The framing of US-China AI competition as a barrier to AI governance treats two things as equivalent that are actually distinct. Competition and cooperation on shared catastrophic risk are not mutually exclusive. The United States and Soviet Union competed intensely across every dimension of national power for four decades while simultaneously negotiating and maintaining a series of arms control agreements that constrained the weapons most capable of ending civilization. The conditions under which that cooperation was possible are worth understanding for what they can tell us about the AI safety problem.
What made US-Soviet cooperation possible
Several conditions were necessary for arms control agreements between the two Cold War powers. None of them required that the competition end, or that the two sides trust each other in any general sense.
The first was a shared and concrete understanding of mutual catastrophe. Both the United States and Soviet Union had observed the effects of nuclear weapons — the US through use and testing, the Soviet Union through testing and through intelligence on Hiroshima and Nagasaki. By the late 1950s, both sides understood that a full nuclear exchange would be catastrophic for both. This was not a theoretical assessment; it was grounded in physical evidence of what nuclear weapons did. The Cuban Missile Crisis in 1962 made this understanding visceral and immediate for political leaders in a way that years of strategic analysis had not.
The second was a sustained informal dialogue that maintained intellectual and personal connections across the political divide. The Pugwash Conferences on Science and World Affairs, beginning in 1957, brought together scientists from both blocs to discuss arms control before any formal negotiation was possible. This informal network built the ideas and relationships that formal negotiations later formalized. When US and Soviet officials sat down to negotiate the Limited Test Ban Treaty in 1963, they were not starting from zero.
The third was careful scoping. The Limited Test Ban Treaty (1963) banned nuclear testing in the atmosphere, outer space, and underwater, but not underground. It addressed a problem, atmospheric fallout from testing, where both sides had a genuine shared interest in a solution, without requiring either side to accept constraints on their weapons programs or verification of their arsenals. It was a narrow first agreement that built confidence for larger ones to follow. SALT I (1972), the ABM Treaty (1972), and the Chemical Weapons Convention came later, each building on the architecture of the last.
The fourth was that verification was built in from the beginning, not added later. Each agreement was designed with explicit verification mechanisms: national technical means for satellite reconnaissance were incorporated into the SALT framework, with both sides agreeing not to interfere with the other's monitoring satellites. This allowed each side to verify compliance without trusting the other's declarations. Agreements that required trust to function were not agreed; agreements that could be verified without trust were.
Where the US-China comparison holds
The United States and China share structural features with the US-Soviet dynamic that are relevant to AI governance.
Both are developing frontier AI systems with potential military and economic applications that neither side wants the other to have. Both have stated, at official levels, concern about the risks of advanced AI developed without adequate safety constraints. Both face a scenario — an AI system that behaves in catastrophically unintended ways — that would be harmful to both countries regardless of which country's system produced it. An unaligned superintelligence developed by a Chinese laboratory would be dangerous to China as much as to the United States, and vice versa. The risk category of greatest concern to AI safety researchers is not rivalrous in the way that nuclear weapons capability is. This creates a genuine structural basis for shared interest.
There is also a growing informal dialogue. The Bletchley AI Safety Summit in November 2023 included Chinese government participation, the first time Chinese officials engaged in a multilateral forum specifically focused on frontier AI safety. Chinese AI researchers have engaged in international safety discussions. In late 2024, the US and China reached an agreement to begin bilateral talks on AI risk. These are early and fragile, easily disrupted by other aspects of the bilateral relationship, but the infrastructure for dialogue exists.
US-Soviet arms control did not require the two sides to trust each other. It required them to agree that certain catastrophic outcomes were worse than the constraints needed to prevent them, and to design verification systems that made compliance observable without relying on good faith.
Where the comparison breaks down
The US-Soviet relationship had a clarity that the current US-China relationship lacks. Both Cold War superpowers agreed, at the official level, that nuclear weapons were weapons and that a full nuclear exchange would be catastrophic. The physical effects of nuclear weapons were documented, witnessed, and understood. The threat was concrete.
The AI catastrophic risk remains, for most political leaders, anticipatory rather than experienced. No government has incorporated advanced AI existential risk into its national security framework with the same weight that nuclear risk has carried since 1945. The scenario of an unaligned superintelligent AI causing civilizational harm has no historical instance behind it. This makes political mobilization around prevention harder, because politicians respond to demonstrated harms more reliably than to anticipated ones.
The economic dimension of the US-China relationship also complicates matters in ways that had no Cold War equivalent. The US and Soviet Union had largely separate economies. The US and China are deeply economically interdependent, with supply chains, technology transfer, and investment flows running in both directions. AI capabilities are central to economic competition in this relationship, not just military competition. This makes AI harder to bracket off from the general competitive relationship and negotiate about separately, because restrictions on AI development also affect economic prospects in ways that restrictions on, say, medium-range nuclear missiles did not.
The current state of dialogue
As of mid-2026, US-China AI dialogue is cautious and intermittent. Bilateral discussions that began in 2024 have continued at the working level, focused primarily on information-sharing about AI risk assessments rather than binding commitments. Both governments have publicly endorsed the goal of preventing catastrophic AI outcomes while simultaneously maintaining competitive AI development programs. At the multilateral level, both attended the 2023 Bletchley Summit and the 2024 Seoul Summit, signing declarations that acknowledged shared concern about frontier AI risks without creating binding obligations.
US export controls on advanced semiconductor technology have been a persistent source of bilateral tension. These controls, which restrict China's access to the most advanced chips used in frontier AI training, are motivated primarily by competitive considerations rather than safety ones. From a governance perspective, they represent a form of unilateral compute governance with some of the same practical effects as treaty-based compute limits, though without any of the reciprocal constraints or verification mechanisms that treaty-based governance would include.
What conditions would enable progress
Three conditions have historically preceded meaningful arms control progress between rivals, and each has an AI equivalent.
A defining moment that makes the risk concrete and political. The Limited Test Ban Treaty came after atmospheric testing by both sides produced measurable global fallout that concentrated public opinion against continued testing. An AI governance equivalent might be a serious incident, a deployment of a highly capable system that behaved in significantly unintended ways at scale, that makes the risk politically visible to domestic audiences in both countries simultaneously. Governments that have been resistant to preventive action become more receptive after near-miss experiences.
Technical confidence-building measures before binding commitments. Before SALT I, both sides used satellite reconnaissance to monitor the other's missile programs, with tacit acceptance that this monitoring was occurring. The AI equivalent might be transparency agreements around training run notifications, shared evaluation standards, or cooperative safety research. These measures build confidence that the other side is engaging in good faith, creating the political foundation for larger commitments.
Third-party facilitation that allows both sides to participate without appearing to concede to the other. Several successful arms control initiatives benefited from multilateral framing: both sides could engage with an international process rather than making bilateral concessions to each other. A multilateral AI governance process that includes both the US and China, hosted by a neutral body, may be the pathway to agreements that bilateral negotiations framed as US-China encounters would not produce.
"The US and China do not need to trust each other to cooperate on AI safety. They need to agree that catastrophic AI outcomes are worse for both of them than the constraints needed to prevent those outcomes."
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
The question of whether the US and China can agree on AI safety is ultimately a question about whether both governments will conclude, before an adequately capable AI system exists, that binding mutual constraints are preferable to an unconstrained race. History suggests this conclusion is reachable, given the right conditions. History also suggests that governments rarely reach it through calm deliberation in advance. They reach it when the alternative has become sufficiently vivid to change the political calculation. The policy question is whether the governance window will remain open long enough for that change to occur.
Common questions.
Yes, in limited ways. Both countries are NPT parties and accept IAEA safeguards on their civilian nuclear programs. Both attended the 2023 Bletchley AI Safety Summit and signed the resulting declaration acknowledging shared concern about catastrophic AI risks. Bilateral AI safety discussions began in 2024. Neither country has accepted binding constraints on its AI development programs, but the infrastructure for dialogue exists.
Probably more difficult. The US-Soviet nuclear risk was grounded in documented experience of what nuclear weapons did; the AI catastrophic risk remains anticipatory for most policymakers. The economic interdependence between the US and China creates a different competitive dynamic than the primarily military Cold War competition. And AI is more dual-use: civilian and military AI development are far harder to separate than civilian and military nuclear programs.
Yes, and this is probably the more realistic near-term path. Notification agreements, shared safety standards for high-risk AI applications, and cooperative safety research are all achievable at lower political cost than a comprehensive treaty. These steps would build the confidence and institutional relationships needed for larger commitments later — similar to how the Hot Line agreement (1963) and Limited Test Ban Treaty preceded SALT I by nearly a decade.
A governance gap at the most consequential level, with regulatory activity continuing in other jurisdictions. The EU, UK, and others would develop frameworks for AI deployed in their markets, creating indirect pressure on US and Chinese companies. But without binding commitments from both leading AI-developing nations, safety governance for frontier AI would remain structurally incomplete regardless of how many other countries join a treaty.