AI Governance Research as the Proving Ground for U.S.–China Cooperation

1990 Institute Prize Runner-Up

Within hours of taking office on Jan. 20, 2025, President Donald Trump revoked Executive Order 14110, the Biden administration’s centerpiece framework for artificial intelligence safety. [1] Two weeks later, at the Paris AI Action Summit, the United States and the United Kingdom refused to sign the joint declaration on AI governance. [2] By June, the U.S. AI Safety Institute had been renamed the Center for AI Standards and Innovation. [3] By December, even the international coalition of AI safety institutes established after the first major international AI summit in 2023 had dropped the word “safety” from its name entirely. [4] In the span of a single year, Washington’s posture on AI shifted decisively from governance to deregulation, from shared risk management to technology-stack competition.

Yet in the same week that Washington was pulling back from Paris, Chinese researchers from Tsinghua University, the Beijing Academy of Artificial Intelligence, and the Chinese Academy of Sciences launched the China AI Development and Safety Research Network. [5] And on March 10, 2026, researchers from the U.S.-based Safe AI Forum sat down with Chinese scholars at Fudan University to compare domestic regulatory mechanisms for frontier AI and discuss international coordination. [6]

This paradox demands a precise framing. What has survived the political freeze is not AI governance cooperation in any policy sense. The two governments are not coordinating on regulation, standards, or enforcement. What has survived, and intensified, is AI governance research cooperation: the joint production of technical knowledge, shared vocabulary, and risk assessments by a transnational network of institutions. This distinction matters. Research cooperation cannot substitute for policy coordination, but it creates the epistemic preconditions upon which policy coordination can eventually be built. This essay maps that infrastructure, explains why it has proven resilient, identifies the structural tensions that limit its policy impact, and proposes concrete mechanisms to bridge the gap between research outputs and governance outcomes.

What’s Broken: The Collapse of Policy-Level Cooperation

Any honest assessment of U.S.–China AI cooperation must begin with what has failed. At the policy level, the two countries are further apart on AI governance than at any point in the past decade.

Washington’s pivot is not merely rhetorical. The Trump administration’s AI Action Plan, released in July 2025, directs agencies to promote American AI worldwide and to counter Chinese influence in international governance bodies. The systematic renaming of “safety” institutions signals a substantive reorientation: “Safety” implies shared vulnerability to a common risk; “security” implies competitive advantage to be maximized. U.S. policy now treats the American technology stack, including frontier AI systems, as a strategic export to be promoted globally, not a shared risk to be governed multilaterally. At the UN Security Council in September 2025, Michael Kratsios, director of the White House Office of Science and Technology Policy, rejected “centralized control and global governance” of AI while China endorsed multilateral oversight, a reversal of the roles each country played just two years earlier. [7]

This divergence is reinforced by the political economy of AI policy research in Washington. The dominant voices shaping U.S. AI policy are not safety-oriented research institutions, but security-competition think tanks—organizations funded by government contracts and defense-adjacent donors, oriented toward maintaining U.S. technological primacy. The researchers who participate in bilateral Track II dialogues and the policy professionals who brief congressional committees on AI strategy operate in largely separate professional worlds with limited interchange.

Yet it is precisely this hostile policy environment that makes research-level cooperation both more important and, counterintuitively, more feasible. Research cooperation does not require diplomatic agreements, government budgets, or political sponsorship. It can be funded by private foundations, convened by universities and independent organizations, and can produce outputs that retain their value regardless of the state of bilateral relations. The contrast with U.S.–China climate cooperation is instructive.

Climate cooperation depended heavily on top-level political sponsorship, from the Kerry–Xie channel to Obama–Xi joint declarations, and contracted sharply when that sponsorship was withdrawn under Trump’s first term. AI governance research cooperation was never built on that model, which is why the second Trump administration’s hostility to multilateral AI governance has not dismantled it. The question is whether anything substantive has been built on this more resilient basis.

What Survived: The Research Infrastructure

The answer is yes, and the evidence is more substantial than most observers realize.

The most durable channel is the Brookings–Tsinghua bilateral AI dialogue, launched in October 2019 between the Brookings Institution and Tsinghua University’s Center for International Security and Strategy. Twelve rounds, convening delegations of nearly 40 experts, have survived a pandemic, a trade war, the Taiwan Strait crisis and two changes in U.S. presidential administration. [8] Funded by the Berggruen Institute and Minderoo Foundation rather than by either government, the dialogue was structured from the outset to operate independently of official bilateral relations. Its most tangible output, a jointly-developed glossary providing parallel American and Chinese definitions for AI terms in national security, was published in August 2024 and most recently updated in January 2026. [9] This glossary does not harmonize policy; it enables two different regulatory systems to communicate accurately about the same phenomena. That is a modest contribution, but precisely the kind that survives political disruption because it serves both sides’ domestic needs.

The International Dialogues on AI Safety, convened by the Safe AI Forum, represent a parallel track focused on technical rather than policy dimensions. Five sessions since October 2023, at Ditchley Park, Beijing, Venice, Shanghai, and London, have brought together leading AI researchers from both ecosystems to produce joint technical assessments. [10] The Beijing session in March 2024 yielded a consensus statement identifying five technical “red lines,” including autonomous self-replication and AI-enabled weapons development, that participants from both countries agreed should not be crossed. [11] The Shanghai session in July 2025, co-hosted with the Shanghai Qi Zhi Institute, addressed AI deception risks, with Turing Award laureates Yoshua Bengio and Andrew Yao both present. [12] These dialogues derive their authority from their participants: The researchers at the table are the same individuals who advise their respective governments on AI policy.

Bridging these ecosystems are institutions that hold co-production relationships with both sides simultaneously. The U.S.-based Carnegie Endowment for International Peace has co-produced research on AI safety as a global public good with Concordia AI, a Beijing-registered social enterprise founded by Brian Tse, a former affiliate of Oxford’s Centre for the Governance of AI. Concordia AI also hosted the AI safety and governance forum at the World AI Conference in Shanghai in 2024 and 2025, serving as the conference’s official AI governance advisor in 2025. [13][14] The Singapore Consensus on Global AI Safety Research Priorities, the final output of a conference hosted by the Singapore government, was published in mid-2025 with 88 co-signatories from 11 countries. The consensus demonstrated that researchers from rival nations could converge on a shared scientific assessment of the field’s most urgent challenges even as their governments diverged. [15]

This infrastructure did not emerge by accident. Three features of AI governance as a problem domain explain why research cooperation has proven more resilient here than in climate, public health, or arms control.

First, the technical risks are genuinely shared and nonrivalrous. A misaligned system developed in San Francisco poses risks to users in Shanghai, and vice versa. DeepSeek’s release in January 2025, matching Western frontier performance at a fraction of the cost, confirmed that neither country can maintain a durable capability monopoly. This has a counterintuitive governance implication. The more the two countries’ AI capabilities converge, the more they are exposed to the same categories of technical risk, and the more valuable a shared understanding of those risks becomes. Research cooperation on risk identification serves both sides’ domestic governance needs, even if the two governments never coordinate their regulatory responses.

Second, AI governance lacks an equivalent to the Intergovernmental Panel on Climate Change or the World Health Organization. No mature international scientific infrastructure exists for assessing AI risks, which means the research network described above is not supplementing government capacity but substituting for its absence. Identifying failure modes, developing evaluation benchmarks, and proposing safety standards are inherently collaborative technical tasks that benefit from cross-border researcher participation, regardless of whether the resulting knowledge feeds into coordinated or separate national policies.

Third, the personnel networks built through institutions like Oxford’s Centre for the Governance of AI, whose alumni now hold positions spanning Georgetown’s Center for Security and Emerging Technology (CSET), the Carnegie Endowment, the UK’s AI Security Institute, Google DeepMind, and Concordia AI in Beijing, create a connective tissue that is structurally resilient to political disruption. [16] Even if any single bilateral channel were shut down, the professional relationships embedded in this network would sustain cooperation through alternative pathways.

Bridging the Gap Between Research and Policy 

The infrastructure is real, but so are its limitations. The central problem is transmission. How do research outputs gain policy relevance when the two professional communities that produce them, the safety-oriented Track II world and the security-competition policy mainstream, operate on parallel tracks?

This transmission problem takes different forms on each side. In the United States, the gap is institutional and ideological. Track II research cooperation is funded by private philanthropy, principally from foundations committed to “AI safety” as a global public good: the Berggruen Institute, Minderoo Foundation, Open Philanthropy, and similar donors. The policy research that shapes congressional hearings and executive branch strategy comes from a different ecosystem entirely. These two communities share some personnel but operate on largely separate institutional tracks. The Brookings–Tsinghua glossary is more technically precise than anything the U.S. government has produced on AI terminology, yet no mechanism feeds its definitions into the standards-setting processes of the National Institute of Standards and Technology (NIST). The result is a paradox within the paradox: The most substantive U.S.–China AI cooperation is happening precisely where it has the least direct policy influence.

In China, the gap takes a different form. Chinese AI governance institutions, including the China AI Development and Safety Research Network launched in February 2025, operate closer to the state than their Western counterparts. [17] This proximity gives Chinese researchers more direct policy influence domestically, but raises questions about institutional independence that limit the trust Western counterparts can invest in joint outputs. The network includes the China Academy of Information and Communications Technology (CAICT) and the China Center for Information Industry Development, both agencies with direct government ties. For Western researchers, the question of whether a Chinese co-author’s positions reflect independent analysis or state-directed messaging is not a hypothetical concern but a practical consideration that shapes what kinds of joint projects are feasible.

Three transmission mechanisms could address these gaps, and each is already visible in embryonic form.

The first is standardization. Creating designated liaisons between Track II participants and national standards bodies, NIST in the U.S. and TC260 in China, would give research outputs a concrete pathway into policy without requiring diplomatic agreement. The glossary’s definitions of terms like “frontier model” and “autonomous capability” are already more precise than what either government uses; the gap is not in the research but in the institutional plumbing.

The second is personnel flow. When Jade Leung moved from Oxford’s GovAI to become chief technology officer of the UK’s AI Safety Institute (since renamed the AI Security Institute), she carried concepts, risk frameworks and professional relationships formed through cross-border research. [16] This is the most organic mechanism through which research cooperation influences policy. Both governments could accelerate it by recruiting from the Track II community for AI governance roles, rather than drawing primarily from the security-competition policy world.

The third is structured policy briefing. For the United States, the conceptual shift required is to recognize Track II research cooperation as intelligence, not concession. The Brookings–Tsinghua dialogue is the primary window through which American researchers understand how Chinese AI governance thinking is evolving. Congress should ensure that research funding is not contingent on avoiding Chinese collaboration and should actively welcome Track II participants to brief policymakers.

For China, the priority is institutional transparency. Publishing research methodologies, opening governance structures to external review and continuing to participate in international peer processes, as Chinese researchers did by co-signing the Singapore Consensus, build the credibility that makes joint outputs usable.

Finally, both sides have an underexploited opportunity in third-country capacity building. The countries most in need of AI governance capacity, in Southeast Asia, Africa, and Latin America, are poorly served by Washington’s technology-export model or Beijing’s multilateral-framework rhetoric alone. Singapore’s hosting of the Conference on AI in April 2025 demonstrated that third countries can serve as convening platforms that are acceptable to both sides. The existing research network already includes institutions from Singapore, India, Brazil, and South Africa. Channeling joint effort toward governance capacity building in these regions would give both countries a tangible stake in maintaining the infrastructure, while producing a genuinely multilateral public good that neither could deliver alone.

The Choice Ahead

In October 2025, Trump and Chinese President Xi Jinping met on the margins of the Asia-Pacific Economic Cooperation summit and reached a one-year trade truce; at the summit itself, Xi proposed a “World AI Cooperation Organization.” [18] No bilateral framework for AI cooperation has followed, and none may during this administration. This essay has argued that the right response is not to wait for one.

The research infrastructure connecting American and Chinese AI governance institutions is not a substitute for policy cooperation. It cannot harmonize regulations or resolve strategic competition. What it can do is maintain a shared epistemic foundation—a common vocabulary, a body of jointly produced technical knowledge, and a network of professional relationships—so that when political conditions shift, the groundwork will already be in place. The mechanisms to connect this research to policy, through standardization pipelines, personnel flows into government and structured briefings, are identifiable and actionable. None requires a diplomatic breakthrough. All require the deliberate choice to invest in institutional infrastructure rather than leaving it to chance.

The history of U.S.–China relations offers few examples of cooperative infrastructure that has strengthened during a period of acute strategic tension. The AI governance research network is one. It was built not by diplomats but by researchers who recognized, years before their governments did, that the risks posed by frontier AI systems do not respect national borders. Whether policymakers on both sides choose to treat this infrastructure as a strategic asset or dismiss it as an insignificant byproduct of academic exchange will shape not just the bilateral relationship, but the governance of the most consequential technology of the 21st century.

Endnotes

[1] The White House. (2025, January 20). “‘Initial Rescissions of Harmful Executive Orders and Actions’ (Executive Order 14148)“, revoking Executive Order 14110. The replacement order, “Removing Barriers to American Leadership in Artificial Intelligence” (Executive Order 14179), was issued January 23, 2025.

[2] France 24. (2025, February 11). “US, UK Do Not Sign Paris AI Summit Final Statement.” The declaration was signed by 61 countries and blocs.

[3] U.S. Department of Commerce. (2025, June 3). “Transforming the U.S. AI Safety Institute into the Center for AI Standards and Innovation (CAISI).”

[4] The International Network of AI Safety Institutes was renamed the “International Network for Advanced AI Measurement, Evaluation and Science,” announced December 9, 2025.

[5] Singer, Scott, Karson Elmgren, and Oliver Guest. (2025, June 16). “How Some of China’s Top AI Thinkers Built Their Own AI Safety Institute.” Carnegie Endowment for International Peace.

[6] 复旦发展研究院 [Fudan Development Institute]. (2026, March 14). “全球人工智能治理工作坊：共探前沿人工智能国际治理路径” [Global Artificial Intelligence Governance Workshop: Exploring Pathways for International Governance of Frontier AI].

[7] The White House. (2025, July 23). “Winning the Race: America’s AI Action Plan.” The quoted rejection of “centralized control and global governance” of AI comes from OSTP Director Michael Kratsios’s statement to the U.N. Security Council, 10005th meeting, September 24, 2025.

[8] Hass, Ryan, and Colin Kahl. (2024, April 5). “Laying the Groundwork for US-China AI Dialogue.” Brookings Institution; Tsinghua CISS. (2024). “The China-U.S. Track II Dialogue on AI and International Security: Interim Report.”

[9] Brookings Institution and Tsinghua CISS. (2024, August 30; updated 2026, January 20). “Glossary of Artificial Intelligence Terms.”

[10] International Dialogues on AI Safety; Safe AI Forum (SAIF), saif.org.

[11] IDAIS-Beijing. (2024, March). “Consensus Statement on Red Lines in Artificial Intelligence.”

[12] IDAIS-Shanghai. (2025, July 25). Co-hosted with the Shanghai Qi Zhi Institute and Shanghai AI Lab.

[13] Concordia AI, Oxford Martin AI Governance Initiative, and Carnegie Endowment for International Peace. (2025, March). “Examining AI Safety as a Global Public Good“; Concordia AI and Carnegie Endowment for International Peace. (2024, August). “The Future of International Scientific Assessments of AI’s Risks.”

[14] Concordia AI. (2025, July 27). “AI Safety and Governance Forum at the World AI Conference 2025.“

[15] Bengio, Yoshua, et al. (2025, June). “The Singapore Consensus on Global AI Safety Research Priorities.” arXiv:2506.20702. 88 signatories.

[16] Personnel flows synthesized from public career records: Allan Dafoe (GovAI → Google DeepMind), Helen Toner (GovAI → CSET → OpenAI board), Jade Leung (GovAI → OpenAI → U.K. AI Security Institute CTO), Brian Tse (GovAI affiliate → Concordia AI), Fynn Heide (GovAI → Safe AI Forum co-founder).

[17] Singer, Elmgren, and Guest (2025), see note [5]. Carnegie notes that CnAISDA includes CAICT and CCID, agencies with direct government ties, and describes it as a “loose coalition” rather than an independent institute.

[18] Reuters. (2025, November 1). “China’s Xi Pushes for Global AI Body at APEC in Counter to U.S.” On the one-year trade truce reached at the October 30 Trump–Xi meeting in Busan, see CNN, October 31, 2025.

Image Credit: Marcel Grabowski / UK Government (2023). Via Flickr, CC BY 2.0.