Washington Opens an AI-Risk Channel With Beijing

The United States is preparing to discuss shared artificial-intelligence risks with China, creating a diplomatic opening that would have seemed unlikely only days ago. Treasury Secretary Scott Bessent told Axios that Washington is willing to discuss ways to avoid common dangers and prevent the two countries’ AI ecosystems from splitting completely. The talks are expected to take place in New York this weekend alongside broader negotiations on trade, rare earths and other economic issues. (axios.com)

The language is deliberately limited. Washington is not proposing a joint pause in AI development, a global regulator or a treaty that would constrain the leading laboratories. Instead, the administration is signaling that national competition and risk management can coexist. That distinction matters because the White House and congressional Republicans have repeatedly argued that slowing American AI development would hand strategic advantages to China.

A new channel inside an old rivalry

The immediate development is less a breakthrough than a change in diplomatic posture. Until now, the dominant U.S. framing has treated AI safety primarily as a domestic question and a contest of technological leadership. Bessent’s statement adds a third layer: some risks may be too cross-border to manage independently, even when the countries remain strategic competitors.

Axios reported that the discussions could cover both open-weight and closed-weight models. That is significant because the two categories create different governance problems. Closed models are controlled by companies that can restrict access, monitor usage and impose contractual rules. Open-weight systems can spread more rapidly, be modified by downstream users and become difficult to recall once released. A conversation that includes both would move beyond general statements about “responsible AI” toward the practical question of which safeguards can work across incompatible commercial and security environments. (axios.com)

The agenda also reflects how quickly AI has become entangled with national security. Officials are increasingly concerned about systems that can assist cyberattacks, accelerate biological research or automate influence operations. Those risks do not stop at a border, but neither do they lend themselves easily to trust-based cooperation. A government may agree that a capability is dangerous while still wanting its own laboratories to reach that capability first.

That is the central contradiction facing the talks. The United States and China may share an interest in preventing catastrophic misuse, yet each sees advanced AI as a source of military, economic and political leverage. Cooperation therefore has to be narrow enough to preserve competitive advantage and concrete enough to produce benefits that neither side can obtain alone.

Why it matters

The opening matters because AI governance has been losing ground to the logic of an arms race. The Associated Press reported that AI companies and researchers have found rare agreement about the need for stronger safety measures, but that domestic competition, commercial incentives and political resistance make coordinated action difficult. The same obstacles apply internationally, with an added layer of distrust between Washington and Beijing. (apnews.com)

A bilateral channel could establish modest but useful forms of risk reduction. The two governments might exchange information about AI-enabled cyber incidents, define common terminology for dangerous model behavior, or create emergency contacts for situations in which a system appears to be causing harm across jurisdictions. They could also compare testing methods without disclosing proprietary model weights or sensitive military capabilities.

Even limited protocols would be valuable. In fast-moving technologies, accidents often become crises because officials lack shared definitions and reliable communication. A hotline does not end rivalry, but it can reduce the chance that a cyber incident, model release or fabricated political message is misread as a deliberate state action.

The talks could also pressure private laboratories to clarify their own operating standards. If Washington and Beijing begin asking companies to document model evaluations, report major incidents or restrict certain forms of access, firms will face stronger incentives to build those processes before a crisis forces them to do so. That would be a meaningful shift from voluntary principles toward operational expectations, even without a formal treaty.

But the political signal may be as important as any technical agreement. The administration’s willingness to discuss shared risks suggests that “AI safety” is no longer being treated solely as an argument for slowing innovation. It is being reframed as part of economic security and strategic stability. That language may make safeguards more acceptable to policymakers who reject the idea of a general moratorium.

The limits of cooperation

The first uncertainty is whether the meetings will produce anything beyond a statement of intent. The reporting indicates that discussions are planned, not that the two governments have agreed on a framework. The difference is crucial. Diplomatic meetings can create a channel without resolving the underlying disputes over chips, model access, data, military applications and export controls.

A second uncertainty is enforcement. Even if officials identify shared principles, neither side may be willing to let the other verify compliance. The United States could fear that technical exchanges expose sensitive capabilities or weaken export restrictions. China could view U.S. safety proposals as a disguised mechanism for preserving American technological dominance. Both governments may prefer ambiguous commitments that are politically useful but difficult to test.

There is also a disagreement over what counts as the primary danger. Many American officials frame Chinese leadership in AI as the largest strategic risk, while safety advocates emphasize loss of control, misuse and systemic harm regardless of which country leads. Beijing, meanwhile, has criticized calls for a slowdown as fearmongering and confrontation. Those competing priorities could make even basic language contentious. (apnews.com)

The upcoming talks therefore should not be read as evidence that the AI race is ending. They are better understood as an attempt to install guardrails while the race continues. That may be the only politically realistic path: preserve competition, limit the most dangerous failure modes and build communication mechanisms before the technology becomes harder to govern.

The test will be whether Washington and Beijing can move from shared vocabulary to shared procedures. If they can, even a narrow agreement on incident reporting or crisis communication would be consequential. If they cannot, the new channel may still prove useful by making the boundaries of cooperation clearer. Either way, AI diplomacy has entered a more serious phase—one in which strategic rivals are beginning to acknowledge that some risks are too large to manage entirely alone.