Nothing false, still misleading: when an AI summary leaves a group's favorite unnamed
Rob Gordon
Submitted to AI x Epistemics Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
AI mediators are starting to sit between people who decide together, and they are mostly judged by what their text says or how people rate it. In a live game, four Prolific strangers privately pick one of four options for a shared decision and write why. An AI that sees the reasons but never the picks writes one text for the group. Each member then guesses the group's favorite (its top pick: the most-picked option, ties included) and wagers on it from a shared pool that carries across rounds and pays out as real money.
In post hoc analyses across ten groups (40 people), the text named an option whenever a supporter had named it, but only 5 of 19 times when supporters argued for it without naming it. In three of 30 scenarios, from two groups, a text with no false statement left the favorite unnamed. Members whose own pick was a top pick kept it 83% of the time when the text named it, and 1 time in 9 when it didn't. On those three scenarios people got 1 answer of 12 right; a model shown each member's screen got 11.
The planned test, adding each member's exact words to the text, showed no difference in accuracy or wagers (about 75% right in both), though only a very large gap could have shown, and the wagers tracked the pool rather than confidence. In 300 re-runs of the real reasons, omissions varied with the reasons, not the format.
Reviews
This is an ambitious project which is in function more of a pilot project than a findings paper. The central question is what happens when AI intermediates between people who have to come to a decision and how does it affect what they believe about one another? Truly an interesting thought experiment and an even more interesting test case. But there aren’t real findings from this pilot, and little evidence that points to much. I appreciated that there were real people in this project, though would have liked some human coding of the results (rather than only AI).
Cite this project
@misc{gordon2026nothing,
title = {{Nothing false, still misleading: when an AI summary leaves a group's favorite unnamed}},
author = {Rob Gordon},
year = {2026},
month = nov,
note = {Submitted to AI x Epistemics Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/nothing-false-still-misleading-when-an-ai-summary-leaves-a-groups-favorite-unnamed-z76pgb16h5}},
url = {https://apartresearch.com/sprints/projects/nothing-false-still-misleading-when-an-ai-summary-leaves-a-groups-favorite-unnamed-z76pgb16h5}
}Browse all projects