Last verified: September 20, 2026. The United States and China are, for the first time since Trump’s second term began, building a dedicated government-to-government channel on AI — not an export-control fight or a single lab’s misconduct, but a proposed working-level dialogue on shared AI-safety risks, expected to surface around President Xi Jinping’s September 24, 2026 visit to Washington. No formal name, charter, or signed text exists yet; this piece covers what has actually happened, what is proposed, what both governments have explicitly excluded, and why officials on both sides are managing expectations down. It also covers where this kind of state-to-state proposal sits — and where it doesn’t yet sit — against NIKOLAI, CASRAI’s own frontier-AI-safety crosswalk.
What’s actually on the table
The clearest public marker so far is a September 16, 2026 pre-summit meeting between U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng, where AI was folded into a broader economic agenda rather than given its own dedicated session. According to reporting from China Tech News, the American side’s priority is establishing a joint AI “working group” — officials with both technical expertise and negotiating authority — though that group has not been formally stood up. The agenda both sides have floated publicly covers:
- Rogue AI models and agents — shared concern about systems acting outside intended bounds.
- Non-state bio-weapon risk — preventing AI systems from lowering the barrier to biological weapons development by non-state actors.
- Infrastructure and cyberattack risk — AI-directed threats to critical infrastructure.
- Information-sharing on AI safety incidents — a lighter-weight version of the threat-intelligence exchange already discussed between labs (see below).
- “Avoiding bifurcation of our two systems” — language used by U.S. officials to describe the broader goal of keeping some technical channel open even as the two countries otherwise decouple AI supply chains.
Separately, Nikkei reported on September 7, 2026 that Commerce Secretary Howard Lutnick had explicitly ruled out one topic: semiconductor and chip export controls are not part of these AI-safety discussions, and are being negotiated, if at all, through a different channel entirely. That is a deliberate scope decision worth underlining, because it is the detail that most cleanly separates this dialogue from CASRAI’s existing export-control coverage — see Compute Governance: Export Controls, FLOP Thresholds, and Chip-Level Oversight and Wolf Amendment: NASA Funding and Chinese Collaboration, both of which cover restrictions this dialogue is not attempting to touch.
How we got here: from a 2024 nuclear pledge to a 2026 restart
This is not the first time Washington and Beijing have talked about AI risk directly. In November 2024, at the sidelines of APEC in Lima, the two countries’ leaders affirmed that decisions over the use of nuclear weapons should remain under human, not AI, control — a narrow but real prior data point. Formal AI-specific talks then went dormant. China Tech News reports that the current restart traces to a Trump-Xi agreement made in Beijing in May 2026 to resume dialogue, with the September 16 Bessent–He Lifeng meeting as the first visible follow-through and the September 24 Washington summit as the next marker; the same reporting names November and December 2026 as windows for further meetings if the September summit produces anything to build on.
Whether it produces much is an open question the U.S. side is not pretending otherwise about. American officials quoted in that same reporting say neither government expects a breakthrough from the September 24 summit itself, though a parallel statement reaffirming the 2024 nuclear-control consensus, alongside a general statement of shared concerns, is considered plausible. One scholar’s more blunt framing, as quoted by China Tech News: “I’m not that optimistic for progress beyond a statement of principles.”
What’s deliberately off the table
Beyond chip export controls, the reporting surfaces several structural obstacles specific to why this dialogue is starting narrow rather than broad:
- Mutual espionage accusations. Both governments have separately accused the other of AI-related data theft, which limits how much technical detail either side is willing to share even inside a “safety” channel.
- Asymmetric framing of AI’s purpose. U.S. officials frame the effort partly in strategic-advantage terms; Chinese counterparts have pushed back that safety framing can function as cover for technology protectionism, and have emphasized AI’s economic role instead.
- Diplomatic process mismatch. Analysts cited by China Tech News note China’s diplomatic process tends to run on tightly scripted, pre-cleared statements, which sits awkwardly against how fast frontier-AI developments actually move.
A separate, non-governmental development sits alongside this official track and is easy to conflate with it, so it’s worth naming precisely: on September 17, 2026, Brookings and Shanghai’s Fudan University jointly published “Advancing Human Control of Military AI,” a track-two research proposal — not a government position — urging Presidents Trump and Xi to agree that only humans, not AI systems, should be able to initiate a cyberattack on each other’s nuclear command, control, and communications (NC3) networks or critical infrastructure. The report’s authors point to a February 2026 war-game simulation in which AI models chose to use nuclear weapons in 95% of scenarios, and frame the stakes directly: “The governance of military AI is not a negotiation that can afford to wait — it’s a race against time.” As of this writing, neither government has adopted the proposal; it is research input into the dialogue, not a term either side has committed to.
How this differs from CASRAI’s other US-China AI coverage
CASRAI already covers two adjacent but distinct US-China AI threads, and this dialogue is neither of them. It is not export control (chip and compute restrictions, covered in the guides linked above), and it is not lab-level misconduct or safety-framework compliance at a single company (see, for a different kind of cross-border AI coordination entirely, A Cross-Lab AI Safety Standards Body? What’s Confirmed So Far About the Anthropic-OpenAI-Google Talks, which covers labs coordinating with each other, not governments). This piece is specifically about two national governments discussing a shared safety channel directly, which is also distinct from the academic consensus document covered in CASRAI’s IDAIS-Beijing Statement, Explained — that statement was signed by individual scientists, not negotiated by either government.
Where NIKOLAI Fits In
CASRAI maintains NIKOLAI, CASRAI’s own independent, unendorsed reference dictionary for frontier-AI-safety terminology — not a standard, and not something the U.S. government, the Chinese government, Brookings, or Fudan has adopted, endorsed, or been consulted on. The most relevant of NIKOLAI’s 64 elements here, verified directly against the live element page on September 20, 2026, is Commitment (element N9, track N9 — Commitments and governance). NIKOLAI defines a commitment as “a public statement by an identified party (a developer, a government body, or a named individual signatory) to do or not do something, carrying an explicit modality (binding / unilateral / conditional / aspirational), a scope, an effective or target date, and a status value that can be tracked over time” — the element’s own definition is written broadly enough to include a government as the committing party, and NIKOLAI’s existing crosswalk for it already documents state and multi-stakeholder-level examples (the EU’s GPAI Code of Practice, and Seoul Summit signatories) sitting alongside individual lab pledges from Anthropic, OpenAI, Google DeepMind, xAI, and Meta.
The US-China dialogue covered in this piece is a genuine contrast case rather than a match, and it’s worth being precise about why: unlike the lab pledges and multilateral statements NIKOLAI’s Commitment element already tracks, nothing described above yet has a named party, a locked modality, or a target date attached to it. A working group that hasn’t been formally established, a summit whose organizers are not expecting a breakthrough, and a track-two research paper neither government has adopted do not, on their own, add up to a Commitment record — that would require an actual dated pledge from either government, which as of September 20, 2026 does not exist. If the September 24 summit or a later session produces one (even a narrow reaffirmation of the 2024 nuclear-control language), it would be a real candidate for a future shadow-mapping row. This is CASRAI’s own reading, not something NIKOLAI, the U.S. government, the Chinese government, or any signatory has filed or agreed to; that changes only if an organization files its own Mapping Declaration.
Frequently asked questions
Has the US and China signed a formal AI safety agreement as of September 2026?
No. As of September 20, 2026, there is no signed US-China AI agreement. A joint AI working group has been proposed by the American side but not formally established, and U.S. officials involved in preparatory talks say they do not expect a breakthrough from the September 24, 2026 Trump-Xi summit itself.
What did the US and China previously agree about AI and nuclear weapons?
In November 2024, at APEC in Lima, the two countries’ leaders affirmed that decisions over the use of nuclear weapons should remain under human control rather than being delegated to AI. That remains the one concrete prior data point in this bilateral track, and a restatement of it is considered one of the more likely outcomes of the September 2026 summit.
Are semiconductor or chip export controls part of these AI talks?
No. U.S. Commerce Secretary Howard Lutnick has said export controls are excluded from this dialogue and are handled, to the extent they are negotiated at all, through a separate channel. CASRAI covers export-control policy separately — see Compute Governance: Export Controls, FLOP Thresholds, and Chip-Level Oversight.
Is the Brookings-Fudan “Advancing Human Control of Military AI” report an official US or Chinese government position?
No. It is a track-two research proposal jointly published by Brookings (a U.S. think tank) and Fudan University (Shanghai) on September 17, 2026. It recommends that Presidents Trump and Xi agree that only humans should authorize AI-enabled cyberattacks on nuclear command-and-control systems, but neither government has adopted it as policy.
Does NIKOLAI have an element specifically for state-to-state AI dialogues?
Not exactly. NIKOLAI’s closest element is Commitment (N9), whose definition explicitly includes government bodies as possible committing parties, and whose existing crosswalk already includes state and multilateral examples. But a Commitment record requires a dated, tracked pledge from a named party — something this dialogue does not yet have. NIKOLAI documents vocabulary; it does not evaluate or predict the outcome of ongoing diplomacy.
