The reason more information can lower accuracy

Forecasters are given five relevant facts and asked to predict an outcome. A second group receives the same five facts plus fifteen more of marginal relevance. The second group is markedly more confident. Their accuracy is no better, and on some tasks it is worse.

What everyone sees

More information is assumed to be strictly better, on the reasonable ground that a fact can be ignored if it is not useful. The assumption fails because irrelevant material is not ignored: it is averaged in, and averaging a strong signal with weak ones moves the estimate towards the middle.

What is actually happening

Nisbett and colleagues described the dilution effect, in which non-diagnostic information weakens the use of diagnostic information. Oskamp found that clinicians’ confidence rose substantially with additional case material while accuracy did not. Work on lens-model judgement, following Dawes, shows that simple models using few strong predictors frequently outperform richer ones, because added variables contribute more noise than signal.

Why it stays hidden

The effect hides because confidence is the only feedback available at the moment of judgement. Accuracy arrives later, if at all, and is rarely paired with a record of what was known. Requesting more information also looks diligent, so the behaviour is socially rewarded independently of its effect, and no institution keeps score of the version with fewer inputs.