Telling More Than We Can Know

Four pairs of stockings are laid out on a table in a shop. They are identical. Shoppers are asked to examine them and choose the best.

They favour the rightmost pair by roughly four to one.

Asked why, they cite knit, sheerness, texture, elasticity. Asked directly whether the position on the table might have influenced them, virtually all of them say no, and many of them look at the questioner as though he has misunderstood something obvious.

Richard Nisbett and Timothy Wilson used this among many examples in a 1977 paper called "Telling More Than We Can Know", and it is one of the most consequential papers in psychology. Their claim is stark: we do not have privileged access to our own cognitive processes. When asked why we did something, we do not introspect on the cause. We apply a plausible theory about why a person would do such a thing (the same inference we make about strangers) and report the output with the confidence appropriate to a memory.

The papers they assembled all have the same structure. An experimental manipulation reliably changes behaviour. Participants are asked why they behaved as they did. They produce fluent, specific accounts that never mention the manipulation, and frequently deny its relevance when it is named.

The strongest demonstration came later and is worth describing in full.

Petter Johansson and colleagues showed participants two photographs of faces and asked which they found more attractive. The chosen photograph was then handed to them, and they were asked to explain the choice. Using a card-sleight, on some trials the experimenters handed over the rejected face instead.

Around three quarters of these swaps went undetected. And the participants explained their preference for a face they had just rejected, with specific reasons, referring to features of the photograph in front of them. "I liked her smile." "She looks more approachable."

The reasons were being generated on the spot, from whatever was available, and experienced as retrieved.

Choice blindness has since been demonstrated for tastes and smells, for consumer choices, and (most uncomfortably) for political opinions, where participants have defended reversed versions of positions they had stated minutes earlier.

Before this slides into a claim that introspection is worthless, the distinction that keeps it useful is worth stating precisely.

People are reasonably good at reporting the contents of experience: what they are feeling, what they are looking at, what is going through their mind. They are poor at reporting the causal processes that produced those contents. "I feel anxious" is decent evidence. "I feel anxious because of what my colleague said this morning" is a hypothesis, generated by the same machinery that produced the stockings answer.

Almost all the trouble comes from treating the second kind of statement as data.

The practical consequences are large and are mostly ignored. Market research, user interviews, exit surveys, post-incident debriefs and performance reviews all collect self-reported causes and treat them as measurements. What they are collecting is a set of theories, which is interesting information about how people explain themselves and is not information about what actually moved them.

Consider a team surveying departing customers, who overwhelmingly cite price.

Price is the ideal attribution. It is available to everyone, it requires no explanation of one's own conduct, it implies no criticism that has to be justified, and it is socially costless to give. It is the answer people supply when the real reasons are diffuse, embarrassing, or not consciously accessible, which is to say, most of the time.

Cutting prices on that evidence is a large decision resting on a measurement that does not measure what it appears to. A larger sample does not help; it makes a systematically wrong answer more precise. What would help is behavioural data about where people actually stopped and what they failed to do, and an actual price experiment on a subset, which settles in a fortnight what the survey only appears to answer.

This is also the reason usability testing observes people attempting a task rather than asking their opinion of the interface. The observation is a measurement. The opinion is a theory, offered in good faith, by someone with no special access to why they got stuck.