Nature reported on August 6, 2026 that a machine-learning model built to predict molecular boiling points kept disagreeing with a widely used chemistry reference database — and that when theoretical chemist Sebastian Pios traced the numbers back to the original literature, the database turned out to be the party at fault. The reported errors include a typo in an old paper and a measurement mistake from roughly a century ago. No paper documenting the verification has been published as of August 30, 2026, and neither the number of errors nor the database’s name has been disclosed. This video explains how an error survives for decades inside a trusted number, why AI is simultaneously finding errors and manufacturing them (fabricated citations), why AI audit performance depends heavily on what you ask it to check, and the real bottleneck: the party that finds an error and the party with the authority to correct the record are different people. It also covers how Japanese researchers reacted — largely by debating accountability rather than the audit itself.
Full article: https://sekahan0623.com/en/global-reactions-en/ai-audits-science-errors-japanese-reactions/
X (Twitter): https://x.com/sekahan_0623
This program is produced with synthesized narration (Voice: Google Cloud Text-to-Speech (Chirp 3: HD)).
