The Global AI Governance Standoff: A Prisoner’s Dilemma?

As artificial intelligence (AI) grows more capable and widespread, the absence of global AI governance becomes harder to ignore. For policymakers, addressing AI governance requires coordinating multiple stakeholders – industry representatives, consumer associations, and academia – for a comprehensive understanding, while ensuring regulations keep up with the rapid pace of AI evolution. Before examining U.S.–China–EU competition, understanding why global AI governance is essential. 

Global AI governance is necessary for various reasons, all stemming from the transnational nature of AI. First, AI development and supporting technologies – such as microchips, graphic processing units (GPUs), data centres, big data – rely on globalised supply chains and firms. Second, AI-related risks cross borders, as malicious actors such as terrorists, criminals, or hacker groups can use AI to conduct cyber-attacks. This concern prevented Anthropic, a major AI company, from widely releasing its latest Mythos model, claiming that many institutions remain unprepared for cyber attacks enabled by this exceptionally powerful model. Other international AI threats include job displacement, pollution, and water consumption. Third, the AI race itself poses problems: China’s AI rise demonstrates that states with less regulation gain a competitive advantage, creating incentives to undercut rules in the name of winning the AI race. Some Chinese authors emphasised this issue of postponing regulations to avoid falling behind, calling it an ‘Ethical scissors gap.’ This final point reveals one side of the prisoner's dilemma central to this debate. While international governance could mitigate fears of falling behind and increase global benefits by addressing these issues, each player has short-term individual incentives to defect from commitment.

The role of the U.S.

Currently, three actors play a significant role in shaping the future of AI governance. The U.S. is the most advanced country in the field and the one that started the AI boom. One of the most groundbreaking papers on AI, ‘Attention Is All You Need,’ was published in Long Beach, California in 2017, and the country is home to the most important players in the field, such as Nvidia, Google, Anthropic, OpenAI, and Meta. The U.S. approach to AI regulation and governance emphasises both private-sector initiative and avoiding inflexible over-regulation. More specifically, it relies on a multi-stakeholder, decentralised model with enforcement left to sector-specific federal agencies, such as the Food and Drug Administration (FDA) or the Securities and Exchange Commission (SEC). The American approach essentially distinguishes between specific uses of AI, rather than establishing an overarching framework based on universal principles to determine what constitutes lawful use of AI. 

At an international level, the U.S. strategy centres on three pillars. The first is ensuring the export of American AI to its allies, preventing them from turning to rivals through technological lock-in. The second involves countering China by closing loopholes in semiconductor export controls and countering Beijing’s influence in international governance bodies, such as the United Nations. The third concerns opposing burdensome regulations emerging from international governance bodies that do not align with American values. Beyond bilateral and multilateral diplomacy, the U.S. AI vision also dominates in organisations such as the Organisation for Economic Co-operation and Development (OECD) or the Group of Seven (G7), which over-represent U.S.-aligned countries. The OECD AI principles, for instance, revolve around trustworthiness and human rights, closely mirroring U.S. values. 

At the latest G7 summit, American AI CEOs sat as if they were heads of nation-states. On this occasion, Dario Amodei from Anthropic, Sam Altman from OpenAI and Demis Hassabis from Google DeepMind all agreed on the need for a U.S.-led cooperation initiative on AI between democratic countries. While this seems to reinforce Washington’s position, this corporation-led attempt can also be seen as a way to limit the influence of the U.S. government itself over these companies. For instance, the U.S. recently suspended Claude Fable 5, one of Anthropic’s frontier models, to foreign nationals over security concerns. This sparked concerns not only among countries reliant on U.S. AI providers, but also among the firms themselves. In conclusion, it is still early to say whether the influence of the multitrillion-dollar U.S. AI companies will end up helping Washington’s efforts, or whether it will evolve into a set of new, non-state, AI governance leadership contenders.

The role of China

China, the most prominent challenger to U.S. AI dominance, saw DeepSeek release its ‘R1’ reasoning LLM in 2025 – a model matching U.S. performance built at a fraction of the cost. In June 2026, a different Chinese lab, Z.ai, released GLM 5.2, an open weights model that once again narrowed the gap with GPT 5.5 and other U.S. models that had pulled ahead. Academically, China surpassed the U.S. as the country with the most first authors in AI-related research in 2026. At present, Chinese AI efforts are directed mostly towards practical applications, such as manufacturing and healthcare, while the U.S. still leads in the race towards artificial general intelligence (AGI). Domestically, China uses a centralised approach and state-led initiatives to support AI progress. Regulations follow a vertical model targeting specific issues, such as recommendation engines, rather than a horizontal framework adapted to specific cases. 

In terms of global governance, China has issued a series of proposals, such as the Global AI Governance Action Plan or the AI+ International Cooperation Initiative. These diplomatic tools define the vocabulary Beijing seeks to use in international AI debates, centring on terms such as ‘sovereignty,’ ‘people-centred approach,’ ‘inclusiveness,’ and ‘sustainability.’ 

The last two words have already proved politically divisive: at the 2025 Paris AI Action Summit, the U.S. and the UK refused to sign the final declaration over concerns that language on “inclusive and sustainable AI” lacked practical ambiguity, national-security priorities, and the risk of over-regulation. Since 2024, China has built a coalition within the UN called the Group of Friends, gathering 80 countries from the Global South to push initiatives aligned with these principles. While forums such as the G7, the OECD, and BRICS+ largely reflect the preferences of advanced market democracies and the Global South respectively, the UN offers a more universal arena in which both sides can seek global legitimacy.

The role of the EU

The case of the EU is peculiar: compared with China and the U.S., Europe is objectively behind in AI technical development. Nevertheless, the EU can still leverage its large consumer market to impose rules and global standards unilaterally, as foreign companies must comply to access it – a dynamic known as the Brussels Effect. The EU focuses primarily on ethics and consumer protection, with values broadly aligned with the U.S.’s vision of fostering AI trustworthiness and accountability. One key milestone in this regard was the release of the AI Act, a comprehensive regulatory framework about AI. The Act mandates disclosures aimed at making AI safe and transparent, with specific requirements for different levels of risk. 

Beyond regulation, the EU launched the AI Continent Action Plan to increase European AI readiness by investing in AI-related infrastructure, such as AI factories and data centres, as well as human capital. Despite the EU’s broadly shared commitment to liberal values and trustworthiness, some differences with the U.S. do emerge. The EU's stated commitment to trust and transparency – described by the AI Continent Action Plan as distinctive and rooted in EU values rather than merely liberal ones – is stronger and less compromised than that of the U.S., which consistently subordinates these principles for the sake of innovation. America’s AI Action Plan, for instance, mentions that ‘the United States needs to innovate faster … and dismantle unnecessary regulatory barriers that hinder the private sector in doing so.’ Nevertheless, it remains too early to determine whether this approach will outlast the current presidency. 

The EU regulatory approach in terms of data and AI is surprisingly similar to China’s. Both frameworks invoke transparency, proportionate data usage by firms and state organs, and mitigation of bias. The key differences lie less in these technical principles than in their political and institutional framing. China emphasises cooperation by all “organizations and citizens” with national intelligence efforts and national security organs, while AI systems must ‘uphold Core Socialist Values’ as a baseline, preventing instability and threats to national security. In contrast, the EU uses its rights-based ideology as an upper limit. Thus, even when national security justifies restrictions on data rights, such measures must remain compatible with fundamental rights and freedoms. Additionally, data protection in the EU is subject to independent supervision, while in China, oversight is enforced by the state cybersecurity and informatisation department.

Is the dilemma solvable?

Global AI governance continues to receive widespread attention among governments, with countless analyses and proposals. However, because the main actors pursue incompatible regulatory models, compete for leadership, distrust each other’s intentions, and operate through a fragmented regime complex, cooperation has so far remained weak. The prevailing zero-sum logic, whereby a state either wins the technological race and imposes its own rules or falls behind, creates a powerful disincentive to cooperate, even where the stakes are clearly shared.

In the past, prisoner’s dilemmas concerned with technology did produce global standards, but not because states converged around shared values; they did so because interoperability, trade, safety, or mutual vulnerability made fragmentation too costly. AI follows a similar pattern, but with a larger obstacle: its standards are not merely technical but political, encoding fundamental choices about the values of the parties involved. Two earlier cases illustrate that humanity has already navigated technology-related prisoner’s dilemmas without one side achieving total victory.

The ozone crisis illustrates the economic dimension. States initially had incentives to keep using cheap chlorofluorocarbons (CFCs) to avoid losing competitiveness, but cooperation became possible once the issue was narrowed to specific substances, scientific evidence became undeniable, poorer countries received time and support to adapt, and trade rules made staying outside the agreement costly. Restraint was transformed from a unilateral sacrifice into a shared industrial transition.

The nuclear case illustrates the sovereignty and security dimension: nuclear technology promised civilian benefits but also existential risks during superpower rivalry. States did not overcome competition through trust, but through an imperfect bargain. Some nuclear hierarchy was tolerated, peaceful nuclear access was preserved, and inspections reduced uncertainty enough to make restraint compatible with national security.

Neither case maps perfectly onto AI, which combines both dynamics at once. Still, both show that competing states can move beyond zero-sum logic when cooperation is framed not as surrender, but as a way to reduce mutual vulnerability while preserving national advantage. 

Conclusion and recommendations for the EU

Given that the EU is already struggling with technological sovereignty, planning for a prolonged zero-sum AI leadership competition is the wisest approach. The EU should therefore design its future AI regulations to push the global race towards terrain more favourable to challengers, while maintaining its transparent and ethical image.

Open-source AI models could help the EU catch up technologically, as they lower the barriers to training, adapting, distilling, localising, and deploying AI systems. However, open-source AI is not a clearly defined category. The Open Source Initiative (OSI) defines open-source AI models as free to use, study, modify, and share across all their components. Many so-called open-source models are only open-weight, releasing model parameters, but not training data, code, or the model architecture. Mistral is a good European example, described as releasing model weights without the used data or the training process. The company itself claims several of its models are “open-weight” and available under permissive licences. 

More broadly, having clear, shared terminology on AI is itself a necessary condition for functional global governance. For actors aspiring to lead on global AI governance, being the first to secure a universal terminology agreement – one that does not exist yet – will be an important milestone. 

On the question of open-source AI, China has already demonstrated how open-source models can favour a challenger in the AI race. In its 2025 Action Plan on Global Governance of AI, Beijing included the creation of a transnational open-source community and an open-source compliance system among the plan’s objectives. Supporting this direction could serve the EU well, as it aligns with both its practical objective of strengthening European AI and its ethics-oriented governance ethos. It would also help prevent the global AI ecosystem from hardening into two blocs, with risk-averse democracies locked into American platforms and revisionist or non-aligned states turning to China for technological autonomy from the U.S.

However, defining open AI too strictly – requiring broad disclosure of training data and training procedures – risks creating compliance burdens that hurt European challengers more than global incumbents. Dataset disclosure, for instance, can expose firms to copyright litigation. Even where AI training might eventually be found lawful, litigation costs can function as an entry barrier, putting European firms at a competitive disadvantage against foreign players that build scale in less-regulated home markets before having to meet EU standards. The AI Act already requires a soft disclosure of training data: a sufficiently detailed but not technically exhaustive summary, covering data categories and major datasets used, without requiring individual files.

The challenge lies in balancing sufficient openness to let EU AI catch up against the risk of excessive disclosure requirements. The EU should embrace openness more than the AI Act currently does, under which open models are treated more as a regulatory exception than a strategic catch-up tool. If correctly implemented, this approach could allow the EU to set a credible governance standard for open AI: more trustworthy than China’s, given its pluralist rather than state-curated regulatory process, whilst reducing its dependence on American platforms. 

The views expressed in this article belong to the author(s) alone and do not necessarily reflect those of European Guanxi.

ABOUT THE AUTHOR

Giacomo Savarese 司马睿 is a Data & AI professional at Accenture Song and a Bocconi graduate specialising in Big Data and Machine Learning. After working at Alibaba.com South Europe on SME digitalisation and automation, his research has focused on the intersection of AI, technology adoption, and China–Europe relations. He is particularly interested in how governance frameworks can make emerging technologies both more trustworthy and more widely adopted.

This article was edited by Nina Thinnes and Stefano Bertoli.

Featured Image: A smartphone displaying the DeepSeek AI chat interface / Matheus Bertelli / Pexels / Creative Commons Attribution 4.0 International / Free for use