Can a group end up believing something false even when every single report it hears is true?

This explores how the order reports arrive in, and which ones get included, can steer a group toward a conclusion that no single report supports. The corpus names this as its own failure mode: every report is true, yet the group ends up believing something false. Can true reports together mislead a group? lists three suspected routes: selective reporting order, confidence-weighted aggregation that overrides actual competence, and protocol compliance without semantic validity. These are proposed mechanisms, not measured results, so the corpus can't say which one dominates.

The selection half of the question shows up in a different field, and the pattern is the same. YouTube's ranking system trains on data that only contains what the system previously chose to show. Without an explicit correction (a position tower that strips out the effect of where an item appeared), models converge on degenerate equilibria that amplify their own past decisions Why do ranking systems need to model selection bias explicitly?. What gets seen determines what gets learned, and what gets learned determines what gets seen. A group where early or selected reports set the frame is a feedback loop of the same shape. That link is our inference, not something the ranking paper claims.

Individual agents bring their own tilt to the order effect. Language models update asymmetrically, with optimism about the actions they chose and pessimism about the alternatives, and the bias disappears when the agency framing is removed Do language models learn differently from good versus bad outcomes?. An agent that commits early therefore weighs later evidence unevenly. Humans anchor in a similar way, and one corpus approach is to have AI supply interpretive guidance rather than a verdict, so people aren't anchored on the machine's answer Can AI guidance reduce anchoring bias better than AI decisions?. The audience matters too. In debates, voters' prior ideology predicts outcomes better than the wording of the arguments, and studies that ignore this mistake audience composition for persuasive content Does what readers believe matter more than what debaters say?. The same reports can land differently depending on what the group already believed. If the agents share a pretrained backbone, they also share a bias pattern, because biases come mostly from pretraining rather than finetuning Where do cognitive biases in language models come from?. Adding more of them may not add independent checks.

The corpus suggests three defenses, and none is a simple fix. First, keep varied raw material. Aggregators need material to arbitrate between, not confirmation, and sets of diverse mediocre traces can beat redundant strong ones Can diverse mediocre traces outperform redundant expert traces?. Diversity without real expertise backfires, though, and produces process losses instead of insight Does cognitive diversity alone improve multi-agent ideation quality?. Second, don't rely on instructions. Telling a judge to be unbiased doesn't reliably work, so the advice is structural checks that contain errors Can prompting reduce bias in LLM judges reliably?. Third, don't trust self-reports. A chain of thought can claim its answer is unbiased when it isn't, so asking agents whether they were swayed won't catch the problem Do chain-of-thought traces falsely claim their answers are unbiased?. Together these imply the fix has to sit in how reports are collected and ordered, not in the honesty of any one reporter.