26
Examination of witnesses
Sam Daws and Isabella Wilkinson.
Q108 The Chair: Good morning and welcome to the International Relations and Defence Committee of the House of Lords. We have been looking at the multilateral system, and, specifically now, we are looking at how it relates to artificial intelligence.
I have received apologies from Lord Houghton and Lord Darroch. Our special adviser, Andrew Ehrhardt, is in Boston today and is joining us by Zoom in the middle of the night. I thank you very much for making yourselves available today for our interrogation and addressing the subject of how multilateral institutions might deal with the problems and difficulties of AI. I remind you and others that this session is being streamed live on Parliament television. A transcript will be taken, and you will be able to make any minor changes to it. You will get a copy of the transcript. I remind members that, if they have relevant interests, they should declare them. At the very beginning I declare that I am a senior counsellor with the Washington-based Cohen Group.
I will ask the first question. As you know, the committee is looking at the whole multilateral system as it applies to the United Kingdom. We hope to make recommendations on the UK’s policy in relation to the various multilateral institutions existing in the world today. That is why we are looking at the whole question of artificial intelligence in the context of the global picture and the way in which multilateral institutions have already looked at, or may in the future look at, these things.
What are the key governance issues related to AI that multilateral engagement really needs to address? I will ask Mr Daws to answer the question first.
Sam Daws: Certainly, and it is a real pleasure and privilege to be here. There is such a broad list of governance challenges but I will pick out just a few of them. The first would be safety: agreeing common methods for evaluating advanced AI, sharing risk information and responding to serious incidents. We have seen in the news in just the last few days the importance of agentic AI and frontier capabilities in the cyber domain. The second would be international security from cyberattacks to AI-enabled biological threats, and military AI more generally. The third is regulatory interoperability. There are different approaches in the EU, China, US, UK and so on. Those differences are legitimate, but we need to make sure our systems remain compatible.
The next one—this is often neglected but is increasingly important—is cultural and linguistic diversity. AI must not be just for greater efficiency; it also needs to reflect the wisdom of humanity as a whole. There is information integrity, which is co-ordinating responses to election interference that AI can supercharge. There is science and public health. There is huge potential as reasoning models develop and through eventually recursive self-improvement to have a cascade of new discoveries in mathematics, science and medicine. Governance is important to harness that in the most effective way.
The last one—there are many more, but I will just stress these—is the environmental sustainability side of it. AI data centres use a lot of electricity and water. That is the downside. On the positive side, through digital twins modelling the weather and earth sciences, through the amazing efficiencies that AI will achieve across many different industries, and through new discoveries in energy sources and other areas, AI will probably be a plus overall, but we still need to monitor and, where possible, reduce its energy output.
The Chair: Thank you very much. I should have said at the beginning that, when you start speaking, you might care to introduce yourself so that the committee and the wider public outside who might be watching will know who you are as well.
Sam Daws: Shall I do that now?
The Chair: Yes, please.
Sam Daws: My name is Sam Daws. I am a senior adviser to the Oxford Martin AI Governance Initiative at the University of Oxford and the founding director of Multilateral AI. I am a former senior civil servant in the Foreign Office and Cabinet Office. Before that, I worked for the UN, including as first officer to UN Secretary-General Kofi Annan.
The Chair: Thank you very much.
Isabella Wilkinson: Good morning all. What a pleasure and a privilege to be here. By way of introduction, my name is Isabella Wilkinson. I am a senior research fellow at Chatham House, an international policy institute, where I am part of our digital society programme. My work focuses on the global governance of emerging technology and disinformation. I am also a doctoral student at the Blavatnik School of Government at the University of Oxford, where my research looks at how companies work together on AI security. My research is supervised by former UK cyber chief Ciaran Martin, and professor and expert in technology governance Roxana Radu.
To answer the first question, if it is okay to dive straight in, I echo Mr Daws’s diagnosis of some of the key governance challenges that many of you will have been confronted with, especially over the past few months and particularly over the past few days. It is helpful for the committee to visualise key governance challenges as falling into three categories. I will walk through each of those categories and, hopefully, this will help us structure some of the answers in the session going forward.
The first type of governance challenge is dealing with shared risks. As we have seen over the past few weeks, few days and even few hours, risks generated and triggered by AI are cross-border. They are inevitably globally shared. They are cross-domain; they might jump from one part of the economy to the other, from one domain to another. These include malicious risks—we have seen concerns about cybercriminals developing a bioweapon—and loss of control risks. This is when AI models and systems are behaving exactly as they were designed to, but they are performing in ways in which their designers no longer expect them to do. This is called a loss of control risk. Finally, there is a systemic risk. These are risks associated with the adoption or diffusion of AI systems that lead to risks such as discrimination or labour displacement.
To recap, there are malicious risks, loss of control risks and systemic risks. We are facing a governance challenge because all these risks are interlinked. We have to be very strategic about how we design governance at multiple different levels. This is a multilateral challenge.
The second category of governance challenges that we are confronting at the moment relates to trust and information, to echo what Mr Daws said. There is a dearth of trusted information about AI, its risks, its impacts and the opportunities that we may be able to collectively and multilaterally harness. There is an emerging scientific evidence base, but it is by no means globally sufficient. We are missing a testable picture of risks and opportunities.
The third category is preparedness for disruption. As we have covered briefly, AI risks are inevitably cross-border and cross-domain. A key governance challenge that companies and countries will face is how to prepare for any range of AI-related and AI-driven disruptions. In other words, a key governance challenge is our lack of a crisis toolbox—networks, strategies and resources—to deploy in the event of a major crisis, or even a small one that triggers broader disruption further down the line. In short, we are facing quite a severe state of gridlock in multilateral AI governance due to these interlinked governance challenges. I would refer the committee to the UN’s preliminary international scientific report on AI that platforms global emerging consensus on the state of risk and links it to many of the different governance challenges that I have mentioned in my remarks.
The Chair: Thank you. That is extremely useful to us and well-structured as well. Lord Ahmad will ask a question.
Q109 Lord Ahmad of Wimbledon: Welcome to you both. It is good to see you again, Sam. We see a variety of international organisations—you mentioned one of them, the United Nations, but also the G7, OECD and the Council of Europe—all looking at this whole issue of common international governance standards. Previously when I served in government looking at cyber governance, there was always a challenge: the Chinese version, the American version, the European model, and which would apply. How do you characterise the current state of play in different countries and the appetite that there is for international governance?
Isabella, you touched on the fact also that initiatives are under way at the UN and whether they are effective. Each country, in a way, is pushing its own agenda. In that context, perhaps I can start with your good self. Building on what you have already said, is there space for one organisation, or indeed a series of organisations, to follow a particular governance model?
Isabella Wilkinson: I broadly agree with your assessment. We see a multitude of different governance initiatives at different levels—different regional groupings with different aims and different modalities. It is helpful to characterise this as a global patchwork that is currently stuck, and in a state, as I mentioned before, of intense gridlock. There are many reasons for this gridlock: some of the governance challenges that I set out in my first answer and some of the competition dynamics between the US and China that all of you will be incredibly familiar with.
The bottom line here is that multilateral AI governance as pushed forward in a variety of different regional groupings, international institutions and other arrangements is nascent. It is still taking its first steps. There is much dialogue and much talk, but, as it stands, our case studies of implementation and actioning these commitments and regulations are just years old. We are talking about an incredibly young space.
A key example of this is the UN’s inaugural meeting on AI governance. In July, it convened over 3,000 participants and over 100 UN member states to discuss ways forward for AI governance. It was a valuable platform and some people called it a dialogue of dialogues, but it did not lead to a binding commitment. What we face right now in the state of play is a variety of different spaces for dialogue and consensus building, but a lack of connective tissue to actionable implementation. This is not an impossible bridge to cross, and I look forward to exchanging ideas on how to do so.
I am happy to walk through some areas of success, failure and inertia, if that would be helpful, or I can hand over to Mr Daws.
Lord Ahmad of Wimbledon: Yes.
Isabella Wilkinson: Great, I will go ahead. Please cut me off if I am taking too long. Within this global patchwork, this somewhat severe gridlock that we have sketched out, there are some spotlights of success. I would categorise these as successes in evidence and documentation as a basis for governance. From a technical perspective, it is really exciting to see emergence of global best practice on how AI incidents are reported on. The way that we report on what goes wrong and how we mitigate it is becoming increasingly standardised. As we have seen in other areas of governance, this is an essential risk mitigation measure. An example of this, as I mentioned before, is the preliminary report from the Independent International Scientific Panel on AI and the International AI Safety Report, which kicked off after the 2023 Bletchley summit. Another example is the standardisation of approaches to transparency around AI, which the G7 and OECD have very much been leading over the past few years.
What we see here as a model of success is a theory of change. Small groupings come together and agree that a technical best practice in governance works, and increasingly it snowballs and scales to become multilateral best practice. We have seen this success in evidence in documentation, although it is still nascent, and we should learn from this impact pathway for the areas of failure and inertia in multilateral AI governance, which I am going to cover now.
Nothing about “pacing the frontier”, which is terminology you would have all encountered over the past few days, is internationally mandated. The bottom line here is that there are some binding rules. For example, the EU AI Act has reporting requirements on AI models and providers, but nothing is internationally mandated. This is a massive governance gap. I am sure we will discuss this later in the session. There are areas of inertia in multilateral AI governance where there is growing global consensus that something should be a priority, even glowing global outcry, but a lack of action due to this gridlock.
A really chilling example of this is the human rights abuses that AI systems may enable: the integration of AI in a military domain, leading to severe civilian harms in places ranging from Gaza to Iran. It is not to say that there has not been movement on a smaller level. The Council of Europe convention on AI puts human rights front and centre, but, returning to our theory of change, there needs to be connective tissue between consensus and shared language in smaller groupings and multilateral consensus.
Let me finish my remarks by underlining the fact that we are in a moment of intense crisis and turbulence in multilateral AI governance. As we have learned, and I am sure this committee has learned, from many other areas of governance in the past, these conditions create a window of opportunity for dealing with the gridlock. I would love us to return to this idea later in the session.
Lord Ahmad of Wimbledon: Thank you. Sam, if I can just pick up a couple of those, there is a general point that the whole international rules-based order is chaotic. Is there an appetite to use multilateral governance structures to pursue this? Everyone believes something should be done, but others would say, “Yes, it should be done according to our way”, and others are not quite clear which method of governance to follow.
Sam Daws: Yes, there is a distinction in governance and regulation between voluntary principles and enforcement. I agree with Ms Wilkinson that it is both fragmentary and patchwork. I have been quite surprised by the degree of universal—through the UN—, regional and minilateral initiatives that there have been over the past few years. The UN ones are in part due to the fact that António Guterres was the first UN SG with a science and engineering background. He set up a UN advisory board. As we have heard, we have had a scientific panel and a largely talk shop-focused global dialogue but valuable in itself.
I would probably divide it across those three things. Certainly under the previous US Administration, we have seen a willingness to set principles at the UN through two UN General Assembly resolutions in lockstep with the Chinese. There was a resolution first in 2004 on AI safety and then one on AI capacity building. The US put forward the first backed by China. China put forward the second backed by the US. There was a real recognition early on and a desire between those two countries to have some kind of framework. At the regional level, the EU has had a very much regulation, risk-based, law-based approach, while other regions have focused more on AI sovereignty, and, in the case of Africa, focused on datasets, privacy, languages; similarly with ASEAN, Central Asia and Latin America.
The most interesting area, in answer to your direct question, is the innovative minilateral initiatives that have taken place. You have the Western-based ones. The G7 has been mentioned. There is the OECD, which expanded adding in Brazil and India by integrating the GPAI, Global Partnership on AI. The Bletchley process that Prime Minister Rishi Sunak initiated was an important minilateral that has continued through different incarnations in different countries—Seoul, Paris and New Delhi. Next year it will be in Singapore. The Singapore Consensus, which is a research consensus, brings together the US, Chinese, UK and other researchers from 13 countries, focusing in the last incarnation on the risks from agentic AI. So you have Western-based minilaterals and neutral minilaterals crossing over. The Digital Cooperation Organization, an initiative from the Kingdom of Saudi Arabia, is an important one that crosses different regions. The Netherlands and the Republic of Korea have a minilateral on responsible AI in the military realm. Then you have the more Chinese-centred ones: the Shanghai Cooperation Organisation’s AI engagement; BRICS Plus; the Digital Belt and Road Initiative; and most recently—I was in Shanghai in July for it—the launch of the World AI Cooperation Organization, which builds on a Chinese focus on the diffusion of AI technologies, not necessarily at the frontier.
There is a real mixture of different approaches from the universal, the regional to the minilateral. I agree that we need connective tissue between them, and we need to embrace the private sector because it is key to solving this. There are things like the Frontier Model Forum, the private sector’s involvement in the various safety summits in terms of reporting, and we have frameworks such as the G20 and APEC, which could also be good places to embrace private sector involvement, but this is very much in its infancy.
Lord Ahmad of Wimbledon: Thank you both.
Q110 Baroness Crawley: You are very welcome. It is good to hear that there is myriad connectivity, even if it is not strongly underpinned by law at the moment. However, what are the really hardcore challenges to multilateral co-operation on issues such as artificial intelligence and common international standards? We were given a very useful diagram on page 8 in our briefing, which was the stack, in which it has the key governance bodies for each layer of the stack: the social layer— ChatGPT, Gemini and so on; the logical layer—OpenAI and DeepMind; and the infrastructure layer—Amazon, Microsoft, Google and so on. Are the challenges to agreeing standards more in the top layer than in the technical layer? Are they more in the logical layer than in the human rights layer? Where are the challenges?
Sam Daws: I would begin by saying that there are three overarching challenges that pertain to the whole stack. The first is the geopolitical and dual-use nature of AI. The frontier models that we are talking about have civilian uses, but they can also be used in military contexts. Nearly all the international agreements—the UN, OECD, the Council of Europe and so on—have a sentence saying, “This applies just to the civilian use of AI”. That is one of the fundamental things. Feeding into the dual-use nature, we have the geopolitical competition between the US and China, where the US has imposed export controls on the export of high-end chips and EUV lithography equipment to etch the chips in conjunction with the Netherlands. There is a lot of tension around the geopolitical space.
There is also a diffusion competition between the US and China. China has taken the approach of offering a full-stack, low-cost, interoperable approach that is focused very much on what they call being tuned to context, culture and language, so that it might be tuned for Southeast Asian education or African agriculture. There is a real attempt by China to get, effectively, market share and to be helpful to the developing world and to China in setting its standards as the global standards through diffusion. Its approach has also been focused on integrating AI not at the frontier but through embodied AI and manufacturing robotics into every different industry. The US approach, the scaling approach, has been more focused on the aspirations for artificial general intelligence and models that can basically do everything. They have capacities in robotics, in brain-computer interface and so on. It is not an either/or, but there are different dynamics. The Trump Administration woke up to the diffusion approach of China and, therefore, took some steps to move away from Biden’s restrictions on exporting high-end chips to countries such as the UAE and Saudi Arabia because of a recognition of competition. That is the first level.
The second is commercial sensitivity. Governance is so difficult because individual companies, whether they are American, Chinese or British, are unwilling to reveal their model weights or the datasets used to train their models because it could undermine their profitability.
The third is the nature of the technology itself. It is opaque and it is fast moving. Some of the best scientists and technologists I speak to inside the companies that are developing these models do not really understand how they come to their conclusions—what happens within the black box. To govern or regulate that is, therefore, inherently difficult.
We are seeing the changing nature of the technology with agentic AI. It is not just a question of how you test, evaluate, secure and ensure that a model is safe before it is released. That is super important, but that is one part of the puzzle. With agentic capabilities, models now do things and they interact in the world. Agents work with each other in unexpected ways. They collude or co-operate with each other to obtain their instructed end result. We are seeing emergent qualities of deception of agents tricking humans in order not to be switched off or tricking humans in order to fulfil their assigned purpose. It is misalignment problems. We need also to govern AI in the full life cycle: after its release, to see how it develops in the wild. We are going to have 1 billion agents interacting in the wild, and bad actors, whether they are a terrorist group or organised crime or a bad state actor, will also harness some of these tools. There are multiple challenges with governance and regulation.
Isabella Wilkinson: That is incredibly well covered by Mr Daws.
Baroness Crawley: Thank you.
The Chair: Lord Grocott has a supplementary.
Q111 Lord Grocott: With the perspective of your background and experience, would you like to reflect for a moment or two for our benefit on some of the language being used at the moment about AI wiping out the human race within a measurable period of time? Another thing that springs to mind is this. I was on a committee looking at AI in defence, and the phrase “killer robots” came up from time to time in the evidence. Where are you on that spectrum? I will make it as specific as this. Should we be really worried, mildly worried, or at the stage of academic reflection? If it is “really worried”, how imminent is it?
Isabella Wilkinson: I am happy to jump in. In my estimation, the hypermilitarisation and anthropomorphisation—sorry, that word is a bit of a mouthful—is incredibly unhelpful and inappropriate to the moment that we find ourselves in. Relying on this sort of language and this sort of discourse not only obfuscates governance solutions but it has a polarising impact. Let us walk through some of these impacts and consider different framings that may help us reimagine the moment that we are in and try to unblock some of the gridlock that we have diagnosed currently in the session today.
Overmilitarisation feeds into zero-sum delusions. We have seen the regurgitation of Cold War and arms race framings around the US-China race for technological supremacy. Indeed, it may be a race for technological supremacy, but it assumes that attaining artificial general intelligence or artificial superintelligence, which for the most part is an unmeasured benchmark, means that the game is up and there is nothing else to do. It is much more appropriate to reframe to words that we have discussed in the session today: a patchwork, the existence of multiple paradigms and multiple pathways, pushing back against the assumptions that some sort of technological supremacy and endpoint is inevitable, and that we are on a determined or predetermined pathway to that point. It is inappropriate also because it leads us into the assumption that the world is divided into Cold War-esque blocs—the US versus China, US technology and infrastructure versus Chinese technology and infrastructure. Again, it makes sense that people tend to divide into these sorts of blocs when they are talking about how to govern, develop, deploy and diffuse, but doing so obfuscates the existence and the reality of alternative and third pathways: the development, say, of open source models in the global majority, UK and European aspirations for building and developing technology in the national interest.
This leads us into a powerful reimagining, pushing us away from catastrophising, anthropomorphising, militarising, and towards something rooted in openness and plurality. It may be more helpful to reimagine the development of AI capabilities as a sort of moonshot. It may be helpful to bake in the acceptance of plurality to conversations about governance, pushing back against an inevitability narrative and reclaiming some agency. As we have seen over the past week, reclaiming agency and sobriety in discourse is as important in discourse as it is in the development and deployment of the technology itself. I would refer the committee to a really interesting anthology on reimagining the AI arms race that pitches a variety of different framings.
The Chair: Thank you very much. Mr Daws, do you want to answer the question?
Sam Daws: That is a fantastic answer. I am extremely worried but I am also optimistic. The risks of the ability for bad actors to develop bioweapons, which is one of the scariest examples, are very real. The loss of control issue of AI misalignment is very genuine, and we are seeing evidence of it. There is a genuine willingness to address it across different stakeholders. There are solutions.
I agree also on the narrative. Often, Cold War narratives only go so far when it comes to AI. It is a very different phenomenon from nuclear. If the risks are down stream, it means that every country in the world needs to be engaged if we are going to solve it. It is a bit like terrorism; you cannot have a space in Somalia, Sudan, Yemen and so on where terrorists are able to regroup. You need every Government to be empowered around the world to address the possible fallout from AI. The narrative around the Cold War and the arms race can be self-fulfilling. It can encourage secrecy. It can weaken safeguards. It can help create the risk that it predicts. A better framing is strategic competition with shared security obligations; that would be my approach.
There are some great examples from the Cold War and since of what we have done when there is not trust between parties: confidence-building measures developed in the arms control, cyber, bio, chemical and nuclear weapons areas. The UN Institute for Disarmament Research has been doing some good work on which of those is transferable in the AI domain and which is not. Confidence-building measures—red telephones, as it were—are good.
There was an encouraging part of President Xi Jinping’s speech at the World AI Conference. It is the first time that a leader of a major country has articulated the need for prevention, for addressing loss of control and for emergency management co-operation. We are seeing some signs. President Trump in his last UN General Assembly speech talked about the need for international co-operation on the nexus of AI and bio risks. We are seeing an agenda where there is willingness as long as the concerns of all sides, as long as this does not impede innovation, are respected.
Lord Grocott: If there is no agreement on these complex issues, what is the worst-case scenario?
Lord Ahmad of Wimbledon: Pack your bags.
Sam Daws: It is a great question. Nothing is inevitable. What we have seen with the development of AI is that it is jagged. As with some models, they cannot tell you how many r’s are in the word “raspberry”, but they can solve Olympiad-level Olympic problems. I doubt that we can ever say that AI has reached artificial general intelligence or artificial superintelligence. I do not think there will be a moment when we know that one country has complete dominance in all domains—security, economic and so on—over another. That uncertainty is one of the reasons these analogies around arms control and “winner takes all” break down. The worst-case scenario is pretty grim, but there are plenty of other challenges that we know that are difficult, wicked problems, such as climate change, that the world is grappling with, and we may or may not solve them. AI is not top of my list in terms of that existential risk, but it is something that we need to take seriously.
Q112 The Chair: The whole nuclear weapons side of things is surrounded by mystique, secrecy and protocols, as well as certain cost factors, that have given the international organisations some degree of traction, but that does not apply to chemical and biological weapons where there is largely an unregulated space. Why should you be optimistic that you can control artificial intelligence in a way that we have not even been able to do with chemical and biological?
Sam Daws: It has been very difficult partly because there is a huge amount of economic value committed to the success of AI. Companies, as we have seen in the last few days, are waking up to the fact that safety is essential if they are not going to lose the economic value that they put into it. If the public, Governments and parliaments lose trust in AI, that is very difficult. I think the incentives are better aligned.
You made a great point about nuclear, chemical and biological weapons. There is a disparity between the nuclear and chemical regime. You have the non-proliferation treaty for nuclear and the Organisation for the Prohibition of Chemical Weapons for chemical. They are large organisations with large secretariats. The Biological Weapons Convention, by contrast, has only four full-time staff members and 10 secondments. It is tiny in comparison. There is a real need. The UK, with its leadership in life sciences and Porton Down, and looking at the risks from biosecurity and biosafety, has a real contribution to make here in developing support for the Biological Weapons Convention, specifically because AI both increases the positive benefits in the life sciences and makes it far more likely that biological weapons may be produced by bad actors. We could talk later about it perhaps, but there are some interesting proposals from Kazakhstan on creating an international agency for biosecurity. Existing work is being done to set up a scientific committee within the BWC. All those are important. The incentives are there.
Q113 Baroness Prashar: Thank you so much for your very helpful answers so far. My question is specifically about the UK. We have heard from a number of witnesses that the UK possesses the necessary capability to effectively shape the multilateral conversations on the future of AI. Do you agree with that assessment? If you do, what are the core capabilities and attributes that the UK brings to the table? Are the Government making the best use of it? If not, how can they make the best use of it?
Isabella Wilkinson: I have full confidence in the UK’s capacity and agency to shape multilateral AI governance, but the UK Government and their supportive ecosystem must tread carefully and act strategically, and use this window of opportunity, visibility and salience to advance agreement and consensus on priority areas, which I anticipate we will get to later in the session.
In my approximation, the UK has three USPs, three unique selling points, when it comes to its engagement on AI multilateralism. The first is its world-leading technical expertise. The UK’s AI Security Institute and its supportive research and innovation ecosystem is, undoubtedly, world leading, and it has a high level of trust and authority. This is a huge part of the UK’s offering to potential partners and allies when it comes to AI multilateralism. We cannot forget the supportive ecosystem that enables the curation and platforming of this technical expertise. I am thinking here of innovation hubs and start-ups in Oxford and Cambridge as well as the Advanced Research and Invention Agency of UKRI.
The second component of the UK’s USP on AI multilateralism is its diplomatic heritage. I am pleased to hear that Bletchley has come up time and time again in our session today. Bringing together representatives from the US and China to talk about AI security and safety is a massive win. The UK has continued to engage, as Mr Daws has pointed out, in innovative minilateralism in the years that followed.
The third area where the UK has, I would argue, a unique selling point is in its partnerships. The UK boasts a strong alliance with the US on AI and a degree of alignment in the way aspects of models are tested, although there is a little bit of controversy here with Anthropic’s decision not to let the AI Security Institute test its latest model, which I would be pleased to discuss later in the session. None the less, the UK-US partnership on technology and AI, in addition to the UK’s role as a proactive middle power engaged in innovative minilateralism working as a sort of bridge with Europe, is a key strength.
Sam Daws: I agree with all of that. There are real examples where the UK has led the world in science diplomacy. The acculturation of antimicrobial resistance was a double act between Dame Sally Davies, the Chief Medical Officer at the time, and Dame Karen Pierce when she was ambassador in Geneva. They did the real legwork of going regionally to Africa, the Asia-Pacific, Latin America and so on, and getting agreements in veterinary science, animal health and agriculture through the different UN bodies and then taking it to the UN General Assembly. We have real experience of how you get buy-in from around the world for crucial issues. We have huge expertise in AI safety from the AI Security Institute and so on. Biosafety, as I mentioned before, is an area where we have particular added salience, allied with our huge university expertise in both life sciences and AI governance.
If I was to be critical, because it is also important to look at where the gaps might be, in my experience of speaking at a lot of summits in different parts of the world, the UK has not been joined up between the Foreign, Commonwealth and Development Office and the former DSIT, the Department for Science, Innovation and Technology. Often, you have DSIT representatives who do not have any experience in multilateral engagement, and they are therefore quite fearful of engaging with the Chinese because they are worried about what they can and cannot say, whether it is a security issue and so on. We need to bring the Foreign Office and Cabinet Office expertise and comfort in dealing with very sensitive issues to the multilateral offer. I have seen really encouraging signs with the new BIST that has replaced DSIT and with the Cabinet Office now overseeing the AI Security Institute, which allows an all-government approach. The omens are very good that we can have that joined-up approach, but it is absolutely essential that the Cabinet Office steers this and ensures that we have a credible offer, not just on the AI safety side but on the AI development side, not just because China is really leading on that but because developing countries and countries in Central Asia and other parts of the world do not see the world through the AI safety lens. They think, “Well, if it’s mainly the problems caused by companies from the US and China, what’s our role in that?” You have to engage people where they are, and from much of the world that is around issues of AI sovereignty and how they can manage dependencies across the full stack of AI.
The Chair: Lord Bruce has a question that in many ways follows on from that.
Q114 Lord Bruce of Bennachie: You have answered a lot of the questions, but I have a couple of things I would like to say. In the July GDP report, it unexpectedly rose by 0.4%, and that was attributed to productivity improvements delivered by AI. That may be just a one-off, but, if it is the first green shoot, it is potentially a very strong one if that was to continue.
I want to link two things. One is the domestic importance of AI and how the Government prioritise it linked to our international engagement. You have indicated that you think we have capacity, but we have dismantled our aid budget and the personnel deployment that goes with it. We are reducing our diplomatic service in numbers. It is a bit of a bee in my bonnet as to whether we need to redeploy our diplomatic capacity, particularly the overseas part of it, in terms of looking at emerging economies and what we, with our AI capacity, can do that is beneficial to them and us, and looking at our like-minded allies, and, there, where we can jointly develop both safety systems and mutually beneficial technologies. How do we use our diplomatic capacity to achieve that? Given it is changing, how much does it need to change?
Isabella Wilkinson: Mr Daws, you are most likely, due to your career, better placed than me to comment on this one.
Sam Daws: Okay, I will try. It is a really great question. I suppose it is part of that whole “What’s our offer on development?” For a long time, ODA has been overtaken by foreign direct investment and remittances in terms of the volume of assistance. It is not really about aid any more. We have a tremendous in-kind contribution that we can make. One avenue is our AI Security Institute partnering with the only AI security and safety institute in Africa, which at the moment is in Kenya— partnering with Kenya, partnering perhaps with Singapore and with other African countries, and spreading out across the African Union. There are very specific ways in which we can work with institutions. There, it is probably best done in partnership with the different UN agencies that have national capability programmes. The UN Development Programme, UNESCO, the ITU and UNIDO are the four UN agencies that are all engaging at a national level in developing countries. That is a key way to do it.
You are right that AI will, I hope, continue to drive productivity. In the most optimistic example of that, it will completely change the nature of the debate between the developed and the developing countries—things that are holding back agreement on climate change and other things—by bringing a huge number of resources into developing countries through efficiencies, developing country priorities, health, education, the delivery of digital public services, smart cities and so on. All these can see an absolute sea change. They do not need frontier AI; they can use free, open-source models. They do not need the latest technology to achieve a lot of these economic efficiencies. We need to prepare for a world—and this goes to the heart of the Foreign Office restructuring—where we may see a huge beneficial uplift in the resources available to all countries through the adoption of AI and other technologies. We have talked a lot about gloom, but there is a possible, very positive economic picture.
I have done a lot with Central Asia and Kazakhstan. I will return there in a few weeks’ time. It has an AI Minister, an international advisory panel for the president on this area, and it is revolutionising its whole country in ways that the UK cannot imagine. It is really focusing on that. Before the Iran War, the path of the UAE and Saudi Arabia to moving beyond oil and gas was through AI. Below the surface, it is not necessarily visible on these conversations around the risks and China and the US. There is a whole lot happening on AI at the less frontier level, which is transforming the world click by click.
Isabella Wilkinson: Perhaps I could jump in to zoom into two tools in the UK’s diplomatic toolbox on AI, drawing on Mr Daws’s fantastic intervention. The first relates to a theme that I mentioned earlier in this session, which is the unique role the UK can play in building capacity. Capacity building has come up time and time again in global governance on AI debates. It is clearly a priority for countries in the global majority. It is front and centre of China’s World AI Cooperation Organization. The UK can play a leading international role in building capacity in many of the ways Mr Daws detailed.
A specific way it can do so speaks to an earlier question raised about preparing for a multitude of different scenarios. In the AI governance space, we have seen the emergence of really exciting and innovative future AI scenarios that bring together key policymakers from across sectors and borders to give them the tools, awareness and networks that they need to deal with anything from an AI-driven disruption in shipping and trade to an AI-driven national security crisis. The UK can work with partners, with other middle powers and with countries in the global majority to develop these sorts of scenarios and this training to not only improve preparedness but ensure that countries facing disruption have the regional and international diplomacy that they need, that familiarity, those networks, as we have discussed before, in order to navigate it on the one hand and build governance in conditions of crisis and turbulence on the other. A specific tool that the UK can use in this emerging diplomatic toolkit on AI looks at scenario-based capacity building. I would be happy to refer the committee to future work and emerging best practice on this point.
The second thing that I would draw your attention to is standards. Standards are a relatively dry area of global AI governance, but they are incredibly important. They pertain to the best practice and the standardisation of how models are tested and evaluated, how the security of open-source models and the complex supply chains that feed into them is guaranteed. The UK can play a leading international role and broker agreements surrounding a democratic approach to the development and deployment of AI standards. I would absolutely push the committee to centre this area as a relatively important tool in this emerging toolkit that we have discussed just now.
Lord Bruce of Bennachie: That is very helpful. I am going to ask for a comment. We have the capacity, but maybe we need to produce a more coherent, joined-up approach to our external relations. Is that essentially what you are saying? We can do it, but we probably need to be clearer about what our offer is to the wider world.
Isabella Wilkinson: Absolutely. There are multilateral windows of opportunity to do so, namely the UK’s G20 presidency, which would be great to discuss later in the session.
Q115 Lord Lamont of Lerwick: We talked about the patchwork pattern that there is. Are we going to see more convergence and agreement among select groupings as opposed to large-scale agreement via the UN? We had given to us the Brookings Institution paper that talked about the need for a decentralised network approach rather than a centralised approach.
Isabella Wilkinson: Absolutely. It is likely to see convergence in these smaller groupings on the one hand but, from a global bird’s eye perspective, a degree of fragmentation. This is expected. We have seen this in other governance areas. It is a theme that we have discussed earlier in the session today: the coalescing of different coalitions and agreements based on shared interests, shared values, as Mr Daws covered. The guiding theme of sovereignty is front and centre of many of these minilateral agreements and declarations. On the one hand, fragmentation in any area of governance is expected and really should be welcomed. There is no silver bullet in global AI governance. We should not all be governing AI in exactly the same way. There are important efforts led by the UN to improve interoperability—the weaving together between different regional approaches or issue-based approaches to global AI governance.
However, there are areas where fragmentation is harmful. We have seen over the past few weeks declarations coming out of the Shanghai Cooperation Organisation, the World AI Cooperation Organization, the US-led Pax Silica and various regional groupings. On the one hand, it may be relatively convenient to see these as competitor groupings and absolutely at loggerheads with each other, but on the other it is important to accept the regional multitude of different approaches and seek to build that connective tissue between them.
I would refer the committee absolutely to a recent report by the UN’s recently founded AI Governance for Humanity Lab, which mapped out the state of interoperability—again, a systematic approach to weaving together the overlap points or the junctures between these different governance approaches to ensure that harm is reduced, that there is maximum coverage, that nothing is slipping through the cracks, and to ensure that patchwork governance is more than the sum of its parts. As I mentioned before, many of these regional agreements—for example, China’s WAICO or the Bishkek Declaration coming out of the Shanghai Cooperation Organisation—are incredibly nascent, and so it is worth watching this space to see where they develop, gauging fragmentation risk, and then, where appropriate, seeking to build interoperability between either competitor or loggerheads approaches to AI governance.
Sam Daws: That is a great answer. In the area of technical standards, that is something on a global level through the ITU, the International Organization for Standardization and the more technical IEC and IEEE in particular. The British Standards Institution can play a really important role. One of the areas for UK influence is through standards. That sets a benchmark. We also have to engage with CEN and CENELEC, the EU standards organisations. There is a risk that, because what the EU AI Act determines becomes law for those countries, global standards are affected by the Brussels effect and effectively adopt CEN and CENELEC standards as international standards. Therefore, BSI and others and individual UK companies need to be engaged in these European standard-setting processes as well as the global processes.
You make a really good point in that you have these patchwork and fragmented minilateral groupings. There does need to be interoperability. A useful report that the committee may want to look at was produced by a new UN university body in Valencia, Spain, on interoperability across all sorts of different dimensions, from culture to safety and security, and that is part of weaving these together.
We need to think about how China is engaged with some of the western-focused organisations in ways that do not harm western or UK competitiveness or security. When the OECD merged with the Global Partnership on AI, it brought in countries like India and Brazil, but China still sits outside of that. China is a partner with the OECD in certain areas but not in AI. This could be a real opportunity with the UK’s G20 presidency to find a way to connect the nascent Chinese standards being projected through their minilaterals with the western ones, and, secondly, Chinese engagement with the global network of AI Safety Institutes, which has been renamed NAAIMES. China has attended alongside some of the meetings of that network in terms of individual academics rocking up, but they have not been formally involved. We need trusted dialogues between the Chinese network of AI safety and security and development institutes and western AI safety and security institutes. That cannot come soon enough.
Q116 Lord Alderdice: Sam, you very kindly participated in an event that I organised in this very room last Thursday on the governance of AI, and in particular, because of the involvement of the secretary-general of the IPU, on the role of parliamentarians in developing guardrails for AI. I wonder if I might ask both of you to address this. Which areas of AI governance and the agenda do you see as most likely to produce some kind of agreement? In which areas will we find that the prospect for consensus is very limited? If we are going to focus on things that we can deliver on, which are more likely to be successful and which are more limited?
Isabella Wilkinson: It is helpful in embarking on this answer to think of agreement and consensus as a spectrum. On the one hand, we have voluntary coalescing around shared principles and norms, non-binding, consensus driven, but fundamentally lacking teeth. This is important, none the less. On the other side, we can see hard regulation, binding rules. We are seeing likely spotlights of success on any range of agreements on any parts of the spectrum at the moment. Much of this is due to the highly salient policy window that we find ourselves in.
Let us take a step back and consider how change happens when it comes to AI multilateralism. As we have discussed before, as we have covered in our deep dives on different tools and mechanisms that the UK is involved in, we see the development of consensus best practice at a small level, maybe just a technical forum or an international network or minilateral grouping, which becomes platformed and strengthened, ideally, at a regional multilateral level. Where we see potential agreement likely, again, this might be more on this side of the spectrum—the voluntary normative side; it may, indeed, be on an element of catastrophic risk. We have seen calls for international binding rules on artificial superintelligence re-upped over the past week.
However, I would draw the committee’s attention to low-hanging fruit areas of global AI governance that have been garnering a lot of attention and coalition building over the past few months and, indeed, the past few years. These are issue areas where countries, despite differences in their political systems and despite competition dynamics, can agree it is a shared priority. One of these is looking at child protection and child safety from the risks generated by the use, adoption and diffusion of AI systems. Another may be around deepfakes, which is a prevailing concern for countries regardless of their political system, and gets us into interesting territory on what constitutes trusted, accurate information with integrity in the AI age.
Another low-hanging fruit area is on capacity building. We are seeing a great deal of global consensus on best practice for capacity building and the prioritisation of it in a variety of different minilateral and regional groupings—perhaps short of a global agreement, but a growing global normative recognition of its importance to closing this ever-elusive global AI gap between the haves and the have-nots.
Another area that is ripe for agreement is something that I would call local models: the development of potentially open-source models trained on local data and local languages aligned with the local or national interest. This is where I would gauge that agreement in the next few months could be likely, noting that said agreement might look slightly different. We may not have binding rules, but we may have the crucial, small-scale consensus and normative alignment, which can then be platformed to something concrete and with teeth. This is a vital process.
Let us look now at areas where agreement is unlikely. We have covered this, so I am not going to spend too long on it. These are areas laden with competition and national values that may be at loggerheads with each other. We have seen the difficulty of brokering minilateral or even international agreement on the use of AI in the military domain despite the emergence of many norms and red lines about unacceptable use at a smaller level. Earlier, I mentioned deepfakes as a subset of risks in the information environment and disinformation. Regulating disinformation at an international level is seen as a priority, but it is an area laden with such values and such loggerheads between political systems that we have seen long-term inertia.
In short, we can characterise areas in AI governance where agreement is unlikely if they bring us into an information-sharing and strategic advantage dilemma, where the disclosure of information or the relinquishing of agency is totally misaligned with the national interest. Mr Daws, do you have anything to add?
Sam Daws: One key area where progress has been made, and it will continue, is technical standards. We have talked about that.
Going back to Ms Wilkinson’s division between malicious use, accidents and systemic risks, on the malicious use, there could be progress on AI-enabled biological risks. Safeguards around autonomous weapons are much harder to negotiate. The UN has been negotiating for 10 years on this issue, on what are sometimes colloquially called killer robots, and that is partly because they are actively involved in the Ukraine-Russia war and other areas. Iran is very opposed to these things. In the disarmament negotiations, it is basically based on consensus. You can have spoilers that prevent progress.
In the accidental category, there is real opportunity now for shared evaluation approaches, shared risk management frameworks and so on. That has to be multidisciplinary, and it has to be multi-partnership involving the private sector to be successful.
The last one is AI for science. There is a huge benefit for humanity to harness the discoveries that will be cascading our way in advances in pure mathematics, medicine and material science. We have a duty to ensure that geopolitics interferes as little as possible in the scientific exchange, which can then revolutionise our lives. That is the real prize. It is often forgotten that there is a cost to mistrust between the West and China, and that cost is shared scientific co-operation.
Lord Alderdice: Thank you very much.
Q117 Baroness Fraser of Craigmaddie: This committee, as you know, is looking at multilateral organisations and their role in AI governance, but I want to focus on the role of the private sector organisations, which both of you have touched on. You have said that we might be at a time where there are shared incentives for the private sector, partly to keep the trust of the public and of Governments, because, if they do not, the vast economic benefits that they have garnered to date will be at risk. Interestingly, we have seen recently OpenAI, Anthropic and I think Google DeepMind coming together to try to convene around international standards, which, Mr Daws, you have just highlighted as an opportunity, but their view of a standards body is not necessarily the view of other organisations. It goes back to Lord Ahmad’s comment earlier that everybody agrees something should be done but it is proving jolly difficult to agree what that something should be and whose version we should take.
In the world of regulation, we can point to failures in regulation of private companies on online safety. X did not come to the Secretary of State’s meeting yesterday. Copyright is another one, as well as LLMs and scraping. We have failed to implement our copyright standards. You mentioned Anthropic. Is there a mechanism, whether it is the Frontier Model Forum or something else, for private sector involvement? If we agree that these organisations should have a formal role in global governance debates, how do we ensure that the right ones are coming to the table and that action follows agreement?
Isabella Wilkinson: It is a fantastic question. Thank you so much, Baroness Fraser. We have been grappling with the role of the private sector in global governance for decades. We have had scholars such as Susan Strange in the early 1990s talking about the retreat of the state and global governance or international relations formerly as the playground of sovereign states being challenged by these incredibly powerful multinational corporations. Many of the anxieties that you note in your question are ones that we have been facing in other areas of governance for a long time. Over the past few years, we are encountering unprecedented market power of these new AI companies and older technology companies, with disproportionate influence over capability. The frontier is wholeheartedly privatised. There is disproportionate influence over discourse as well. What constitutes superintelligence? What constitutes things like recursive self-improvement? What do we know about the frontier and how models are performing along it?
Very interestingly and excitingly, we have seen disproportionate influence over governance mechanisms, many of which on a technical and policy level are incubated in a corporate setting and then developed and platformed. For a long time, we have had a binary division between regulation and approaches to governance developed by states only and regulation developed in a private setting. This division absolutely has to break down, because what we are seeing is a great deal of hybridity, in that approaches to, say, documenting what models can do—model cards—are developed by the developers of those models themselves; best practice is developed among, say, companies and the technical community; and then eventually you will see some of this best practice reflected in regulation. That is to say there is a huge influence over the global governance of AI, noting that, if you govern a model, of course it has cross-border and cross-domain impacts. None of this governance is stuck to one policy area, just one model or just one jurisdiction.
We have seen private sector crowding when it comes to many of the new governance proposals swirling around over the past summer. Dario Amodei’s essay on pacing the frontier, which many of you will be familiar with, calls for an independent third-party evaluator network. We have seen Demis Hassabis of Google DeepMind calling for a US-led AI standards agency. Sam Altman and Elon Musk have thrown their weight behind Dario Amodei’s proposals as well. On the other hand, we have seen Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg coming out in opposition to these calls for regulation.
The puzzle that we find ourselves in is what to do here. Perhaps the private sector does not need a formalised role in global AI governance. I would argue they are already there. They are already shaping the debate, mediating the discourse, and creating many of the rooms that these critical decisions about the governance of technology are taking place in. The puzzle that we face and the critical question that we need to answer and on which Governments need coherent approaches is how to ensure accountability and maximise governance agency when coming into confrontation or perhaps partnership with these new concentrations of corporate power at the global level. There are ways of doing so.
My bottom line here would be that the global governance of AI does need to be designed with the private sector in the room. There is very little that we can do without it, but it cannot be a relationship based simply on voluntary agreements, trust and good will. There has to be a degree of agency and reciprocity. We see the co-development of trusted governance of high-trust technical instruments in different areas of global AI governance. We know we can do this, but it is a matter of incubating. It is also a matter for the UK to consider where, co-ordinating with other middle powers that maybe do not have access to the most powerful frontier capabilities, and which are encountering these incredibly powerful AI companies, they might be able to co-ordinate collectively on things like access and potentially on soft regulation.
I would love us to return to the point and the proposal for setting up an independent network of third-party evaluators. These are independent evaluators who would go into AI companies and test the models to see different safety and security risks that can be mitigated before they are released. The UK can play a central role in any of these governance proposals. We have the technical expertise and diplomatic capital to do so. It could be a fantastic route to approach this critical relationship with the private sector and global AI governance with the two principles that I mentioned earlier: accountability and agency.
Baroness Fraser of Craigmaddie: Can I quickly come back to you on that last point? It is all very well to say, “Let’s send people in to look at the safety aspects of a model before it is released”, but is there not also an issue that the companies that are developing it are saying, “It was safe when it left me, but you are deploying it in a way that you have chosen to”? It is not just as simple as that.
Isabella Wilkinson: Absolutely; thank you so much for raising that point. Testing and evaluation is by no means a silver bullet. A lot of the time, the environment or the part of the economy or the community where a model is deployed will bring up a variety of different risks that are really difficult to test for at the testing and evaluation stage. This is why it is so important to promote independence, integrity, and the development of best practice in how evaluation is done. Evaluation has to take place in AI companies, but we need a variety of other mechanisms to test, for example, model impacts once it has been released and then identify a risk threshold or a notification threshold if model behaviour is causing X amount of harm. This is often called technical intervention or a kill switch, which I would be happy to explain further or refer the committee to some technical work on. I would underline that, in order to test and evaluate the safety of models pre-deployment and after they have been released, you need a supportive, independent ecosystem driven by agency and accountability. The UK can build this ecosystem, but it has to act incredibly decisively and has a window of opportunity for doing so.
Sam Daws: Baroness Fraser, taking your last point first, it really points to the need for clear attribution of responsibility among different parts of the whole supply chain. One example is open-source models. If they are modified and harm is caused, is the original open-source developer responsible or the company that did the modification? How do you prove that?
Agentic AI has the same thing. Having agent identification and agent tracking is really important so that you know where an agent is at any particular time and what it has done. The problem with a kill switch—although it is a great idea and it may be necessary—is that an agent may be embedded in very sensitive technology at that time. You need a kill switch, but you also need the ability to row back permissions over time, taking it back to when it did not start acting in a non-aligned or misaligned way.
To answer your original question, the traditional way that western companies have engaged in multilateral processes has largely been voluntary through the Frontier Model Forum, the G7 Hiroshima AI process and through the standards bodies that we have talked about. It has been exchanging best practice, and it has all been valuable.
They have also made commitments as a result of the Bletchley summit to the follow-up to Bletchley, which is Seoul. They made commitments for the Seoul summit. Chinese companies in that case also did. Chinese companies are not part of the Frontier Model Forum and they are not part of the other processes except for the standards ones. It was one of the few areas where they could also put forward voluntary commitments. It could be a coincidence, but the unfortunate happenstance was that the Chinese companies that put their head above the parapet and made commitments that were in the Seoul framework were then hit by export controls by the US. I have been to China three times in the last nine months to talk to AI experts in academia and elsewhere. They feel that they are walking on very sensitive ground, wanting to accord with what the Chinese Government want and need and what the international community wants. It is quite difficult for those private actors.
When we talk about the private sector, given that Chinese capabilities on the frontier are between six months and two years behind, according to different estimates, they are going to be very close to where we are, so they need to be involved. The Chinese Government already place certain obligations on them: an algorithmic registry, new legislation of deepfakes, companion bots and so on. In some ways, Chinese companies are more regulated in some of the safety demands than us. If we want interoperable approaches, we really need to engage with the Chinese private sector as well.
Q118 Baroness Blackstone: Do you think that the global governance of AI will continue to take place under existing organisations—from what you have told us, there seem to be a lot of them—or do you see the possibility or indeed the desirability of some new, either one or more, international organisations to take this on?
Isabella Wilkinson: My advice is that we must at all costs avoid reinventing the wheel. We have valuable architectures already taking root in the global AI governance landscape that can be better resourced, can be strengthened and can improve their mandates, but I would for the time being advise against a wholly new institution. An example could be, as Mr Daws mentioned earlier, strengthening the International Network for Advanced AI Measurement, Evaluation and Science—NAAIMES, if I am getting the acronym correct. That was the renaming of the International Network of AI Safety Institutes. This is a critical network for setting the science around AI safety and developing best practice on testing and evaluations. If its mandate was strengthened, if it was better resourced, if it could improve its global membership, this could be a brilliant place for co-ordinating information sharing in the event of a disruption or even for standardising approaches to model risk monitoring.
We have seen in civil society the development of spaces for dialogue or technical forums, which again are really valuable architectures and could be bolstered if we wanted to action some of the governance proposals we have been discussing today. One of the leading scientists in technical AI governance and a Chatham House associate fellow, Stuart Russell, runs the International Association for Safe and Ethical AI, which is a critical technical gathering, setting best practice on different technical governance mechanisms and then scaling this internationally. We have seen different Track 1.5 efforts and scientist-to-scientist dialogues in the AI safety space, and of course the successor to the Bletchley summit, the Geneva AI Summit, taking place in June next year.
I mention all these different architectures to say there are concrete opportunities to better resource them. There are strategic opportunities to improve interoperability between them and to set up permanent mechanisms. When I was in Geneva at the UN Global Dialogue, there were some fascinating proposals to give the Geneva AI Summit a permanent secretariat that can track progress and implementation around some of the voluntary commitments that Mr Daws mentioned, for example on frontier company AI safety commitments.
It is worth learning from what has worked well in the past and collecting a patchwork of governance analogues from other areas. We have mentioned the environment, which many folks borrow from when they think about the potential future role of an international scientific panel on AI setting out the state of risk and how different technical specialists are mitigating and tracking them, and developing some sort of information sharing in order to set that global best practice.
I would also draw the committee’s attention to mechanisms such as the Financial Stability Board, the scenarios for future planning that the World Health Organization does, and the International Atomic Energy Agency. Again, these are not perfect analogues. They should not be copy-and- pasted into the complexity and tempo of the global AI governance landscape. I would reiterate that, as we are seeking not to reinvent the wheel and thinking about promising governance architectures that we can strengthen and better resource, we have a history guidebook of lessons not only for governance mechanisms that might work but many governance mechanisms that have been developed in times of great crisis and turbulence. For policy recommendations on exactly how to do so, I would refer you to a Chatham House report Breaking the Deadlock on AI Governance, which sets out how to draw from historical case studies to hopefully bolster some feasible global AI governance mechanisms.
The Chair: Thank you for that plug of a Chatham House document. As a distinguished fellow of Chatham House, I can appreciate that. Mr Daws.
Sam Daws: AI is particularly difficult institutionally to house because it is so far-reaching and general purpose. It reaches in a way that interconnects issues in ways that humans eventually will not even be able to understand because AI will be able to work out the interconnections better than we can as it gets more intelligent. It is because of human nature or certainly political nature that we tend to be quite siloed in our approach to things. Governments have a Ministry of Health and a Ministry of Defence and so on. That is reflected in the international architecture at the moment with different UN bodies, NATO and so on.
The most likely direction of travel, one where there are the most incentives—because every UN and other entity tries to increase its mandate and its resources, which is basically in its DNA—is that for work displacement the International Labour Organization will take the lead, for the health aspects of AI the World Health Organization will take the lead, and so on. The challenge at the moment is that the international organisations are coming under huge pressure to deliver more for less in the future because donors’ money is also limited. The next Secretary-General, who will be appointed sometime between today and 31 December of this year, will have to decide how to embrace AI as a policy domain, whether, as with Guterres, it is quite central, what kind of architecture they want to recommend, and how much political capital they want to put into setting up new systems.
For the bits of AI that do not fall neatly into a health basket or a security basket and so on, there have been various attempts, as Ms Wilkinson said, to develop an IAEA for AI that would have enforcement powers to require states to do something. That was considered by the UN advisory board on AI, and it decided not to put it forward. There is no appetite at the moment for a UN body with enforcement powers to tell countries or companies what to do, for understandable reasons. So we have a bit of a challenge. We may end up with a body like that in a few years’ time, but the precondition for that would be a major incident, a major cyberattack, that was so significantly economically disruptive that there was a willingness to relinquish a bit of sovereignty in order to set up such a body. We are not there yet and heaven forbid that the actual crisis is a lot worse than a simple economic disruption, but it points to the real difficulties in designing the architecture.
My one guidance for the committee in considering it is: do not start from an organigram of what looks sensible on paper. Multilateral reform is around aligning very different priorities. UN reform means so many different things to each country. It is not around an ideal academic solution. It is a political compromise around conflicting priorities. AI has many different conflicting priorities within it that make the formulation of such a new institute difficult. It probably will need to give something to everybody. As with the IAEA, it prevents the proliferation of nuclear weapons. In return, it supports the peaceful development of nuclear technology. I could see a new agency in the future monitoring the risks from AI in real time but at the same time supporting developing countries with their AI development. In areas like cybersecurity, for geopolitical reasons the priority should be for the UK to work with our close security partners such as NATO, G7 countries and so on. There are some things where the global framing can be interoperable, but you need to work with close allies. There are different configurations.
Baroness Blackstone: Thank you very much. I have a very quick follow-up. You have said, “Don’t reinvent the wheel”, although you have also come up with one possible longer-term solution to getting better global governance. I am a bit stuck between “Don’t reinvent the wheel. Keep going as we are”, yet there is a gridlock in global governance, and you keep using that phrase. So, somehow or other, we have to find our way through those two positions.
Sam Daws: We need to be ambitious and say, “Look, this would be the ideal”, and see if we can push it, and if the UK can help garner some agreement. It will be incremental in nature. It will be built up through, first, private sector voluntary agreements. Regulation is already having some effect: for example, the state laws in California passed in recent weeks and the Brussels effect of the EU AI Act, and in particular now the general purpose AI provisions, which have come into force while the other EU AI Act provisions have been postponed because of the AI Omnibus initiative of the EU. You are seeing companies beginning to bring in safety controls or at least safety declarations and transparency that meet both California’s requirements and the EU’s requirements. We will not see a perfect evolution but different regulatory steps. It was quite noticeable that the Vice-President of the US struck a conciliatory chord recently in some of his remarks in being open to certain forms of regulation of AI. It is very much a fast-moving situation and the UK should be as ambitious as it can be in terms of the optimal situation and have a clear pathway to get there.
The Chair: Thank you very much. Lord De Mauley has the next question, but he has to leave, so you have to be very brief in your reply to his question.
Q119 Lord De Mauley: China has cropped up a number of times in the discussion today. You mentioned earlier the World Artificial Intelligence Cooperation Organization. What a snappy name. How does its approach differ from those of western-led organisations? What are the implications for us and our allies?
Isabella Wilkinson: I defer to you given the time you have spent in China.
Sam Daws: Okay. The Chinese approach is a focus on sovereignty. It is a focus on control. It is a focus on development. The World AI Cooperation Organization seeks to be an international version of what is happening at home. If you look at China, there is a fifth five-year plan, and its own AI Plus initiative, which is about implementing AI into every economic and social sphere in China. Its offering to the world is a low-cost, open-source, full-stack approach that reflects the local context, culture and languages of a country. The implications for the West are that China is continuing with its strategy of wanting interoperable standards worldwide that reflect China’s own domestic intentions. The West has been poor in giving a similar offer to the world. The answer is not to be scared of China but to reflect a similar offering to developing countries that reflects the specific western approach to open societies, democratic values and so on. The UK could even consider joining the World AI Cooperation Organization depending on the formulation and then influencing it from the inside. We need to be nimble in our approach to these things and not fear everything China does as being purely a China strategic play. Partly it will be, and partly it will be a genuine desire to improve the lot of the developing world. When we meet with Chinese representatives, it is really important to acknowledge the good intent behind a lot of Chinese measures, while also recognising the difference in approach and the need to be very clear about where our red lines lie in security and in other areas.
Isabella Wilkinson: The UK can play a really important role in supporting interoperability between anything that is agreed and platformed using China’s WAICO and other regional or even multilateral initiatives. The place to do so is absolutely the United Nations. You have seen it take bold strides, pitching the UN Global Dialogue as a dialogue of dialogues, a hub of different partnerships and initiatives. It is in the UK’s strategic interest to support interoperability on the standards front—it is a valuable way of doing so—and in diplomatic engagements, and to ensure that the strategic engagement is also connected to the UN, being that critical connective tissue.
Q120 The Chair: My final question was going to be, “What do you think this committee’s recommendation, since it is reporting to the Government, should be?”, but you have basically answered that question now. Is that right?
Isabella Wilkinson: I am happy to reiterate two recommendations that should be front and centre for the committee. The first reiterates points that I have raised earlier on security and on supporting an independent evaluator ecosystem, leveraging UK technical authority on testing and evaluations. The second relates to UK diplomatic engagement. The UK, as we have covered, has incredibly strong convening power, evidenced by Bletchley and by its heritage on science and technology diplomacy, and it has opportunities ahead to really generate minimum multilateral consensus on priority areas in AI. The G20 presidency could be a great opportunity to do so, connecting the safety and security imperative to economic outcomes, but the UK should not wait for the guise of the G20 presidency to push for agreement on priority areas. I would draw the committee’s attention to, potentially, biosecurity, as Mr Daws has outlined.
Sam Daws: My answer would be very similar. I am really glad that we are very aligned. The UK has a huge offer, a huge contribution, to make in AI safety, AI security and AI for science, and, diplomatically, really perfecting our different engagement strategies. We need a warm relationship with the US with the aim of ensuring that US frontier models are accessible to the UK AI Security Institute.
We need regulatory alignment—not identical—with the EU to the extent that it will allow UK businesses to prosper and to further our objectives and differ where that is of advantage.
We need engagement with China on a whole series of tracks, particularly between our AI Safety Institute and its security institute and its equivalent. As a possible bridge between China and the West more broadly, we occupy a uniquely beneficial position outside the EU, with good connections to the US and with excellent connections across the Commonwealth. The last area for diplomatic engagement is our offer to the rest of the world. There, we need tailored plans for Central Asia, Southeast Asia, Latin America, the Gulf and Africa. There will be slightly different contributions that we can make. These are not things that will cost a lot of money. There is a tremendous amount of expertise in the wider UK ecosystem that we can contribute to, ensuring that AI helps humanity to flourish as well as addresses the risks and the challenges that we have talked about today.
The Chair: Thank you so much to both of you. We are very grateful for all the insights that you have brought here. We are all much better educated on this subject than we were at the beginning of the session, so we are very grateful to you for that. I declare that this public session is now closed.
