takeoff.watch

US intelligence says China has matched the US in frontier AI

Will the US intelligence community state in a public assessment that China has matched or overtaken the United States in frontier AI capability?

Chance by Q4 2031

12%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 14 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 14 Oct 2026 · Methodology v2.1

I put the chance at about 12% over five years. US spy agencies' latest public threat report says America leads in AI and calls China only the strongest competitor. The government's own testers still put Chinese models about four months behind, and political leaders in both parties prefer to say America is ahead. What would most change this is Chinese models clearly and lastingly matching the best American systems, for example if overall AI progress stalls and China catches up.

Where things stand (as of 14 Oct 2026)

What the intelligence community has said publicly. Its most recent public line is that the US leads. The 2026 Annual Threat Assessment (ATA), released 18 March 2026, refers to "the leading position of the U.S. in AI innovation." It calls China "the most capable competitor in the AI space," one that "aims to displace the U.S. as the global AI leader by 2030." The question lists both phrasings as non-qualifying.

  • The 2025 ATA described China's goal of becoming the "most influential AI power by 2030." The only lead it granted China was in niche areas such as surveillance and image recognition.
  • The Defense Intelligence Agency's (DIA) 2026 written testimony contains no comparison of frontier AI capability.
  • The most recent agency product is a joint advisory from the NSA, FBI and the cybersecurity agency CISA, dated 8 Sep 2026, which predates the question window. It says Chinese distillation lets Chinese models "continually close the technology gap." That wording assumes a US lead still exists.
  • I found no qualifying publication since the question opened on 16 Sep 2026.

What the government's own evaluators and outside measures show.

  • The Commerce Department's AI evaluation centre (CAISI, at NIST) is not part of the intelligence community. On 17 Sep 2026 it found the best Chinese model, Zhipu's GLM-5.3, about four months behind the US frontier in cyber capability. It also noted that unreleased US models could be stronger.
  • A joint US-UK assessment put Moonshot's Kimi K3 about six months behind in cyber capability.
  • The White House attributed K3's progress to smuggled chips and to distillation of US models.
  • Outside measures show a narrow gap. Epoch AI puts the gap at about four months, down from seven. Bloomberg Intelligence puts it at 3% on LiveBench (Oct 2026). Stanford's 2026 AI Index called the gap "effectively closed."
  • Most experts still judge that the US leads at the frontier, especially counting the restricted US models that are not publicly released (CFR, CSIS, NYT, Sep 2026).
  • Metaculus forecasters give roughly 28–40% that a Chinese model tops the US on Epoch's capability index by at least 2 points at some point before 2029. That is a momentary benchmark crossing, not an intelligence-community judgment.

Politics and institutions.

  • The administration's message is "whoever wins AI, wins." It explains Chinese gains as theft.
  • The Director of National Intelligence (DNI), Jay Clayton, was also named the administration's AI czar on 4 Oct 2026. His task force reports in about 120 days, but that is a White House product, not an intelligence-community analytic judgment.
  • ODNI has shrunk to about half its former size. That leaves the yearly threat assessment and congressional testimony as the main channels for a qualifying statement.
  • The intelligence authorization bills for fiscal 2026 and 2027 require AI reports to Congress, not public comparisons. H.R. 6909, which would require a National Intelligence Estimate on Chinese AI, covers bias and influence operations, has not been enacted, and would not require public release.

Reference class and model

When a rival clearly leads in a technology, US security agencies do eventually say so; hypersonics and surveillance AI are examples. For frontier AI, the intelligence community has never said this. The wording bar here is strict: "near-parity," "closing the gap" or "only months behind" do not count.

I built the forecast from three scenarios:

  1. The US lead persists (about 55–60%). Chance of a qualifying statement: about 1–2%, mostly from loose wording.
  2. Rough parity on public benchmarks, but the US keeps its edge in restricted models (about 25%). Chance of a statement: about 10%.
  3. Clear Chinese parity or lead (about 15–22%). Routes include a frontier plateau that lets followers converge, a US slowdown, theft of model weights, or a Chinese compute buildout. Chance of a statement: about 40–50%, given an annual publishing lag and political reluctance.

Together these give about 10–13% over five years.

Hazard shape. The chance of a first qualifying publication is concentrated in each March threat assessment and worldwide-threats hearing:

  • 2027 Q1: 0.8%. The US lead is current, and the administration says "we're winning."
  • 2028 Q1: 1.2%.
  • 2029–2031 Q1: 1.8% each, reflecting a possible new administration and more time for parity.
  • Other quarters: 0.2–0.45%, covering DIA testimony, hearing transcripts and ad hoc advisories.

I computed the cumulative series as 1 − ∏(1 − h) in Python. It reaches about 1.7% by end-2027, 3.8% by end-2028 and 12% by end-2031.

Strongest case against my number. Outside measures already show near-parity, and Chinese models lead usage on OpenRouter. There will be five ATAs in the window, and threat documents have reason to emphasise threats. A Trump administration ATA could say "China has caught up by stealing" without admitting a policy failure. Those points argue for something like 15–20%. On the other side, the agencies can see unreleased US models, the US has a large compute lead, and the US labs are increasingly using AI to speed their own research. Both parties also want to say the US leads. I land at about 12%.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has the U.S. intelligence community stated in a public assessment that China has matched or overtaken the United States in frontier AI capability?

Definitions.

  • U.S. intelligence community assessment: an official analytic judgment issued by ODNI, the National Intelligence Council, or a U.S. intelligence-community member agency. An agency's own assessment counts without requiring unanimity across the community. An attributed minority or dissenting agency judgment counts if it affirmatively meets the capability criterion.
  • Public assessment: an officially released unclassified assessment, an officially declassified assessment or its key judgments, or official written testimony or a government-published hearing transcript explicitly presenting an intelligence-community or member-agency analytic judgment. Personal opinions, policy recommendations, and statements merely quoting outsiders do not count.
  • China and the United States: the respective national AI sectors, including private and government developers, as identified by the assessment; this is not limited to state-owned systems or covered developers.
  • Frontier AI capability: the overall technical capability of each country's most advanced general-purpose AI models or systems. The assessment need not use the word “frontier,” but must clearly compare leading-edge model or system capabilities, rather than national AI influence, adoption, or industrial resources.
  • Matched or overtaken: an affirmative judgment that China's leading systems are at least broadly on par with, or more capable than, the leading U.S. systems at the time being assessed. Explicit descriptions such as “at parity,” “has caught up,” or “has surpassed” qualify when they concern these capabilities. A judgment qualified by an intelligence confidence level, or saying China “probably” or “likely” already has parity or superiority, counts; a mere possibility does not.

Resolves YES if a qualifying public assessment explicitly makes the comparative judgment above. A comparison between named models counts only if the assessment identifies them as representative of the countries' respective leading systems and concludes overall parity or superiority. No independent benchmark verification is required.

Resolves NO if no qualifying assessment has been published by the quarter-end. The following do not qualify:

  • Predictions that China will reach parity, conditional scenarios, Chinese ambitions, “cannot rule out” statements, or assessments that the gap is narrowing, small, or measured in months.
  • Calling China the “most capable competitor,” without an affirmative parity or superiority judgment.
  • Superiority only on a particular benchmark, task, application, or subset such as open-weight models, without an overall frontier-capability judgment.
  • Leadership only in publications, patents, talent, investment, compute, manufacturing, deployment, military use, price, efficiency, or market share.
  • Matching older U.S. models while remaining behind the U.S. frontier at the time assessed.
  • Leaks, anonymous-source reporting, unofficial transcripts, or statements by officials outside the intelligence community that are not themselves an official release of a qualifying intelligence assessment.

Resolving source. Official U.S. government publication of the assessment or testimony only, a narrower source rule than Public confirmation (A2). Sources include ODNI's Annual Threat Assessment of the U.S. Intelligence Community, National Intelligence Council National Intelligence Estimates and their declassified key judgments, and official analytic publications of intelligence-community member agencies. Official congressional hearing records may establish qualifying testimony. The 2026 Annual Threat Assessment, released March 18, 2026, is a baseline example, not a qualifying publication. Apply the Resolution window and Resolution date (A3) conventions; for declassified material, use the official public-release date, not the original classified issuance date. Apply the Observation period convention if the claim rests on a published statistic. Later disagreement or retraction does not reverse a qualifying YES.

Download all forecasts (JSON)