Artificial intelligence has entered diplomacy.

The United States and China are preparing formal talks about frontier models, military risk, cyber capabilities, labor disruption, intellectual property, and the spread of powerful systems to non-state actors.

That may sound like another technical meeting between rival governments.

It is potentially much more than that.

For years, AI governance has been discussed as a domestic problem. Congress debates company liability. Europe writes transparency rules. Governments commission safety institutes. Technology companies publish voluntary commitments and argue over whether models should remain closed or be released with downloadable weights.

But the most powerful systems now cross all of those boundaries.

A model developed in one country can be copied, distilled, hosted elsewhere, modified by private actors, used in cyber operations, inserted into military planning, or deployed across millions of workplaces. Its economic consequences can move faster than labor law. Its security implications can move faster than treaties.

Washington and Beijing are beginning to acknowledge that neither can govern this alone.

That is the encouraging part.

The troubling part is that they may try to govern it together without the rest of the world.

The proposed talks are expected to address what counts as a frontier model, how powerful systems should be controlled, whether open-weight releases create unacceptable proliferation risks, and how to manage technologies that could assist cyberattacks or military operations. Even defining the category matters. Once governments decide that a certain class of model is strategically sensitive, that definition can become the basis for export controls, sanctions, licensing systems, surveillance, and restrictions on research.

The line between safety and industrial policy can become very thin.

The United States accuses Chinese firms of extracting capabilities from American models. China worries that Washington may restrict access to its open-weight systems while continuing to limit advanced chips. Both governments describe their policies as necessary for security. Both also have enormous economic incentives to preserve domestic advantage.

This does not make negotiation pointless.

Nuclear rivals created communication channels because the alternative was not freedom. It was unmanaged escalation. AI systems with military, cyber, and biological capabilities deserve serious international rules before a crisis forces hurried ones.

But AI is not nuclear technology.

The infrastructure is more distributed. The private sector owns much of the capability. Models can be copied. Software evolves rapidly. Civilian and military uses are intertwined. Universities, small companies, open communities, and countries outside the two dominant blocs all participate in the ecosystem.

A bilateral agreement could reduce some risks while concentrating enormous authority.

Imagine Washington and Beijing agreeing that advanced open-weight models should not be released internationally. That might slow dangerous proliferation. It might also lock universities, smaller countries, and public institutions into permanent dependence on approved corporate and state platforms.

Imagine them agreeing on watermarking or identity requirements. That could improve accountability. It could also expand surveillance and make anonymous research or political speech harder.

Imagine them defining acceptable model behavior through security priorities developed behind closed doors. The resulting systems may be safer from certain attacks while becoming more obedient to concentrated power.

The right lesson is not that the talks should fail.

It is that they should not become the constitutional convention for humanity’s relationship with machine intelligence.

Other governments need representation. So do labor organizations, technical researchers, civil society, educators, communities affected by data centers, and countries that do not want to choose between American corporate control and Chinese state influence.

The United States and China may be the only actors capable of slowing some forms of escalation.

They are not the only societies that will live with the consequences.

The first serious AI diplomacy should therefore begin with a modest principle:

The countries with the most power may convene the room.

They should not be allowed to own every chair.