A top journal retracts a headline result credited to AI

Current status (as of 14 Oct 2026)

• Nothing qualifies yet. Crossref and the Retraction Watch database list two retractions in these five journals since the window opened on 16 Sep 2026. Neither involves AI:
  • Nature, 2 Sep 2026, "Warming accelerates global drought severity". This is before the window anyway.
  • Nature, 1 Oct 2026, the CCR5 ribosomal-frameshifting paper.
• No AI-credited paper in these journals looks close to retraction. I checked current expressions of concern and editor's notes:
  • Science's June 2026 concern about the "DebunkBot" conspiracy-beliefs paper would not count. The authors say a reanalysis upholds the findings, and the finding is about how well an AI performs, which the definition excludes.
  • The high-profile AI disputes have ended in corrections, not retractions:
    • The A-Lab autonomous-lab paper got an author correction in Jan 2026. Its headline dropped from 41 to 36 compounds and "novel" was removed from the title.
    • Google's AlphaChip paper got an addendum.
    • DeepMind's GNoME crystal paper faces calls to retract it (C&EN, Dec 2025) but has no editor's note.
    • Microsoft's MatterGen paper has been criticised but not corrected or retracted.

Base rate (computed)

• Retraction volume. From the Retraction Watch/Crossref CSV (gitlab.com/crossref/retraction-watch-data), the five journals retracted about 11 papers a year in 2016–2025. There have been about 13 so far in 2026, a high.
  • Median time from publication to retraction is about 2.7 years; 75% happen within 6.6 years.
  • About 64% of retractions are tagged for error or fabrication. Reading actual notices, about half would meet the strict "fabricated or wrong" wording. The CCR5 notice ("artefact… undermine the main conclusion") would pass; "lost confidence" notices would not.
  • That gives roughly 6 notices a year that could qualify.
• Precedent. In more than 10 years there is only one borderline case: Poore et al. (Nature 2020, retracted June 2024), a machine-learning cancer-microbiome study. Its notice says only that "some findings… no longer supported."
• Size of the AI pool. PubMed counts of AI-keyword research articles rose sharply:
  • Nature: 41 in 2022, 101 in 2025, 109 by Oct 2026.
  • Science: about 45 a year; Cell: about 30 a year.
  • I hand-classified 2025 Nature titles: about 30–35% credit a headline result to AI (designed, predicted or inferred, not just performance claims).
  • AI-credited papers are therefore about 3% of flagship output now and probably 5–7% by 2030. Nature now publishes AI-agent discovery papers such as Robin, which says "all hypotheses… analyses and data figures… were produced by Robin."

Model

Expected events = qualifying retractions per year × AI-credited share of papers at risk × relative risk × chance the resolver accepts the case.

• The model reproduces history: about 0.3–0.65 expected cases for 2016–2026, against about 0.5 observed.
• Looking forward, the central parameters give about 0.8 expected events by end-2031 (about 56%). The range runs from about 28% (pessimistic) to about 82% (optimistic).
• I shade this to about 0.61 expected events, or about 46%. Reasons: journals and companies have preferred corrections for AI overclaims; the AI-credit and wording tests are strict; and nothing clearly qualifying has ever happened.
• The hazard rises each year as the AI-credited stock grows: about 1.8% in Q4 2026, then 2–4% a quarter by 2030–31.

Paths

• To YES:
  • An error retraction of an ML-based estimate paper, for example in earth or climate science, where Nature retracts about one error paper a year.
  • A fabricated or flawed lab validation in an AI-design or AI-agent paper.
  • A coding bug or data leakage that invalidates an AI-inferred result. Open code makes this kind of fast retraction possible, as with Science's sponge paper (3 months) and genome-imaging paper (10 months).
• To NO: corrections or addenda instead of retraction, vague "loss of confidence" notices, or the AI not being explicitly credited.

Strongest objection: essentially no precedent, and every high-profile AI dispute so far has been corrected rather than retracted. I partly priced this in by shading well below the model's central estimate.

Sources: Retraction Watch CSV; PubMed E-utilities counts; Crossref API; Poore retraction note (nature.com/articles/s41586-024-07656-x); A-Lab correction (nature.com/articles/s41586-025-09992-y); C&EN 16 Dec 2025 and 29 Jan 2026; Retraction Watch 11 Jun 2026; Nature drought retraction (s41586-026-11027-z) and CCR5 retraction (s41586-026-11149-4); Robin (s41586-026-10652-y); AlphaChip on Wikipedia.

Question. By the last day of each quarter, has Nature, Science, Cell, the New England Journal of Medicine, or The Lancet retracted a paper whose headline result the authors had credited to an AI system, with the retraction notice saying that result was fabricated or wrong?

Definitions.

• Journals: the five flagship journals only: Nature (nature.com/nature), Science (science.org/journal/science), Cell (cell.com/cell), The New England Journal of Medicine (nejm.org), and The Lancet (thelancet.com/journals/lancet). Sister journals, other journals from the same publisher, and preprint servers do not count.
• Paper: a published contribution reporting original scientific or clinical findings, including a research article, research letter, brief report, or case report. News, editorials, reviews, and commentaries do not count.
• Headline result: a principal finding explicitly stated in the paper’s title or abstract. For a paper without an abstract, it must be a principal finding stated in its opening summary or concluding paragraph. A supporting detail alone is insufficient. If there are multiple principal findings, one qualifying finding suffices.
• Credited to an AI system: before publication of the retraction notice, the authors explicitly stated that an AI or machine-learning system produced, discovered, inferred, predicted, designed, or proved the headline result. The attribution must appear in the paper or its supplements, or in a publicly dated author-authored statement identifying the paper and result. The original attribution may predate the question’s opening date. AI need not be an author, autonomous, generative, or from a covered developer. Merely using AI for writing, translation, illustration, formatting, routine preprocessing, or literature search does not qualify. A finding solely about an AI system’s measured performance does not itself constitute a result credited to that system.
• Retract: the journal publishes a formal notice withdrawing the entire paper from the scientific record, whether initiated by the authors or the editors. A formally labelled retraction-and-replacement counts if the retraction otherwise meets these criteria.
• Fabricated or wrong: the notice affirmatively states that the qualifying headline result was invented, falsified, erroneous, invalid, or not supported because of an identified error or fabrication in its underlying evidence or analysis. The exact words “fabricated” or “wrong” are not required. A statement of uncertainty or loss of confidence alone is insufficient.

Resolves YES if a qualifying journal retraction notice identifies the paper and explicitly applies the fabrication or error finding to the same headline result previously credited to AI, or to evidence or analysis that the notice explicitly identifies as invalidating that result. The notice need not repeat the earlier AI attribution. It need not blame AI for the failure: human misconduct, erroneous inputs, or human analysis errors can qualify. Author disagreement does not prevent resolution.

Exclusions (resolve NO). Corrections, expressions of concern, allegations outside the notice, failed replications without a qualifying retraction, and notices stating only that results cannot be verified do not count. Retractions solely for plagiarism, consent violations, authorship disputes, undisclosed AI use, or image manipulation do not count unless the notice also establishes fabrication or error in the qualifying headline result. AI use first disclosed in the retraction notice does not satisfy prior credit.

Resolving source. The formal retraction notice on the relevant journal’s official website or in its official issue is the exclusive resolving source, narrower than Public confirmation (A2). The original paper, supplements, and publicly dated author-authored statements establish the prior attribution and headline result. Apply the Resolution window and Resolution date (A3) to the first qualifying notice or substantive update to it; a later news story or unchanged republication of an older notice does not count. Earlier supporting documents may establish attribution but cannot themselves trigger resolution.