Rapid geopolitical change, institutional weakness and asymmetries between the public and private sectors have combined to make cooperation on AI near impossible.
International cooperation on AI is not only technically, legally and practically difficult, but faces structural barriers that make proactive, coordinated governance unlikely. A substantive and durable system of governance for AI may instead emerge only in response to a crisis situation, when the political costs of inaction clearly exceed the costs of coordination.
2.1 Geopolitical deadlock
As the dominant powers in AI, the US and China have an outsized influence on its development and regulation. The US hosts many of the world’s largest AI and technology companies, and therefore has no incentive to constrain its AI industry through international agreements when it can regulate domestically if it so chooses. It currently prefers shielding industry from any regulation at all. China, meanwhile, cannot accept verification provisions that require third-party access to confidential model weights and training processes. These factors make enforceable agreements politically untenable for both countries at present.
Intense competition between these two powers to develop and deploy frontier technological capabilities has brought discussions on global AI governance to a standstill. The dynamics of this race for supremacy are preventing the emergence of institutions and forums for international cooperation on shared risks, and do little to slow the development of potentially destabilizing AI models and systems.
In AI development, China and the US are far ahead of the rest of the world, implementing a pro-innovation agenda in the hope of securing economic and security advantages. The US outspends China, and together they outspend the rest of the world. China’s AI investment was set to grow by 48 per cent, to $98 billion, in 2025, while reported private AI investment in the US reached $110 billion.
The key goal of this competition is the promise of artificial general intelligence (AGI) or ‘superintelligence’ that would give the victor a national strategic advantage that its rivals could not match – representing both a highly desirable prize and an existential risk should the race be lost.
Calls from the AI safety community to slow down or impose moratoriums on AI development are frequently challenged on these grounds. Political leaders are instead moving in the opposite direction, with the US administration’s AI Action Plan looking to shield industry from hard regulation and President Donald Trump seeking to prevent attempts by individual states to regulate AI.
Intense competition between China and the US to develop and deploy frontier technological capabilities has brought discussions on global AI governance to a standstill.
Other centres of global influence on technology regulation are also in retreat – most notably the EU, where industry pressure, fears over a US backlash, and an anxiousness not to stifle much needed growth and innovation have all contributed to a shift away from safety and towards ‘competitiveness’. The new UK government pledged AI regulation, but has instead focused on securing US investment and a business environment more closely aligned with the US.
Developing nations largely lack domestic frontier AI capabilities and depend on technology transfer from the US or China. Rather than pursuing independent regulatory frameworks, many Global South countries prioritize securing access to AI infrastructure, training data and computational resources. Anxiety among the so-called middle powers over their access to frontier technology and their future participation in a global economy increasingly shaped by AI has reduced their appetite to pursue strict governance.
The costs of AI governance failure are still poorly understood. Despite (and, in part, because of) the enormous excitement around its potential, AI is still a small part of the global economy and a peripheral concern to decision-makers. Unlike the geopolitical deadlock, however, the level of attention does seem set to change in the next 24 months. AI will eventually affect almost 40 per cent of jobs around the world, according to the International Monetary Fund. Even the most conservative estimates, such as those of Nobel laureate Daron Acemoglu, credit AI with an increase in total factor productivity of around 0.5 per cent in the next 10 years.
2.2 Weakened institutions
International institutions lack enforcement power and member states disagree radically on how important and strategic AI will be, leading to different levels of seriousness, budgets and willingness to make trade-offs. The diplomatic channels and foreign policy platforms that would need to negotiate AI agreements remain largely unaware of AI’s implications.
Recognizing the scale and unpredictability of shared risks, many stakeholders have turned to supranational political bodies, multilateral institutions or regional forums – including the UN, the Organisation for Economic Co-operation and Development (OECD), the EU, ASEAN, the G20 and the G77 – to develop shared approaches and guidance. To date, no such institution has delivered substantial progress on global AI governance. Nor do any yet appear well-suited to navigating and mitigating a global AI crisis.
Most international institutions lack the enforcement mechanisms required to constrain state and industry AI powers. Despite considerable UN efforts like the establishment of the Independent International Scientific Panel on AI and Global Dialogue on AI Governance, the organization generally cannot supersede the political will of sovereign nations on issues relating to technology. The speed and technical capacity required to keep pace with technological advancements is equally absent, while the gap is deepened by spending cuts. Finally, the UN’s consensus agreement process is also in danger of negotiating on AI to the lowest common denominator, a criticism levelled at the Summit of the Future and Global Digital Compact. While they play an important role in signalling willingness for dialogue and enabling further cooperation, lowest common denominator negotiations risk watering down content for the sake of broad-based consensus.
The US withdrawal from various international organizations and agreements – including climate accords, human rights systems and the World Health Organization (WHO) – has undermined the existing institutions of international governance. Similarly, the US rollback of international aid and intergovernmental support signals a move away from multilateral visions and goals, including the UN Sustainable Development Goals, while China continues to strengthen its parallel multilateral systems.
Though Brussels has demonstrated its ambition to regulate emerging technology through the EU AI Act, its global strategy has failed to launch and is failing to garner sustained support outside the EU. First, with little semiconductor manufacturing, few frontier AI companies and limited talent, European policymakers are criticized for over-regulating while not having ‘skin in the game’. Second, the so-called ‘Brussels Effect’ – referring to the EU’s ability to set global regulatory standards that others follow – has yet to materialize in AI. Many potential emulators of EU regulatory approaches have turned elsewhere: either away from hard regulation altogether, towards setting priorities in other regional bodies, or to bilateral relationships with the US and/or China.
In a fragmented global order, the gravitational pull of investment is stronger than that of comprehensive policy. Brussels has moved towards a softer approach to technology governance, like the General-Purpose AI Code of Practice. Most major AI companies have agreed to adhere to these guidelines, which align expectations around how to comply with regulatory obligations on transparency, safety and security. But ultimately, the code is a highly contested and non-binding instrument that is consequently vulnerable to geopolitical pressure.
2.3 Asymmetry between public and private sectors
Private companies largely shape the development and deployment of frontier AI systems. Companies like Google and OpenAI build the models, NVIDIA and TSMC manufacture the chips that power them, and cloud providers like AWS, Google and Microsoft supply the infrastructure. These companies are overwhelmingly concentrated in the US and China.
Public–private partnerships are becoming more common. Notable examples include the US’s ‘Genesis’ and ‘Stargate’ investments at home and abroad or the US–UK Technology Prosperity Deal. But these arrangements have not shifted the fundamental reality. Outside of China’s state-directed AI labs, the most consequential decisions on AI development happen inside private companies, in opaque and largely unregulated environments.
Private companies have commercial incentives to push back against costly governance provisions like transparency requirements, capability limitations or verification processes, particularly when competition rewards secrecy and speed as richly as it does in AI. Regulatory capacity gaps mean that, even if they wanted to impose constraints on the private companies developing AI, governments lack the computational resources, technical expertise and legal authority to independently evaluate proprietary models or compel disclosure.
The scale of private investment far exceeds regulatory capacity. Industry estimates put 2026 hyperscaler capital spending at $527 billion globally, while OpenAI’s CEO projected that future frontier models could require $100 billion in capital per training run. By contrast, the EU’s AI Act, finalized in 2024 after years of negotiation, allocated just €1 billion for its enforcement and implementation. When the UK government announced its AI Safety Institute (since renamed the AI Security Institute) in 2023, it committed £100 million over two years – which, despite being a significant investment, is less than major private sector labs spend in a single week. The digital industry’s spend on lobbying in Brussels alone is reported to have increased by over 50 per cent in the four years up to 2025, reaching $175 million in that year.
This investment gap has resulted in limited capacity for effective regulation. Regulatory agencies largely lack the computational resources to independently evaluate frontier model capabilities, relying on developer self-reporting or voluntary access agreements. Agencies can neither hire sufficient technical staff to match industry expertise, nor compel wholesale disclosure of training data, model architectures or safety testing results – all of which are considered by private companies to be proprietary information.
Governments have in recent years largely prioritized investment in AI over regulation. For example, when the US administration of President Joe Biden released its AI executive order in October 2023, the emphasis was placed on fostering innovation and attracting private investment. The order imposed minimal binding constraints on model development. Biden’s successor as president, Trump, has taken further steps to minimize regulatory burdens and maximize investment. Meanwhile, France and Germany successfully lobbied to weaken the EU AI Act’s requirements for foundation models, citing concerns over competitiveness.
The result is an environment where the information, resources and technical capacity of the regulated is vastly superior to that of the regulators.
2.4 Knowledge collapse and threats to shared values
Effective governance of shared global problems depends on the capacity to establish facts, resolve disputes about evidence, and build consensus on when action is warranted, and is made easier through shared values and principles. The latter are both are in decline.
Nations are increasingly unable to agree on fundamental questions because they operate in incompatible information environments and hold increasingly divergent values.
Trust in media, government and scientific institutions has declined sharply across democracies over the past two decades. Media fragmentation means individuals consume fundamentally different information from one another. Regulatory agencies face systematic underfunding and political interference, weakening their ability to produce credible independent assessments. Scientific expertise is increasingly treated as partisan rather than authoritative. The result is an absence of shared processes for resolving disagreement. When stakeholders cannot even agree on what counts as credible evidence, or which institutions have legitimate authority to assess such evidence, collective decision-making inevitably breaks down.
Efforts to establish AI governance are floundering in this deteriorating information environment, while facing additional structural challenges. AI model interpretability is itself an evolving field of science. Technical evaluations of model behaviour remain largely under the ownership of private companies, limiting independent verification outside of a handful of AI security institutes. Developers face strong commercial incentives to withhold information about system capabilities, and strong security incentives to avoid disclosing vulnerabilities.
International agreements cannot proceed without shared basic assessments of risk. Enforcement becomes impossible when compliance itself is continually contested. Support for intervention from the public and policymakers alike evaporates without sufficient consensus about what AI systems can do or should be used for. AI governance requires trusted mechanisms for risk assessment, transparent information-sharing, and the capacity to build consensus rapidly when threats emerge. But none of these things are achievable when the basic infrastructure for establishing shared facts is weakened.
Nations are increasingly unable to agree on fundamental questions because they operate in incompatible information environments and hold increasingly divergent values. The traditional transatlantic alliance between the US and Europe, once rooted in shared democratic principles, is being reconstituted as a ‘coalition of capabilities’, based on transactional exchanges of economic benefits rather than common purpose – the so-called ‘Pax Silica’.
The US-led ‘Pax Silica’ explicitly frames cooperation around ‘shared interests’ in supply-chain security and economic advantage, rather than democratic solidarity or multilateral principles. This change in emphasis reflects a broader shift in which ideological alignment between governments is replaced by capability-based coalitions. Some AI scholars such as Anton Leicht contend that this type of agreement may eventually form the basis for a stronger coalition than one based on shared values, but the transition will be uncomfortable and is not certain.
Taken together, these barriers create an environment where binding international coordination appears difficult, if not impossible, under current conditions. The US and China prioritize national advantage over cooperation; middle powers pursue access over regulation; international and state institutions lack enforcement power and technical capacity; private investment dwarfs regulatory budgets; and allies and rivals disagree over basic facts, while abandoning previous shared values and common purpose.
