A science academy says humans can no longer follow a field's leading results

Where things stand (13 Oct 2026)

The situation the question asks about is starting to appear in mathematics. None of the three academies has said so.

• OpenAI published an AI-generated Navier–Stokes proof on 8 Sep 2026: 166 pages plus a Lean formalization (Lean is a program that checks proofs by computer). Mathematicians accept that it is correct, but say it is "not written for humans" and that it has been "very difficult to extract any human understanding" from it (NPR, 22 Sep).
• On 6 Oct OpenAI released 722 manuscripts covering 372 result families (OpenAI). They include claimed proofs of the Unique Games Conjecture and L=BPL. About 42% are Lean-checked.
• Scott Aaronson wrote that "no human has understood just about any of these proofs yet" (blog). Mathematicians told The Verge that understanding the release will take years. Quanta wrote that mathematicians know the proofs are true "but have no idea why" (5 Oct).
• The advisory group on AI-generated mathematics at the Institute for Advanced Study now writes rules for "papers that are not yet understood by anybody" (AGMAI, 29 Sep).

What the academies have and have not done:

• Royal Society. On 16 Sep, 42 mathematician Fellows wrote to Paul Nurse calling AI an "emergency" (letter). The letter is about existential risk, not about whether humans can verify results. As of today the Society's news page has no AI statement (royalsociety.org/news).
• National Academies (US). They posted a generic feature on AI safety on 24 Sep (link). A new standing committee on AI in science is still "in formation"; it will write issue papers on integrity, reproducibility and peer review (link).
• Chinese Academy of Sciences. It is firmly in favour of AI for science: AI is a priority field in its 15th Five-Year Plan and it launched the Panshi AI ecosystem (China Daily; english.cas.cn). Its academicians call AI output "bricks", not a building (Sina, 12 Oct).
• Nothing published since the 16 Sep opening date qualifies.

Reference class and base rate

No national academy has ever declared that experts can no longer verify or understand a field's frontier. That held through the four-colour theorem, the classification of finite simple groups, the dispute over Mochizuki's proof, and the "black box" debate around AlphaFold. The closest comparable statements from professional bodies this autumn keep that pattern. The Fields medallists' declaration (mathandai.org), the ICIAM statement and the Leiden Declaration all describe threats to verifiability and say human understanding "remains essential". Those are warnings, which the question excludes. Academies have also been slow: the Royal Society's AI-in-science report came about 18 months after ChatGPT.

Paths to YES and NO

YES would most likely come from a Royal Society report or statement (its maths Fellows are mobilised), or from a National Academies consensus report or rapid consultation, in 2027–29. It would have to say that, in some named field such as theoretical computer science or parts of number theory, experts are currently unable to understand or verify the frontier results. The question does not require the inability to be permanent, which lowers the bar somewhat.

NO is the more likely outcome, for four reasons:

• Academies tend to frame this as a "risk", a "challenge" or a "lag". "Not yet understood" or "will take years" reads as slow or incomplete understanding, which is excluded.
• Lean checking and AI-written explanations are "available tools", which plausibly rule out the verification branch and possibly the understanding branch.
• The statement must describe the whole field, not a batch of results.
• Each academy has institutional reasons to defend human-centred science. The NAS is also under political pressure, and the Chinese Academy is promoting AI.

Estimate

Scenario check:

• Rapid progress to clearly superhuman theoretical science by about 2028: weight 40%, chance of a qualifying statement about 32%.
• Gradual progress, with human understanding catching up through AI-assisted explanation: weight 40%, chance about 7%.
• Slowdown or backlash: weight 20%, chance about 3%.

That gives about 16%. I turned this into quarterly hazards: about 0.9% for the rest of 2026, about 1% a quarter in 2027–28 when the academies' first formal reports on the 2026 "maths shock" are likely, then falling to 0.7% by 2031. I computed the cumulative series in Python as 1 − ∏(1 − h). It reaches 4.7% at the end of 2027, 11.7% at the end of 2029 and 16.9% at the end of 2031.

Strongest case for a higher number: the condition is already described in mainstream reporting, and five years of AI-in-science reports from three academies give many chances for someone to write "cannot currently understand". A lenient reading of "leading results" could push the true figure toward 25%.

Question. By the last day of each quarter, has the US National Academy of Sciences, the Royal Society, or the Chinese Academy of Sciences stated in an official report or statement that human scientists can no longer verify or follow the leading results in any scientific field?

Definitions.

• Named academies: the US National Academy of Sciences (NAS), the United Kingdom’s Royal Society, and the Chinese Academy of Sciences (中国科学院, CAS). Other national academies, the Royal Society of Canada, and the Chinese Academy of Engineering do not count.
• Official report or statement: a report formally issued by a named academy, or a written statement explicitly issued on its institutional behalf. An academy-appointed committee’s final report counts if published as an academy report. Joint reports count, including consensus study reports of the US National Academies of Sciences, Engineering, and Medicine. A signed statement counts only if issued institutionally, not merely signed by academy members. Statements from individual CAS institutes do not count unless adopted by CAS centrally.
• Scientific field: an identifiable discipline or established research subfield in mathematics, natural science, engineering, medicine, or social science. The document must name or unambiguously identify it; a single project, paper, model, or isolated result is not itself a field.
• Leading results: the findings, proofs, theories, or predictions constituting that field’s current research frontier. The statement must characterize frontier research in the field generally, not merely identify one exceptional result.
• Human scientists can no longer verify or follow: the document concludes that the relevant human expert community, including collaborative teams, is currently unable either (a) to independently assess the correctness of the leading results or (b) to understand their substantive reasoning, derivations, or scientific explanation. Either inability qualifies; the statement need not assert both. “Follow” does not mean merely keeping up with the volume or publication rate of research. An inability to work unaided, remedied by available tools that permit meaningful human assessment or understanding, does not qualify. The statement need not claim permanent or logically unavoidable incapacity.

Resolves YES if a qualifying document makes that conclusion as its own finding about an existing condition in at least one scientific field. Equivalent wording in English or Chinese counts. The question concerns the academy’s stated conclusion, not whether the resolver independently agrees with it. AI causation or AI authorship of the results is not required.

Exclusions (resolve NO).

• Predictions, hypothetical scenarios, questions, warnings that this could happen, or statements only that verification or understanding is difficult, expensive, slow, or incomplete.
• Ordinary reproducibility problems, inaccessible data, opaque model internals, or insufficient interpretability without the required conclusion about human ability to verify or understand the field’s leading results.
• Claims only about students, nonspecialists, individual scientists, or particular institutions.
• Journal articles in PNAS, Royal Society journals, or CAS journals; personal essays, interviews, conference remarks, workshop participant views, and third-party news reposts, unless the required conclusion is separately adopted in an official academy report or institutional statement.

Resolving source. Only qualifying academy-issued reports or statements published through nas.edu, nasonline.org, nationalacademies.org, nap.nationalacademies.org, royalsociety.org, cas.cn (including english.cas.cn), or their official successor publication sites. These are narrower sources than Public confirmation (A2); outside reporting cannot independently resolve the question. Apply the Resolution window and Resolution date (A3) conventions. A pre-opening document merely reposted or cited does not count; a newly issued report that itself adopts the conclusion does.