A preprint finds that a single fake news article in an AI forecasting system's search results can reverse 74% to 82% of the forecasts that originally opposed the article's position, while average accuracy barely changes.

Concept illustration: a frame topped by a black bar lettered ARXIV holds two side-by-side scenes on one shelf line; on the left, a stack of beige pages with grey text lines, propped by a black bookend, stands beside an open laptop with a blank screen whose speech bubble shows a black thumbs-up. On the right, a similar stack of beige sheets with a black bookend has a red sheet lettered FAKE NEWS with a white thumbs-down in front, beside an identical laptop whose speech bubble shows a black thumbs-down. Printed on the image: 74%–82% Share of AI forecasts opposing a fake article that the one article reversed, per preprint authors.

According to the paper's abstract, using two retrieval-side defenses together changes 29% of normal judgments and discards 39% of genuine news, and these are the authors' self-reported results. Forecasting systems of this kind search the web before making a call. The paper doesn't name the two forecasting systems tested. Authors Zeyu Zhang and Bradly C. Stadie assume the attacker knows only the forecasting question and has no internal access. According to the abstract, given the architectures of two deployed leading forecasting systems, those systems would retrieve the fake article about 90% of the time. Each defense used alone blocks a single fake article, but one change to the attacker's prompt gets past it. Used together, the two defenses block 4 fake articles, and the attacker gets through by submitting 16 rewritten copies of the article. Because average accuracy barely changes, average accuracy alone won't reveal targeted manipulation. No one else has repeated the results (arXiv 2611.15102, 2026-11-11). How we checked →

Source and checks

This item ran in the brief of Nov 13, 2026, alongside 16 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "A preprint finds that a single fake news article in an AI forecasting system's search results can reverse 74% to 82% of the forecasts that originally opposed the article's position, while average accuracy barely changes.." https://secondsource.io/en/research/2026-11-13-1/. Accessed access date.