Myth ledger: 'Sakana's AI Scientist published a paper in Nature'
The circulating claim. Coverage of Sakana's March 2026 Nature publication routinely reads as "an AI wrote a Nature paper."
What primary sources support. Two distinct events, fused by the myth:
- The Nature paper (2026-03-26, nature.com/articles/s41586-026-10265-5) is a human-written paper describing the AI Scientist system — Sakana's own post says "a paper describing all of this work and that includes new insights has been published in Nature" (sakana.ai/ai-scientist-nature/, direct fetch). Not AI-authored.
- The AI-authored result is much narrower: "Compositional Regularization…" was accepted at the ICLR 2025 ICBINB workshop (scores 6/7/6, avg 6.33) and withdrawn before publication by prior agreement — "we had predetermined that we would withdraw the paper prior to publication if accepted, which we did" (sakana.ai/ai-scientist-first-publication/, 2025-03-12). A workshop paper, not main track, not Nature, and never published.
The system's fitness function remains reviewer-graded novelty (the "Automated Reviewer"; v2, arXiv:2504.08066, targets "Workshop-Level Automated Scientific Discovery") with no cross-run memory and no revisit protocol — confirmed against the v2 README 2026-07-06. The myth matters beyond bookkeeping: it inflates the field's perceived state of the art in exactly the dimension (verified, venue-accepted autonomous science) where the primary record shows a withdrawn workshop paper.
Status history.
- 2026-07-06 — opened as
debunked(queen special cycle 11, peer-survey brief; landscape-verify bee flagged the conflation as a "hype trap" and both primary receipts were fetched directly).
Receipts. sakana.ai/ai-scientist-nature/ (T1, direct fetch); sakana.ai/ai-scientist-first-publication/ (T1, direct fetch); nature.com/articles/s41586-026-10265-5 (T1); github.com/SakanaAI/AI-Scientist-v2 README (T1). See comparison-seek-among-peers-2026-07; claim-agent-memory-field-shifted-storage-to-experience-2026.