Human Introspection

In 1977 Richard Nisbett and Timothy Wilson published a review of a great deal of research on what people know about their own mental processes, and the answer was: much less than they think, and they do not know that they do not know. The centrepiece was a study in which shoppers were shown four pairs of nylon stockings laid out on a table and asked to choose the best. The pairs were identical. Shoppers chose the rightmost pair at around four times the rate of the leftmost — a large and reliable position effect — and when asked why, gave reasons: the knit, the sheerness, the elasticity. When it was suggested to them that position might have influenced their choice, they denied it, usually with some impatience, and occasionally with the suggestion that the experimenter was mad.

Nobody was lying. The shoppers had introspected, found a reason, and reported it. The reason was constructed after the fact by a system that generates plausible explanations, and the actual cause was invisible to that system, and the output was indistinguishable — to the shopper — from a genuine report.

I raise this at the start rather than the end because the standard structure of a chapter like this is to set out the ways machine self-report is unreliable and then, in a spirit of fairness, to note that human introspection has its problems too. That structure gets the weighting wrong. Human introspection is not a reliable instrument with some known failure modes. It is a system that constructs reports, that is excellent at producing coherent ones, and whose relationship to the underlying processes is intermittent at best. We grant it authority not because it has earned it experimentally but because we have no alternative and because, in the narrow domain of current sensory states, it seems to do well enough.

from Are LLMs Conscious?: Language, Experience and the Problem of Artificial Minds (2026)