The CERN nuclear physics laboratory was founded in 1954 to pool expertise and resources for fundamental research, in service of the common good, on a scale beyond the means of any single country. Might a similar model work for AI governance today, enabling its huge challenges to be tackled in an environment of depoliticized cooperation?
A moment for governance
The question of how to ensure artificial intelligence (AI) systems operate safely, ethically and legally is a challenging one. Risks and harms can originate and proliferate at different stages of an AI system’s life cycle. If poorly designed or hastily deployed, AI systems can cause a range of harms to people and society. Job displacement, exacerbation of societal inequalities, and amplification of toxic content or racial stereotypes are all potential outcomes. Risks can also arise from misuse or catastrophic accidents, from AI systems malfunctioning and causing injury, and from poor supply-chain practices.
Recent advances in generative AI systems and ‘foundation models’ (AI models capable of a wide range of tasks) have exacerbated concerns about these risks. Powerful AI capabilities such as text or image generation are more readily accessible to everyday users. Among policymakers, civil society organizations and industry practitioners, advances in AI have created a sense of urgency about the need to govern these new technologies effectively.
Local and national governments around the world are grappling in different ways with the challenge of governing AI. In the US, several cities have passed laws prohibiting or restricting the use of facial recognition systems. At a regional level, the European Union has just finalized its AI Act, a comprehensive product safety law that will affect many AI systems. These distinct measures address a common challenge: how to govern a suite of technologies that affect different sectors, involve complex supply chains, operate across borders and can raise a wide variety of risks. Leading AI labs like Anthropic, Google DeepMind, Microsoft Research and OpenAI have joined calls for a coordinated international effort to improve safety.
A ‘CERN for AI’
Efforts such as the UK’s AI Safety Summit, and the follow-up AI Seoul Summit, have aimed to start an international discussion around AI safety and have led to the creation of national AI safety institutes in the UK, the US, Japan and Canada. There is a shared ambition to create an international AI safety network, and these institutes have begun to sign bilateral cooperation agreements. However, there is not currently a single, coordinated global institution seeking to promote or research safer AI. Recent research has explored how different existing models for international governance might be applied to AI, including whether institutions such as the Intergovernmental Panel on Climate Change (IPCC) or the International Atomic Energy Agency (IAEA) have features that could be borrowed or adapted for use in this emerging field.
One increasingly prominent proposal, circulating among some academics and policymakers, advocates the creation of an international coalition for AI research inspired by the scale and collaborative spirit of CERN.
One increasingly prominent proposal, circulating among some academics and policymakers, advocates the creation of an international coalition for AI research inspired by the scale and collaborative spirit of CERN, the European Organization for Nuclear Research. After the devastation of the Second World War, there was an effort to rebuild European science. Leading scientists, including the Danish nuclear physicist Niels Bohr, lobbied European governments to establish an international laboratory devoted to particle physics. In 1954, CERN was established under the umbrella of UNESCO. CERN is publicly funded by 23 member states (22 European states plus Israel).
CERN is both an international institution and a laboratory. It is famous for discoveries like the Higgs boson and for being the birthplace of the World Wide Web. Its primary function has been to provide the powerful and prohibitively expensive infrastructure and hardware – most notably the Large Hadron Collider – needed to conduct particle physics research. CERN has historically focused on fundamental science rather than on developing technical standards or benchmarks. Its significant resources have enabled it to fund and broker multinational research collaborations. In addition, CERN has been an explicit source of inspiration in the institutional design of organizations in molecular biology and astronomy. For these reasons, some people have argued that a similar institution could tackle the complex challenges of AI safety.
CERN’s broader legacy has been in enabling nations to work together on expanding scientific knowledge. Constructing its particle accelerators and detectors required member states to pool expertise and funding. This allowed them to achieve a scale and depth of scientific work that no single country could have reached alone. CERN represents the post-war ideal of science beyond borders in the service of discovery and peace.
If a CERN-like organization for AI were to exist, with the function of coordinating international AI safety research, it would require several features. First, as with the actual CERN’s particle accelerators and supercomputers, a CERN-like body for AI would need its own technical infrastructure to support computational research. This could include physical infrastructure such as data centres, high-performance computing resources, networking systems, and laboratory facilities tailored to AI work.
Social and organizational infrastructure would be needed to provide operational support, including to manage relationships with commercial labs and nation states, make platforms available for open and innovative research communication, and secure sustainable funding for international research networks and collaborations. A CERN-like body might also foster more interdisciplinary and international research collaboration on AI risks, enabling greater involvement of researchers from countries that are otherwise lacking in computational resources.
If this new organization were to study the safety of frontier AI models, it would also require privileged, structured access to state-of-the-art AI models from industry labs, and access to underlying ‘training’ datasets – used to train AI systems in different capabilities – and other critical materials relating to each model’s design and operation. Leading labs, including OpenAI, Google DeepMind and Anthropic, have already made voluntary commitments to open their models to select researchers for the purposes of safety and independent evaluations, though it remains unclear how meaningful these commitments will be unless underpinned by hard regulatory requirements.
A CERN for AI might be in a unique position to broker access to cutting-edge AI systems, allowing researchers to test and compare the safety, biases and robustness of models beyond what any single lab could achieve independently.
However, a CERN for AI might be in a unique position to broker access to cutting-edge AI systems, allowing researchers to test and compare the safety, biases and robustness of models beyond what any single lab could achieve independently. The convention that underpins CERN grants it status as an intergovernmental organization, with the privileges and immunities that come with that, and provides for direct contributions from governments; all this insulates it from political pressures in a way not possible for even national labs.
Strengths of a CERN for AI
Proponents of a CERN-like body for AI have called for its creation as a way to build safer AI systems, enable more international coordination in AI development, and reduce dependencies on private industry labs for the development of safe and ethical AI systems. Rather than creating its own AI systems, some argue, a CERN-like institution could focus specifically on research into AI safety.
Some advocates, such as computer scientist Gary Marcus, also argue that the CERN model could help advance AI safety research beyond the capacity of any one firm or nation. The new institution could bring together top talent under a mission grounded in principles of scientific openness, adherence to a pluralist view of human values (such as the collective goals of the UN’s 2030 Agenda for Sustainable Development), and responsible innovation. Similar sentiments have been repeated by other prominent actors in the AI governance ecosystem, including Ian Hogarth, chair of the UK’s AI Safety Institute, who argues that an international research institution offers a way to ensure safer AI research in a controlled and centralized environment without being driven by profit motive.
Proponents of a CERN-like model also argue that such an institution could provide vital global public goods for AI safety, which profit-driven private companies might otherwise undersupply. Such goods could include: benchmarks to evaluate model robustness; auditing tools to increase accountability; and datasets to assess harmful biases. Providing all of this would require sector-wide collaboration between governments and AI companies. Crucially, this work could include research into an expanded definition and operationalization of ‘AI safety’ that would cover the full scale of harms that AI systems can cause; such research could be informed by a deliberative process involving a representative sample of humanity, not just commercial labs or academics.
Some proponents of a CERN for AI believe that it may also reduce dependency on private labs for AI safety research, and attract top researchers interested in pursuing projects of greater public benefit rather than those with purely commercial potential. Professor Holger Hoos has described a potential CERN for AI as a ‘beacon’ to ‘attract talent from all over the world’. This could create an alternative hub of expertise outside the private sector.
The existence of a different type of AI institution could provide academics and students with an alternative career option to joining big tech firms. It could also help address asymmetries in political power between industry and academic labs. Currently, significant power in AI development accrues disproportionately to a handful of private labs. A publicly funded international research organization conducting safety research might be more resilient than private sector labs to economic pressures, and better able to avoid the risk of profit-seeking motives overriding meaningful research into AI safety measures.
Hurdles faced by a CERN for AI
Long timelines and cost overruns often plague ambitious big science collaborations. Physics breakthroughs have required enormous hardware investments over years. For example, to build CERN’s Large Hadron Collider, over 10,000 scientists and engineers from hundreds of universities and labs contributed to its design and construction over a decade.
But while current computer clusters for AI research have yet to require such large workforces, constructing data centres and network infrastructure at scale for a new institute will still take time, investment, and reliable access to currently undersupplied specialized chips for AI development. That said, the modular nature of graphics processing units (GPUs) and servers could allow for much faster scaling up of AI infrastructure than has been feasible in previous science megaprojects.
Challenges in AI safety also differ from those of particle physics, so addressing them may require more dynamic, distributed initiatives. For example, CERN itself primarily focuses on pure science. However, AI safety is as much a question of values, ethics and societal impacts as it is a matter of AI systems’ technical capabilities. Focusing on purely technical evaluations of an AI model’s performance can only reveal so much information about its use and potential outcomes. It would thus be essential to ensure that critical perspectives on the impacts and implications of AI are incorporated from the outset into any new institution’s culture and mission. But even this may not be enough: some risks of AI will only become apparent when a particular application is deployed, and may prove challenging for a CERN-like body to address.
In other words, it is possible that a CERN for AI could address only a subset of the risks that AI systems pose. This is due to the vast range of challenges that AI systems can present to actors in multiple domains. Focusing only on technical, model-level fixes such as better learning from human feedback, for instance, could prove a distraction from other essential governance efforts, such as regulation, accountability and public engagement, all of which are also necessary for identifying and mitigating risks from AI systems. Care would need to be taken to involve diverse stakeholders, and to balance capabilities against controls. Inflated expectations for AI governance via a CERN-like model could backfire if they are not realistic about such an organization’s inherent limitations.
Another hurdle could be the issue of information asymmetry between the private sector and any new institution. Given its likely focus on safer systems and providing public goods, as discussed above, rather than purely pushing forward AI capabilities, the new institute would be unlikely to control the most capable AI systems itself. It would therefore be dependent on information-sharing from commercial labs to understand those systems (which will depend on proprietary data, model design and engineering insights, which commercial labs will want to keep to themselves), absent any information-sharing obligations placed upon them (e.g. as is mandatory to meet safety requirements in the aerospace industry). Even if internal developments in commercial AI labs (such as safety concerns) are published openly, there will be a delay between discovery and those findings being shared more widely and acted on. Being ‘behind the curve’ in terms of understanding and having access to the most capable systems may also make working in public bodies less attractive for some. To mitigate this risk, an incentive structure would need to be established that can compete with private industry to attract and retain researchers.
There are also worries that creating a CERN for AI may result in safety researchers working in less close proximity to leading commercial AI labs, thus reducing the ability of such researchers to monitor risks on the ground. It may be that the best safety research is conducted alongside cutting-edge AI research in the private sector, as this could enable a deeper understanding of the systems and processes of the labs involved.
There are worries that creating a CERN for AI may result in safety researchers working in less close proximity to leading commercial AI labs, thus reducing the ability of such researchers to monitor risks on the ground.
These issues also raise the concern that a new CERN for AI could be influenced or captured by big tech firms. To date, the research carried out by such firms has far outpaced public sector capabilities, with the result that major tech companies currently hold disproportionate power by virtue of their resources, expertise and leverage. Preventing narrowly commercial interests from dominating a CERN for AI would require vigilant governance.
That said, the governance structure of CERN could provide a template for its AI-focused equivalent: CERN’s multinational membership and interdisciplinary focus insulate it from capture by special interests, and provide a diversity of input to counter corporate influence. CERN is run by a council of its member states, with two delegates each (one representing government, the other national scientific interests); each member state has a single vote, and the council operates on a simple majority vote for decision-making. This also ensures no single member state can abuse its position within CERN – and provides a measure of protection against risks associated with the actions of individual states, as seen in the council’s suspension of Russia’s scientific observer status in March 2022 after Russia’s full-scale invasion of Ukraine.
Researchers have also raised concerns that giving a centralized institution access to the advanced AI models of leading labs might compromise the security of those labs and models. For example, effective access to design evaluations and benchmarks may require the ability to copy a given model, which could undermine the commercial interests of those labs and enable diffusion of those models before adequate testing. This may be less of an issue for mechanistic interpretability and similar research, which may not require access to the latest models.
Lastly, a CERN for AI would have to grapple with rising geopolitical tensions. It is arguably harder today to start an international governance body than it was in the era immediately after the Second World War. Most leading AI labs are based in the US and China, two countries that are arguably engaged in a ‘new cold war’ that is fuelling a technological arms race between them.
A path forward
A CERN-like institution would not be a replacement for comprehensive national and local regulation and governance frameworks, which would need to address broader challenges such as harmful misuses of AI systems. What is interesting about the proposal, though, is the potential that a new international body could complement the creation of other international governance organizations and instruments, including standards-setting bodies, certification bodies, treaties and domestic legal frameworks.
There is no perfect analogue for AI when it comes to governance, and as future AI safety summits approach, policymakers should evaluate proposals for new international institutions and consider what these can accomplish, building on the efforts of the already established AI safety institutes. A CERN for AI undoubtedly represents one credible possible model to advance targeted elements of AI safety research and provide public alternatives to private sector dominance. With ample resources and global collaboration, it could make valuable technical contributions.
However, we cannot and should not expect one governance model to address the full span of risks and harms from AI systems. It may be that institutions such as the International Civil Aviation Organization, IAEA or IPCC provide better models for solutions to international AI governance. We cannot overlook the risk that a CERN for AI may turn out to be too expensive, too cumbersome, or simply unnecessary. More research is needed to flesh out what the realistic objectives of this kind of institution might be, how it might work, and what kinds of challenges it will be best placed to solve. The path forward rests on collective insight, courage and care in steering AI’s immense potential towards the common good.
