateralism is key. I think true multilateralism, pointing to the phrase it has used, is very astute because it will have its own definition of what true multilateralism is. Its approach to that will naturally position China as a provider of global public goods, especially when it comes to innovation and AI. We can see that in the announcement of the World AI Cooperation Organization, as well as other initiatives and multilateral innovation attempts it is making, such as those through the World Data Organization and others.
These are very early-stage efforts. They have been in the works and talked about for a long time, and they are progressing in form and substance, but the substance is in terms of the benefits to member countries—the actual deliverables and the returns on that “investment”—and what those will actually be for China. When any country positions itself as a provider of public goods for others, it can find itself in a bit of a bind in then being expected by other countries to provide things it may or may not be able to provide, whether that is on the technological front or otherwise, diplomatic, et cetera. It is a key thing to keep an eye on. It is still unfolding. I am really looking forward to hearing Matt’s comments as well and getting into conversation with you all on this.
Matt Sheehan: Thanks very much for having me. To give a bit of background, I am a senior fellow at the Carnegie Endowment for International Peace and co-director of the China AI Initiative there. I have been working on China AI issues since about 2017. Maybe the first four or five years were very focused on the development side, comparing Chinese with American capabilities in AI, how we can measure research quality, output and the quality of the data ecosystem and stuff like that. Beginning in about 2021, I shifted focus from the development side more towards the regulation and governance side. In 2021, China started to roll out its first meaningful regulations on AI. Its policy goes back much further, but binding regulations that impose requirements on companies started around then. I have spent the past five years monitoring all aspects of China’s AI ecosystem with a special focus on governance.
China’s approach to the development and adoption of the technology—in some ways this is a philosophical approach—is to put heavy emphasis on the application. How can we use this technology to achieve our broader economic, especially, but also social and political goals? The core of that is: how can we use AI to lever up our economy? How can we use it to climb global value chains? How can we use it to skip over the middle-income trap that people are afraid they will fall into? That tends to be how China looks at most new technologies, including information technologies.
It does not particularly like or is not particularly interested in pure software or purely financial markets. I forget what the exact translation would be, but they call it the fake or virtual economy—tuō xū xiàng shí. It has a different term for the real economy, shítǐ jīngjì, which is the manufacturing of stuff or physical services delivered in the world. It always puts heavy emphasis on the real economy. It does not want to be like the United States which abandoned that through financialisation. When it looks at a technology like AI, or even 10-plus years ago when it was looking at the mobile internet, it was thinking how to turn this into a tool for upgrading its economy and achieving these broader social, but especially economic, goals. That has guided a lot of its development. It has companies competing at the absolute frontier of AI development. That is a really important sector and where the US is primarily competing, but, if you were to give top leaders a truth serum and say, “What do you care the most about? Is it the absolute frontier or broad, diffuse application for the technology?”, they would lean towards the application side.
Briefly, on how they tend to do that, a lot of people think of the Chinese system as very top-down and a lot of its elements are very top-down. But when it comes to how to stimulate the adoption of a very diffuse technology, that is not something you can perfectly plan from the top. You cannot lay out a perfect road map for where, when and how AI will be used.
In both the 2017 national AI plan, which was the first big development document, and the 2025 AI-plus plan, which is the most recent big development document, the centre is essentially sending out the signal to officials all around the country, “We are prioritising AI development, especially AI applications, and you—Ms Mayor, Mr Governor, the head of the traffic bureau in some random city, or a university president—will be rewarded if you find some way to encourage the diffusion of AI in your area”. It is then up to those officials to figure out how to do that. If you are a traffic administrator, maybe you install AI in your camera networks; if you are a governor, maybe you create a tax-free AI development zone; maybe you launch pilot projects that integrate AI into public services or factories.
That is traditionally the way they tend to incentivise the diffusion of a very diffuse technology like AI. They have had a lot of success with it in the past. I think that this time around it will be a little more difficult, mostly because the local governments that are charged with actually doing this work do not have the same type of cash floating around that they had 10 years ago, due to the collapse of the property market, but philosophically there is a heavy focus on applications. In terms of the mechanism, I think of it as a call and response. Central government sends out a signal, “You will be rewarded if you do AI well in your area”, and then it is up to the localities to figure out how to do that. I will leave aside some questions on frontier AI development or regulation, if others are interested in those.
The Chair: So it is not particularly ideological.
Matt Sheehan: The first concern it has when a big information technology comes into play is whether it will disrupt its controls on information. That is a heavily ideological thing. When China started to regulate AI, its first regulations were on recommendation algorithms, deepfakes and generative AI, and the through line in all of those was: how do we make sure this does not disrupt our existing controls? We know how to control what the media says and what goes on the internet, but how do we control what an AI model says? How do we make sure it says the right thing? That in many ways is almost like the prerequisite to the next step, which is the go-go development phase. China goes through three phases: control, harness and govern. You control the political risks, harness the technology for your broader economic goals, and then start to govern these more diffuse social issues that arise, for example governing AI companions and data privacy. That tends to be the third stage. I feel it has the political risks under control, and is more largely in the harness phase as well as the governing phase right now.
Q30 Lord Darroch of Kew: This is probably for Mr Sheehan first but both of you may want to comment. On the development of AI intelligence, as an expert how would you say that China matches up to the United States? Where is it in terms of the competition between them? How would you characterise that competition? Is this like an arms race? Is it a cold war? How does it feel from where you come from?
Matt Sheehan: It is helpful to split this into competition at the frontier of AI—who is building the biggest, strongest, most capable models—and competition in terms of the diffusion and application of AI, which in many cases relies on having good frontier models, but not necessarily. A lot of stuff done in robotics and manufacturing does not need the most advanced AI. If we separate out those two things, it can be a little bit cleaner.
On the frontier models, nowadays I would say that the general consensus—I imagine Kayla talks to many of the people I talk to on this—is that China’s best model is about six months to a year behind the best US model. That means essentially that, if you take the model that Anthropic comes out with today, in six to nine months China will come out with a model that is about equal to it. I might push that a little further back to maybe nine to 12 months. That is my personal take, but that is the frontier. That is relatively easy to quantify because you can measure the capabilities and do head-to-head comparisons.
The diffusion front is much harder to quantify. The short version is that the US is very successful and quick at diffusing software-based AI, diffusing AI into business processes and into individual workplace optimisation, because we have a long history of paying for enterprise software in companies, individuals paying for enterprise software and having IT departments that integrate into the systems of major corporations. China, does not have a long history of buying software and paying for it, so that diffusion is slower and harder, and it is harder for the companies to make money. They partially do what we talked about earlier: they place their bets more on the diffusion into non-white-collar sectors. Maybe white collar is not quite the right division, but its AI-plus plan focuses on manufacturing, AI for science, healthcare and transportation. I forget the last two of them, but China is placing its bets on that type of diffusion, whereas the United States—I imagine the UK is in a somewhat similar boat on this—tends to place its bets more on directly integrating AI software into our existing software systems for productivity.
Kayla Blomquist: I very much second those points. On the question about the extent to which this is a race dynamic and new cold war, I characterise the US ecosystem-level dynamic as a race dynamic, in that there is a lot of focus very explicitly by the US Government but also implicitly on the Chinese side on the fact that this is a very important technology on which to be a leader in the world. On media ecosystems, in a lot of the incentive systems for industry, government and beyond in each country there is absolutely a competitive dynamic, but, as Matt said, it is important to disaggregate the layers on where there is a direct race.
Sometimes, especially at the frontier model level, I hear the same metric of six to nine months in terms of having the most advanced frontier model, recognising its relative shortcomings in competitive advantage when it comes to compute, as well as the broader ability continuously to develop the most advanced models ahead of the US. China has, therefore, leaned into open‑weighting, broadly called open-sourcing, very focused on part of the models being publicly available and downloadable. A lot of industry players have leaned into that approach over the years. It is evolving and changing, and not every single company or developer is continuing to open‑weight its most advanced models. We have seen movement on that both at firm level as well as country level with proposed draft updates to export controls related to AI and other industries.
As Matt said, when it comes to the diffusion goals and that strategy at the firm level, and then assumed at the country level, around an open ecosystem, diffusing the technology at a practical level goes back to China’s pitch for the world, often saying, “In contrast to leading US companies as well as broader US positioning, we are the open country in comparison”. It has been doing this for quite a while. We have data on public messaging on this going back to about 2015 or 2016. The comparative element and the messaging to the rest of the world is a really interesting component here.
One small additional point on diffusion is that, because of China’s strengths in terms of 5G technology and other digital infrastructures under the historical belt and road initiative being very active throughout the world in building out layers and infrastructures, we have seen some of these models, both proprietary and open-weight, being quite intentionally worked in at a very practical, applied level, whether it is telecommunications infrastructures or weather and flood monitoring services. Some very interesting and now long-standing, but not always talked about, attempts there would also speak to the public good provider positioning that China is weaving at international level.
Q31 Lord Ahmad of Wimbledon: Building on the responses you have given to the two previous questions and pushing that element of philosophy and China vis-à-vis the US, we have seen repeatedly for many years the economic dominance of China, as opposed to the US model, which ties in countries for a very long period. That means that the reliance on any renewal, any restructuring of a debt that may be owed, is tied in specifically and conditionally into the Chinese approach. Bearing in mind the kind of difference you outlined, could you elaborate on that? Starting with Matt, where do you feel that difference arises? We have also had a lot of evidence on private expenditure, where the US outdoes China substantially. Where are we on an understanding of the public-level expenditure that China is putting behind this?
Matt Sheehan: First, on global diffusion and lock-in element as it relates to AI, it is interesting that China’s messaging as being a global public goods provider started off perhaps with not much reality behind it. Essentially, it stumbled into a real version of that through what Kayla described in terms of China releasing its models openly. While Anthropic released its best model that you can access only by going to quad.com, the entire Chinese model is uploaded to the internet. People with enough technical capability and computing power can fully download the Chinese models DeepSeek and Kimi, and can then tweak them and run them locally. They do not have to interact at all with Moonshot, DeepSeek and these other Chinese companies. Essentially, oftentimes that becomes a much cheaper way to do this for either businesses that are trying to save money or low-income countries that are not able to pay all the subscriptions for this. That shift towards releasing the models open-weight has given some reality to China’s claims to be a provider of global public goods on this.
The flip side of that is that there is not a legal relationship between the people who download and use Chinese open-weight models; for the most part, they are not paying for those open-weight models. Once they are downloaded, they are entirely out of China’s hands; it does not have the ability to claw those back from people using them. In many ways, at a rhetorical level, this has been very good for China in giving reality to the claims made about China providing these global public goods, but it is also not necessarily revenue-generating. There may be some knock-on effects that would eventually lead to revenue, but it is not directly revenue-generating and does not necessarily provide China with the kind of leverage that, for example, debt-servicing can provide a country.
This will evolve over time. So far, the primary mode of international diffusion has been these open-weight models. We will see this change. As Kayla mentioned, some of the Chinese companies are trying to close up their model, releasing them not open-weight but in the same way that Open AI and Anthropic do. That will be a major trend here. Maybe some of China’s provision of these global public goods will start to look a little more like past generations with the Belt and Road, where there are major infrastructure projects that involve those types of entanglements. At least for now, I think the nature of how the technology is spreading does not create the same type of entanglements and dependency. Dependency is somewhat debatable because countries might become dependent on having access to these open-weight models, but it is not dependency on the Government in the same way.
Lord Ahmad of Wimbledon: Kayla, is there a long-term strategy behind this approach of proprietary models against open-weight models?
Kayla Blomquist: It is hard to use the expression “long term” in the AI context. I think a lot of industry players are trying to figure out how to monetise it, as Matt said. That is very much happening at both the labs that are traditionally closed-source, the US ones, and those that lean more in the direction of open-weighting because developing advanced AI is so expensive. How do these companies make enough revenue to convince investors to continue investing, but also create revenue over time? There is a lot of experimentation in that right now from both types of companies, but the way that they will find success I think will have extreme consequences for what lock-in, dependencies and flow of investment in future look like.
To go back to the economic development model question, one thing China has done inside and outside the AI context is to pitch other countries on its economic development model. Having genuinely, over time, lifted so many out of poverty, and having its unique approach, it has gone to other countries building soft power, or trying to, and saying, “You, too, can go on this journey and become a more advanced economy and innovator like us”. That is part of some of the pitch on AI. It is infrastructure as well as technology partnerships.
I agree with Matt’s last couple of sentences and want to build on that. The lock-in potential for dependencies on AI models is real but potentially quite different from physical digital infrastructures. There might be more room for intentionality at the adoption of, say, open-weight models. They become really high switching costs only if you integrate it in such a way to make it hard to switch to another model when another one becomes available in some form or another. A lot of security analysis and additional research is needed on this, but the potential for strategic manoeuvring to avoid real lock-in dependency is a bit more there than, say, building a whole 5G telecommunications infrastructure or more physical objects. That is probably good news for middle countries trying to navigate the adoption of different models from both the US and China.
Q32 Baroness Blackstone: On China’s multilateral engagement, first, what drives it and, secondly, what if anything impedes it?
Matt Sheehan: I am trying to think of balancing the AI piece versus others. Starting with AI, I guess its main goal is a reputational one, or soft power goals, and then there are economic goals for China. Its first major foray in this is the creation of the World AI Cooperation Organization, WAICO, in Shanghai last year. What is actually happening there is still very vague; I think it is vague to the creators of the institution and to those countries that have signed up. The general thrust of it is, “Let’s help you build up your own AI infrastructure in a way that can speed up adoption of AI in your countries”. I think the unstated next step of that is, “And we will bring our AI products to bear there and you will be a customer”. It is not entirely self-interested. There is some genuine element of believing in the power of development and that other countries have a right to develop. I do not want to pretend that it is all some master plan, but directionally that is where that is pushing.
As for its engagement with the UN, every document about multilateral engagement on AI will say, “We support true multilateralism, with the United Nations as the core”. Why the United Nations? I think you will get very different responses from asking people in, say, the American strategic community and those in China. People in the American strategic community will say that this is because China knows that nothing meaningful will ever happen at the UN, or there will be no meaningful costs or restrictions imposed on China by way of the UN.
China would say it is because the UN is the centre of just—I am blanking on a word. It is a little bit later here in China and I am jet-lagged. It is the area of legitimate multilateralism that is global. The key factor leading that choice is that China loves the UN because it always has a seat at the UN. China always has a powerful veto seat at the UN and is always afraid of being encircled and excluded. When the United States and UK were jointly creating the International Network of AI Safety Institutes in November 2024, one big question was whether to invite China to join the network. It was complicated, because China did not have an AI safety institute at the time, but the main issue on which I was going back and forth between the places trying to communicate was that this was not another example of China being excluded from a network that will then be turned against China to hold it down, because I think that is its greatest area of sensitivity in global affairs.
Kayla Blomquist: I agree with the point Matt made about what the US strategic communities would say versus the Chinese. I think my take is somewhere in the middle in that, in some way, China has more of a say relative to other arenas. In the UN, it has been quite strategic in terms of placement of critical positions in UN working groups and committees, and also in conjunction with its approaches to positioning itself as a leader of the global South, which lines up with that approach. There is perhaps more representation of the global South, so it might expect to have more of an advantage, or at least less disadvantage, than in some other international fora and possibly for that reason would want to centre the UN for AI governance discussions.
It will be very interesting to see to what extent WAICO and other initiatives at the multilateral level play out as arenas. China is not trying to position them as a competitor to UN spaces. It says it supports both. We will see where some of these discussions take place. That is something to keep an eye on. It is not meant as a competitor to the UN. It intends it to be a complement. There are limitless discussions that we can have at the international level. It will be interesting to see where the momentum builds. It is potentially more likely at the UN level because it is a much more existing infrastructure.
Q33 Baroness Fraser of Craigmaddie: I was struck, Mr Sheehan, by your point that AI is a very diffuse technology, and that China is interested in the application of that technology. This committee is rather struggling to narrow down the application of this technology in the multilateral regulation and governing space. Are there any specific areas rather than organisations, whether that might be safety, arms control or technical standards, that are more likely to see China want to engage in some form of multilateral regulation?
Matt Sheehan: I think technical standards are the best bet there. China is a very active participant in technical standards internationally. It is very active on building up AI-related technical standards domestically. This is one of my main areas of research right now. It started regulating AI primarily through regulatory documents with standards as supplemental to that. As the pace of change has accelerated, it is increasingly relying on technical standards, maybe not as a first draft of regulation but as a first pass at trying to build in some sort of safety checks along the way. That is primarily in its own domestic standardisation process. It is also very actively engaged with—I am going to get some of the acronyms wrong—the ISO and the ITU certainly as a UN-based one. I am trying to remember which group it was. ISO/IEC JTC 1/SC 42, if I have the number right on that, is an AI-focused technical standards group that does a lot of work on safety.
There is a big open question of what kind of role international technical standards can play in AI safety. International technical standards tend to be a very slow process, and they tend to be more about codifying consensus as opposed to paving the way in the future. China’s domestic standards can be a little bit different because they can be done more quickly and they are more directly connected into government demands. It is easier to build consensus domestically on that. International technical standards are a very good starting point, with the caveat that these might not move at the speed that is needed if we are trying to work on safety-related issues. I will leave it there for now.
Baroness Fraser of Craigmaddie: Ms Blomquist, would you agree to focus on something like technical standards? You said that China was not in the same place on safety as the UK and the US. Does it need to get its own house in order domestically before it is ready to engage multilaterally on specific areas? Are there specific areas in which it would be more likely to engage multilaterally?
Kayla Blomquist: That is a good question. In terms of getting its own house in order, that has been something that Matt has done a lot of work on over the years, as have many in the broader AI governance community. It has been the case that a lot of Chinese AI developers do not have as advanced internal safety practices as US labs, and as promoted and developed by AISI here and so on. There is definite room for improvement and need for improvement on that front. On the international engagement on AI safety issues, attempts to create more of an equivalent to AISI domestically have been in place, and there is still movement in terms of creating a cohesive domestic house for a counterpart organisation. There has been a lot of effort, and there is something resembling a counterpart on that.
On the willingness to engage on AI safety as well as preparing for potential international crises and incidents stemming from advanced AI, there is an openness. There is still a lot of work to do in the practical sense, both in China and in other countries, majorly. There have been efforts, also supported by the UK, that have gone on between the US, China and the UK for starting to build these plans and what even a counterpart situation in different forms of crises would look like. There are many forms that they could take, with AI being such a cross-functional technology. There is still so much more work to be done on each side, in co-ordination and figuring out which specific elements of co-ordination are necessary versus each country getting its own domestic house in order. There is a lot of work to be done on this. There is absolutely space for the UK to continue to be a leader and convener as well as pushing for some specificity from China and other countries on this.
Q34 Lord Bruce of Bennachie: I am interested in the offer that China is presenting to developing countries, namely, “We’ve got this great technology. It’s there for you to use. We’ll help you use it. It’s open source and you can retain the data”. That is very appealing, whereas if you look at America you could not trust anything about it because it is only doing it for American interests. It might want to manipulate it. It is closed source. It wants to use your data. So my question is: is there a role for the UK here? We have substantially dismantled our aid programme. Is there a role for the UK to work with developing countries, first to act as an honest broker in this situation and, secondly, to develop some niche capacity of our own that is relevant to that? Is that something that would perhaps restore the credibility of the UK’s engagement with developing countries, which clearly has been seriously undermined?
Matt, you said there is an element of self-interest as well, all the time. Even with our aid budget at its most altruistic, there was always a recognition that there would be a benefit. There is nothing wrong with that. But at the end of the day can you project to say, “We really genuinely want this to be a product that benefits the whole world and doesn’t become another instrument of a second Cold War between the two great powers”? Is there a role for the UK to do that? If so, how could it do that?
Kayla Blomquist: That is a great characterisation of the pitch that China is making to the rest of the world as well. In terms of the UK’s role, being American and formerly in the US Department of State but choosing to be here and to do this work, I absolutely believe that the UK has a unique role to play, in particular not just because of the special relationship with the US and ability to engage in narrow and specific ways with China, but because of the talent and expertise ecosystem when it comes to AI safety, security, governance and so on. The AI Security Institute is such an exceptional example of innovation on the international stage of what kinds of institutions other countries can build out to navigate what is in their own interest and need, as well as for other countries, and really supporting it and building out how that can be modelled and extended in expertise forms to other countries navigating this for themselves.
When it comes to the infrastructure on the aid front with other countries, that is a great question to keep diving into: how then could the UK bring that specialised expertise that it absolutely has, more than almost any other country, to other countries? That, too, will need to navigate many different questions across the whole economy and society. Are these models safe to deploy in our own country? There is a lot of work to be done there from a government aid and support perspective, as well as on a services industry perspective. There are some here in London working on the AI assurances industry, which is both public civil society as well as private sector types of many different evaluations, and plurality is needed with such an expansive technology as AI, leveraging that as a national export in a way that is useful to other countries in this time.
Lord Bruce of Bennachie: Thank you. Matt, what is your take on that?
Matt Sheehan: I am very much in agreement with Kayla. The UK has already built up a deserved reputation and brand as being one of the premier centres for the testing of frontier AI models for different risks. That testing tends to focus on frontier risks. Is this model going to go out of control and defy the wishes of its creator or manager? Can it contribute to the uplift of a normal person’s ability to create dangerous chemical or biological weapons? That tends to be where the testing work by UK AISI focuses a bit more. I could be a little bit out of date on that. That is my memory of it.
Countries around the world are going to care about that, and they are going to want something of an independent voice to do that. I do not think they are going to just trust the models coming out of China as being safe to deploy widely across society, and they might not trust the evaluations that come out of the US. The CAISI, formerly the AI Safety Institute in the US, has evaluated at least one Chinese model, a DeepSeek model. When you read the introduction, it was a very ideological document, laying it on thick about the role of this model in disseminating CCP propaganda and so on. That is a legitimate issue and something that many countries will care about in a certain sense, but I do not think that those documents coming out of the US system will be read as independent and neutral documents.
Another element of this would be, through conversation, gathering insight from developing countries about what they actually care about when it comes to evaluating a frontier model. Are they focused on, in certain ways, these frontier risks and control over AI agents? Do they care more about whether this model is going to infuse CCP ideology into their countries? Maybe places like Vietnam or the Philippines would care more about that. There are a lot of different aspects. Learning from other countries in terms of what they care about and then putting the UK’s existing excellence in testing of models to use, it becomes a global public good and one that the UK is in a unique position to offer.
Q35 Lord Grocott: This develops from what Matt just said. I know this is an oversimplification, but you can think of the US on one side and China on the other as a spectrum, from the “let things rip” as far as the US is concerned to “let’s regulate things” as far as China is concerned. You referred to developing countries—whatever phrase you want to use, whether it is global South or non-aligned. They have had all sorts of different titles over the years. This is probably an unanswerable question, but I will ask it. Where would you put the balance of who is winning on that argument with countries that are not directly aligned with the US or China? Where is the balance of world opinion on this spectrum?
Matt Sheehan: To really answer this, I would need to spend a lot more time in the global South than I do. I am mostly in the US and China. From meeting a lot of officials at these international AI gatherings and from looking at the polling on this, they are overwhelmingly on the side of, “Let’s use this technology to upgrade our economies and to improve our livelihoods”. There is some irony here. The Trump Administration are very much in a “let it rip” type of phase, with no regulation. US companies themselves are actually quite concerned about the risks and investing a lot of time and energy into the safety side of it, even if that is not showing up in results at this time. The fact that the models are served through an API and charged for means that they are not nearly as accessible. While the US might have a “let it rip” policy or philosophy, the models are not actually as accessible and diffusing as much around the world.
China’s pitch is, “Look, we have used technology to lever up our economy several times over, and, while we might put a lot of regulation on the way that these models are deployed domestically in terms of speech, we want to share the models with you, and we want you to use them to lever up your own economy”. That has a lot more traction than the message either from the US Government, which is, “American AI dominance”, or from the frontier AI labs, which is, “We’re going to build these exquisite products, but they’re also potentially very dangerous, and we’re very worried about that”. I do not think either side of that US pitch is really landing necessarily, while the Chinese pitch does. At the same time, people who are the most sophisticated users of the models—people around the world, including in the global South—who really want to work with the best models, if they have the ability are still in large part opting for US models. That is maybe only a sliver of the user base in certain countries. Even in China, people who work for Chinese AI companies building frontier models still often find ways to use the US models because they think that they are better. So it cuts a few different ways.
Kayla Blomquist: I agree. It is a time of flux. Certain trends are taking shape that will maybe over time lead us to a clearer answer, one way or another. It is a time very much of hedging of other countries throughout the global South and other middle powers. As Matt said, if a country, business or individual can afford frontier-model subscriptions or the more fundamental infrastructure investments with US companies, many of them tend to go in that direction. At the same time, many countries, businesses and individuals out there will opt naturally for models that are very good and useful but more affordable and accessible, and often those may tend to be Chinese models. There will be a few years of flux on this with different tiers of advancement of economy.
This is also playing out in a time of uncertainty at the traditional diplomatic allyship level. Pax Silica is the US answer to WAICO and other initiatives. It is very focused on deal-making at the infrastructure level, and we will see the effectiveness of that over time. There are definitely deals in progress and under consideration. It is quite in the middle, even though there is not perfect global public opinion data out there. Keep an eye on what is genuinely usable, accessible and split out from the frontier. There are actors that will absolutely seek that out and only go to what is the most frontier level of advancement.
Q36 Lord De Mauley: You both referred to China’s launch of WAICO while simultaneously maintaining support for the UN as the focus for multilateral co-operation on AI. You started to address how we should understand this approach to institutional co-operation on AI, but could you expand a bit on that and on how significant the development of WAICO is in the wider context of multilateral AI governance?
Matt Sheehan: At this stage, multilateral AI initiatives will be most suited to and effective when it comes to international diffusion of the technology. In the short term and maybe medium and long term, they are not going to be the place where binding constraints on or requirements for AI development come into play. As much as China likes to talk about multilateralism, it is not trying to sign up for any multilateralism that constrains what it can do when it comes to developing and deploying the technology.
Setting aside the governance side of things at least for now, in the multilateral space China is aggressively pushing this diffusion narrative. People always tend to underestimate the long-term success of various Chinese initiatives. We underestimate its technological progress. I am very hesitant to write off WAICO so early on, but it sputtered out of the gate. It was announced in July this year in Shanghai, and an initial list of countries signed up for it. It was not the most impressive list of countries, with all due respect to them. It was not the people whom you would try to get to sign on to your multilateral AI initiative from the start. That was due in part to scepticism from a lot of countries about China’s intentions and the vagueness of what it was offering. I spoke to officials from some countries that were invited, and they had almost no details on what it meant to sign up to this initiative; it just meant that you were signing up. If you are not going to have something relatively concrete to offer, you can still download China’s open-weight models and use them. You do not need to sign up to WAICO to do that. At this point, WAICO feels a little bit more political.
I wish I could fact-check this in real time. One of the initial initiatives from WAICO that would be substantive is in the area of international standards for interoperability of AI agents. China is building its agents. If those agents are interacting with agents in Indonesia, do they have a way to shake hands? Do they have a way to understand the identity of the other side, and what tools and capabilities they have access to? That is a really important question, and China is building the domestic standards on this. My guess is that these technical standards are not being built, or being designed, in a lot of the other countries in WAICO, and China is maybe using this as a chance to export some of its technical standards. That would be good for China because it would become a lot easier to continue using Chinese agents in this way, and it would be good for the countries, in a way, because they have some participation in the standard-making process. It can be an educating process for them because they just do not have the same technical standards bureaucracy. There can be things that WAICO can offer in this space. Maybe that is all still TBD, and still being figured out in real time. It is not quite so clear as, “This is the belt and road initiative. We’re going to build a port. You’re going to pay for it. We’re going to build a railroad, and it’s going to connect to your city”. It is not the same type of deal on offer this time.
Kayla Blomquist: In a different way, paralleling it to the belt and road initiative, there is a little bit of a “possible but TBD” element of “build it and they will come”. It is creating an organisation that is more at the political and signalling level, and somewhat club-like and trying to build out the political influence and clout of China as a global leader in AI. There is a strong element of that.
Zooming in on what could happen and is actually happening under this WAICO umbrella, in addition to technical standards, we have heard at least gestures at a couple of different things. Some of that would be educational exchanges and capacity building at more of a technical development level, in particular with countries throughout the global South and the member countries, some of them creating programmes at major Chinese universities. There is signalling in the direction of China offering some of its models to member countries, but there is a question of how that is different from what is already out on the open-source platforms available for anyone to download. That is also a bit TBD. One specific model, though, that China has been promoting in general and in relation to WAICO is MAZU, its weather-prediction model, which is also in line with more digital infrastructure sales and ties such as internet-of-things technology for flood monitoring as one example. That is an interesting subtheme within this. That goes into the question of what is actually happening within WAICO. There are some things, but it is perhaps not as comprehensive at this stage as some might think.
Q37 Baroness Prashar: Against the background of what you said, would there be any benefit for the UK and other Western countries to engage with WAICO at this stage?
Matt Sheehan: It would be worthwhile to engage in the sense of, “Let’s talk about this. Can you describe what this actually is? Can we figure out what it is?” I do not at this point see a lot of benefit in joining WAICO. Primarily, China would be very happy to sign people up to get them to lend credibility to this, but I do not see what the benefit would be to the UK or other countries at this point from joining. Yes, light-touch engagement to essentially understand better what is going on and what is on offer, but I do not see a need to sign on to the initiative at this point.
Kayla Blomquist: I agree with that. Looking at what is happening under the WAICO umbrella and what are the UK’s strengths and interests, as briefly talked about before, there is the possibility for the UK to do capacity building in, for example, AI evaluations and the AI assurances industry. There conceptually is some possible overlap, but to deploy that type of foreign assistance well does not require doing it within a WAICO umbrella, and perhaps it would be more efficient and in line with more traditional country approaches to foreign aid and development to do it more unilaterally. It will be worth keeping tabs on, in the capacity-building front, what is happening within WAICO.
Matt Sheehan: On the capacity-building front, one of the other things that China said it will offer is training for people from WAICO countries exclusively. I forget what the number is exactly: 5,000 training slots over five years or something like that. That really is a potentially meaningful offer. Most of these countries are not going to train their own foundation models. That is a $100 billion project over the long term. Training people to be able to use and deploy the open-weight models that China has already trained and are on offer could be a very meaningful thing for other countries and meaningful for China in drawing these countries closer together. Maybe that is another area that I would watch in particular to see if it gains real meat on the bones.
The Chair: That is a good point.
Q38 Lord Alderdice: Picking up on that, we have marked out how one of the hallmarks of China’s foreign policy is an attempt to position itself as a developmental leader and champion of the global South, or whatever we choose to call it, and that this is influencing its approach to AI. Looking at it from the other side, do you have any sense yet as to how China is being perceived on this front by those countries to which it offers these opportunities? Is it winning influence at the expense of the United States and the West? Do either of you have any sense of that?
Kayla Blomquist: I have not looked at specific countries recently, but generally in many countries there is distrust of both great powers within AI and through other forms of engagement and investment. However, a lot of countries have often made deals on infrastructure and now moving into AI with China and Chinese companies from a “This is a relatively good deal for us” perspective. In many countries, there is scepticism of some of these offers, but it might not lead to not taking the offers or making the deals because some of these, whether it is AI models specifically or surrounding infrastructure and services, may be more affordable, accessible and on a practical level what a given country may assess that it needs at a given time, regardless of whether there is a strong sense of trust there. I think it will be very mixed across the board.
Matt Sheehan: I have a brief data point and a couple of thoughts. I spoke to one relatively senior government official during the rollout of WAICO who had been asked to join, and their basic response was, “We don’t know what this is, and they can’t answer the question when we ask them, so we’re not going to be joining”. That is from a country that is a little bit better positioned in this overall. For a lot of countries, this may be the only AI-related invitation to the table that they get. People in every country have an overriding feeling of, “We need to understand this. We need to find some way to get on the train when it comes to AI. This is happening one way or another, and we need to get our foot in the door”, or whatever metaphor you want. If you are in one of those countries and China has invited you to be part of this admittedly vague organisation and is saying that it will help you with some training and stuff like that, that might not sound like a great offer but it might be the best and only offer that you really get in a given year.
The offer from the United States is, “Feel free to pay for our models, or join if you have certain critical minerals. Join Pax Silica so that we can guarantee our supply chains here”. Light as it might be on details, it still might have some power. The area where there might be real soft-power gains is within the technical and entrepreneurial communities in these countries who are using and deploying Chinese models. That is where a very tangible thing is being provided that they would not have access to otherwise and it really empowers them to do different things. That is having that power, but that again does not stem from WAICO; it just stems from the open-weight model that it uses.
The Chair: One of my colleagues once said it was impossible to under-flatter a politician. I think it applies to countries as well, which seems to be what the Chinese are doing.
Q39 Lord Lamont of Lerwick: It could be argued that China is endeavouring to fill the space that has been left by the United States, given its attitude to competition with China—although I have to say that the witnesses we have had to this committee have generally denied that China wants to replace the United States in international institutions and prefers pre-eminence or a leading role simply because of the cost and inherent cautiousness. Is it maybe different with AI? That is the first question. Secondly, in general terms, how would a Chinese-led international governance regime on AI differ from an American one?
Matt Sheehan: I will start with the second one. As I mentioned before, I do not think that it is looking for global governance to turn into something intrusive towards China or other countries. The No. 1 principle that China has carried over from its view of international internet governance into its view of AI governance is respecting the sovereignty of countries when it comes to their decisions about how they want to manage the technology domestically. In the internet era, this was internet sovereignty and cyber sovereignty, which began very much from China defending its own position. When China was one of the few countries that had really meaningful controls on the internet, it described those as a sovereignty issue, and said that every country has the right to impose those. That does not mean that China respects everybody’s sovereignty in cyberspace, but that is the principle that it carries over.
If it were to take on some leading or guiding role in global AI governance, central to that would be prioritising sovereignty and saying that it is not the job of a global regulator to come in and say, “You cannot do censorship that way. You cannot regulate your models that way. The core of activity has to be at the national level”. That is maybe its starting point for a global governance regime. From there, it would probably move into focusing on AI for development and interoperability of systems: can we make sure that our systems can be deployed globally and that agents are able to connect with each other in the right way? Questions of frontier AI safety would, in China’s mind, be left to a domestic issue within China and a domestic issue within the US, and something that maybe those two countries are going to talk to each other about, but I do not think they would put that on the table when it comes to a global governance regime.
On the question of whether it is China’s goal to displace the US, I generally agree with the previous witnesses whom you have had in that China wants to carve out a space for itself and wants to have freedom of movement, but it does not want to take on the responsibilities of being an international, even global, security provider. That is partly because of the sovereignty issues it describes, but it recognises that that is quite burdensome. Maybe China’s most successful era of development was one in which the United States still played a role in maintaining a variety of global norms, global institutions and security guarantees for other countries. China was able to develop on its own terms and leverage that international system and the World Trade Organization for its own development. I agree that it is not looking for displacement if it means responsibility. How all that translates into the AI era, I would have to think a little bit more about that to give a good answer.
Lord Lamont of Lerwick: Thank you. It was a very good answer.
Kayla Blomquist: In terms of the sovereignty-centric, internalist definitional approach to global AI governance, I agree with that. To further build on that, with regard to a lot of the promotional materials ranging from academic papers to MFA messaging and others that are involved in the shaping of WAICO, part of the pitch has been on paper and explicitly been, “Your membership in WAICO is not tied to your governance model. It is not tied to you being a specific kind of country and running your domestic affairs in a specific way”. This is a very specific-to-China way of framing sovereignty, definitely going back, as Matt mentioned, to internet sovereignty. Now it is making appearances here in the multilateral and global AI governance space, which is quite interesting.
That connects very directly to the space left by the US. For a long time, China has been very publicly criticising the US on its approach to global affairs and multilateralism, tied especially to the sovereignty and domestic ways of governing point, and then leading into the AI era. That will continue to be part of its selling point.
Is AI different from previous times? I also agree that China has a not-forward-leaning and hesitant relationship to not necessarily aspiring to be a global hegemon in the way that the US or others have been in the past. That is quite strongly upheld by a number of experts. It still has a complicated relationship, as we have talked about today, to positioning itself in part as a provider of global public goods, but cautiously and in sometimes very specific ways. With open-weight models and positioning itself as a leader of the global South, it is almost accidentally positioning itself as being in this role of responsibility.
In terms of the expert and adviser level on AI governance, there has been hesitation around WAICO and other initiatives as to what the consequences of this positioning as the responsible leader and convener of WAICO or other initiatives are. Even if you are narrow in the offerings within WAICO or other initiatives, other countries might come to the table and with their membership now expect certain benefits that China may not even have offered in the first place. In some conversations with people who were involved in the shaping of WAICO at the more academic level, they internally seemed quite concerned about this positioning and its consequences thereon. It will remain a complicated relationship. Some of it depends on how other countries force China into that position now that it has claimed “great power” status, even if it is not trying to claim the “one and only great power” status internationally.
Q40 Lord Houghton of Richmond: My question was answered by what you said to Lord Bruce, so I will go somewhere else with a question. Our committee is international affairs and covers both international relations and defence, and there has not really been a hard defence question about international regulation, particularly in the context of artificial intelligence and autonomous weapons systems. I know, Matt, that in the Carnegie Endowment for International Peace you are international advocates for peaceful outcomes, and I know that you advocate for the retention of meaningful human control within the use of artificial weapons systems, but in many respects that is purely about the softer issues that relate to compliance with international humanitarian law and international legal constraints on the use of force.
What advantageous proposals do you have for regulating the capability itself? Certainly, the United Kingdom and America both see the discriminatory advantage of the better exploitation of AI and autonomy to be things that they want to pursue for natural military advantage. Are you tilting at windmills in a way, or is there something else within the regulatory body other than straightforward safety that you are looking at here? I will come to you, Matt, first, because you represent an endowment that is meant to be doing some of this stuff, and then, Kayla, you might have a supplementary.
Matt Sheehan: I am afraid I do not really have a great answer for this. People at Carnegie do this work and do it very well. My colleague Steve Feldstein is at the forefront of many of these issues. I really do not focus on military applications of AI. My gut instinct—you said at the end about tilting at windmills—is that this is seen as the source of so much potential advantage on the battlefield that it is quite hard to imagine countries really showing restraint in this. That is an amateur answer, I am afraid. I do not really have more for you.
Lord Houghton of Richmond: Kayla, do not worry if you cannot answer, because we are going out to ask these questions in our next bit of evidence research. Is there anything that you would offer?
Kayla Blomquist: This is not my area of specialty, either. However, my colleague Thomas Watts, at the back here, might have some helpful thoughts afterwards. The supplementary point to add, especially in terms of autonomous weapons, is that in figuring these things out coalitions of middle powers are specifically relevant. It would need to be a coalition of middle powers with the expertise to understand the defence applications and the severity of potential consequences, but in a way that is aggregating enough voices in the room to hold any given power to account for irresponsible development or deployment of these new capabilities.
The Chair: Okay. We will pursue this at some later evidence sessions. The rest of what you have said today has been profoundly useful and stimulating. Thank you so much for giving us your expertise, which has clarified a lot of the issues that we will have to deal with as well. We thank you very much for giving us your time and expertise today.
