US–China frontier AI agreement

This question asks how far the US and China get on paper about frontier AI — not talks, not one-sided announcements, but a written document both governments actually put their names to. Three rungs: banning specific AI uses (L1), committing to pre-release safety testing with some exchange of results (L2), and constraining the pace or scale of development itself, including training-run notification (L3). Right now none of these exists. By the end of September 2026 we put L1 at 8%, L2 at 2%, L3 at essentially zero. Five years out, by September 2031, the numbers are 53%, 27% and 12%.

The setting matters. A Trump–Xi state visit lands on September 24, with AI on the agenda, and a Treasury-led AI safety dialogue has been floated for mid-September — though the White House has at points denied one was scheduled, and reporting suggests the AI track may simply be folded into the economic meeting. The May 2026 Beijing summit produced no joint statement at all, only a US fact sheet and a Chinese confirmation that the two sides would talk. That format problem is central: this question needs co-issued text, and recent practice has been duelling readouts. Meanwhile both capitals are talking past each other — Bessent says nobody can pause because China won't, Trump says whoever wins AI wins, and Beijing calls slowdown proposals a containment playbook.

So why not near zero? Because L1 is cheap. A line saying neither side will use AI to help build biological weapons or attack civilian infrastructure costs almost nothing, is already implied by existing treaties, and is exactly the kind of deliverable summits produce — the 2015 Obama–Xi cyber-theft commitment is the template. Over five years and a dozen leader meetings, plus a likely change of US administration in 2029, that gets a real chance. L2 and L3 are different animals: swapping pre-release test results with a strategic rival has no precedent anywhere, and pacing constraints run straight into an unsolved verification problem that neither side has proposed a fix for.

The biggest open question is whether an AI incident severe enough to change both governments' minds arrives — that is the main path to the higher rungs. The nearest-term mover is simply whether September 24 yields joint AI text with a named prohibition, which would settle L1 outright, or nothing at all, which would pull the near-term numbers down.

Question. For each quarter, what is the highest level L such that an instrument at level ≥ L has been in force at any point on or before the last day of the quarter?

Instrument. A written document signed or jointly issued by both the US and Chinese governments. Track 1.5/2 dialogues, unilateral executive orders, and statements by only one side do not count.

Levels (cumulative; higher levels imply lower).

• L0 — Dialogue or joint statement of principles with no specific new obligations.
• L1 — Both governments commit to specific prohibitions on AI uses (e.g., AI-enabled biological weapons development). Must be a new commitment; the 2024 statement on human control of nuclear-use decisions does not count.
• L2 — Both governments commit to pre-release testing of frontier models against named risk categories, with some exchange of results.
• L3 — Both governments commit to constraints on the rate or scale of frontier development: compute thresholds, training-run notification or licensing, limits on automated AI R&D, or a general pause or pacing.

Resolving source. Official text published by the White House, State Department, Treasury, China's Ministry of Foreign Affairs, or Xinhua.

Forecast horizons. Every quarter-end from 2026-09-30 through 2031-09-30.