False Assurance
Mohit Goyal, Ananya Bhattacharya · Team Two Bits
Submitted to AI x Epistemics Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Community Notes has AI note writers, and AI agents read the notes. The platform's quality gates check whether a note looks sourced (links load, raters tick "cites high-quality sources", readers see "Rated helpful"), not whether the cited page supports the claim. A source-check tool could tell readers, but its verdict can itself be wrong. We gave three chat LLMs (Claude Opus 5.5, Claude Haiku 5.5, Gemini 3.8 Flash) 80 Helpful notes whose source support two LLM judges agreed on and human labels confirmed for 78, with right or wrong verdicts, with and without the checked passage (1,920 answers). With a verdict only, the readers inherited the checker's errors: the verdict moved belief by 35–62.5 percentage points and real support by −5 to +7.5, and accuracy fell to 21–30% when the verdict was wrong. The Claude readers mostly withdrew belief as "unsure", not "no". Adding the checked passage cut dependence on the verdict by 26–51 points in all three models, and a wrong verdict shown with its passage left readers no less accurate than no tool. The protection was one-sided: readers rejected false alarms but rarely false assurances, and under a false assurance an unrelated passage worked as well as the real one. A page-level check of 200 paired notes found that 52–56% (human labels) to 68% (LLM judge) of readable cited pages do not fully support the claim, with no detectable AI–human difference. Verification signals for AI readers should carry their evidence and treat "supported" as the risky verdict.
Reviews
This project attends to a timely question that affects our information ecosystems and human levels of trust by evaluating how AI’s “trust” what other AI say. The finding that AI readers mostly just believe an AI source-check tool has implications for the kinds of information humans eventually see and use to inform opinions or act on—even when the information is ungrounded in citations.
While there are methodological limits to the study (and the authors disclose them), this is a solid submission that should be studied further.
My largest difficulty was the presentation with 2 studies presented in the same short paper—the abstract beginning with the second one and the paper opening with the first. It was hard to follow some of the conditions labeled in the tables. In this sense the structure and writing clarity took away from the findings.
Cite this project
@misc{goyal2026false,
title = {{False Assurance}},
author = {Mohit Goyal and Ananya Bhattacharya},
year = {2026},
month = nov,
note = {Submitted to AI x Epistemics Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/false-assurance-5kn9n4xzcv}},
url = {https://apartresearch.com/sprints/projects/false-assurance-5kn9n4xzcv}
}Browse all projects