U.S., China Launch Formal AI Dialogue After Trump-Xi Summit, Treasury Says

WASHINGTON/BEIJING — In the wake of a closely watched summit between U.S. President Donald Trump and Chinese leader Xi Jinping, a comparatively quiet announcement may prove to be one of the meeting’s most consequential outcomes: both governments say they will begin a formal, intergovernmental dialogue on artificial intelligence.

U.S. Treasury Secretary Scott Bessent said Washington and Beijing would “set up a protocol” focused on best practices to keep powerful AI models out of the hands of non-state actors. China’s Foreign Ministry, through spokesperson Guo Jiakun, separately confirmed that Trump and Xi agreed to launch an “intergovernmental dialogue” on AI.

The language signals intent, but the harder test is whether the process becomes a durable channel for risk reduction—or another round of high-level talking points that fails to change behavior. Researchers and policy specialists who have studied U.S.-China technology dynamics say the path to meaningful progress is narrower than the rhetoric often suggests: start small, keep it technical, and avoid turning the talks into a proxy arms-control negotiation.

A window opens—amid shaky governance

The timing is notable. The Trump administration has recently signed an executive order establishing a framework for AI companies to coordinate with the U.S. government on managing model risks, giving Washington a more structured policy platform to bring into international discussions. But confidence in U.S. follow-through has been dented by recent domestic governance turmoil, including a high-profile episode involving Mythos–Fable that underscored how quickly AI policy can be buffeted by industry missteps and political crosswinds.

That combination—new formal authorities alongside visible fragility—helps explain why experts argue for realism about what early dialogue can achieve.

Former U.S. officials and analysts have floated ambitious goals, including reciprocal constraints on frontier AI development. Others have argued for a “crisis management mechanism” that would define red lines across AI-enabled cyber operations, biosecurity, autonomous military escalation, model theft, and third-party use. Yet specialists familiar with past U.S.-China confidence-building efforts caution that pushing for binding limitations too early risks deadlock, especially in a period of strategic mistrust and rapid technical change.

Not arms control—at least not yet

A growing consensus among researchers is that the next phase should not be framed as bargaining where one side limits its AI capabilities in exchange for concessions from the other. Instead, they argue, the talks should be built as a standing, working-level process for exchanging methods: how each government and its technical agencies evaluate model risk, how they test before deployment, and what mitigations actually reduce danger in practice.

That approach reflects a practical reality: many of the most alarming AI risks—such as enabling offensive cyber operations or lowering barriers to chemical and biological weaponization—are easier to discuss as engineering and assessment problems than as geopolitical trade-offs. Even comparatively “mundane” failures like hallucinations, flawed code generation, or faulty recommendations could become catastrophic when embedded in critical infrastructure, supply chains, or military logistics.

In this view, the early deliverable is not a treaty. It is a shared vocabulary for risk and a regular rhythm of contact between specialized experts who can compare protocols—especially around pre-release testing, red teaming, evaluation benchmarks, and incident reporting.

What a serious agenda could look like

Researchers who study AI governance have proposed two core themes for bilateral talks: establishing a shared understanding of risk, and sharing best practices for reducing those risks. Experts say the initial agenda should focus on a narrow list of model capabilities with clear, measurable harms.

One obvious category is cybersecurity: how models might assist vulnerability discovery, phishing, malware development, or automated intrusion workflows. Another is biological, chemical, or nuclear risk: whether models can meaningfully accelerate weapon design, procurement, or deployment planning. A third is reliability in high-stakes settings, where model errors—rather than malicious use—could still cause large-scale damage.

The most useful exchanges, researchers say, would drill into concrete practices: what each side considers adequate red-team coverage, how they evaluate dangerous capability uplift, how they secure model weights against theft, and how they limit “tool access” that could turn a general model into an operational threat. Mitigations could include alignment techniques, controlled deployment “sandboxes,” stronger system hardening, and procedures that ensure models do what users intend—especially when integrated with external tools.

“Parallel” action, not conditionality

A key design principle gaining traction is “parallel” action: both sides take steps they believe are prudent for their own reasons, while using dialogue to learn from each other and reduce misunderstandings. That is different from conditional cooperation, in which safety measures are adopted only if the other party agrees to match them.

The precedent often cited is a 2024 moment when Biden and Xi affirmed that humans should retain control over nuclear weapons decisions. That declaration was not a formal agreement and did not hinge on reciprocal enforcement; it was a synchronized signal of restraint, shaped by shared recognition of catastrophic downside.

Analysts say AI talks could generate similar “mutual understandings” over time—statements of practice that stop short of binding commitments, but still help stabilize expectations and guide internal policy.

For now, the immediate challenge is procedural: ensuring the dialogue has the right participants, a recurring schedule, a secure channel for sensitive technical discussion, and a mechanism for urgent outreach when surprises occur. Without that infrastructure, even well-intentioned leader-level directives can dissipate into bureaucratic inertia.

Why it matters

Advanced AI systems are increasingly implicated in cyber operations, critical infrastructure, and national security decision-making—areas where miscalculation can escalate quickly. A working U.S.-China channel that prioritizes practical risk assessment could reduce the odds of accidental escalation and slow the diffusion of dangerous capabilities to third parties. The stakes are global: if the two leading AI powers cannot translate dialogue into routine cooperation on safety basics, smaller states and non-state actors may shape the risk landscape instead.