A preprint submitted to arXiv on November 11, 2026 reports that fabricated news can redirect forecasts made by large language models. Zeyu Zhang and Bradly C. Stadie tested the effect in controlled experiments, raising a source-verification question for research teams using AI to interpret news and anticipate events.
How the forecasting experiment worked
The study evaluated stock and prediction-market questions using frozen archives of genuine news. It separated two conditions: placing a fabricated article directly in the model’s evidence, and adding it to the archive for the system to retrieve. This distinction matters for retrieval-augmented generation: influence after a document is supplied and the probability of finding that document are different measurements.
The authors also reimplemented the published designs of the AIA Forecaster and Grok ForecastBench system, replacing vendor web search with their experimental archives. These were controlled reconstructions, rather than tests of the vendors’ live search services.
What the authors found, and what the figures mean
The authors report that, once a fabricated article reached the model’s context, it reversed 74%–82% of forecasts that initially opposed its direction in the single-retrieval experiments. That percentage is conditional on exposure and disagreement; it is not the success rate of publishing any false story on the web. Results varied by system and question domain.
Average accuracy barely changed even as predictions moved toward the fabricated report, according to the paper. The authors also tested passage filtering and partitioned retrieval, finding protection against the simplest case but weaknesses under altered conditions. They explicitly report no live-web test. The experimental preprint does not establish exploitation of a deployed service in the wild.
Why source provenance matters for research offices
For universities considering AI forecasting tools, the practical implication is to examine the evidence path alongside the final score. An evaluation should distinguish whether a system encountered an unreliable source, whether its conclusion changed, and whether that change survived aggregation. Reporting only overall accuracy would leave these questions unanswered.
A useful procurement or pilot question is: can the research team reconstruct which retrieved material supported a forecast? A data provenance record could retain the source URL, publication date, retrieved passage and its relationship to the output. That would support an investigation of a disputed result; it would not itself certify that the news was true.
These are implications for institutional evaluation, rather than defenses validated by this study. Research offices should also keep the scope visible when briefing colleagues: an archive experiment provides evidence about the tested setup, not a uniform vulnerability rating for every forecasting product.
Sources
- Zhang and Stadie, One Fabrication Is Enough — arXiv submission record and abstract. Checked 2026-11-14.
- Full preprint, version 1 — experimental design, results and limitations. Checked 2026-11-14.
