Limits of introspection

Why can people explain a choice fluently and confidently while having no access to its real cause?

Ask someone why they hired that candidate, or chose that jam, and the answer arrives immediately and without strain. It has reasons in it. It sounds like a report.

But a report of what? We assume the answer is retrieved — that the mind looked inward, found the cause, and read it out. Suppose instead that no such retrieval is possible. What would the answers look like then? That is the question worth sitting with, because the surprising part of the evidence is not that people are sometimes wrong about themselves; it is that being wrong feels exactly like being right.

Start with what an accurate report would require. To read out the cause of a choice, introspection would need a channel carrying information about the causal process — not about its output. Do we have any reason to think such a channel exists? Consider how much of your processing you can observe. You cannot watch your visual system infer depth, or your language system select the next word, yet both are causing things constantly. Introspection seems to be positioned downstream: it receives products, not procedures.

If that is right, then introspection faces an information limit rather than a motivation problem. The honest person and the liar are in the same position, because the missing item is not being withheld — it was never delivered.

So what does the answering system do when it is asked a question it lacks the input to answer? It could say "I do not know." Does it? Nisbett and Wilson's 1977 review is the classic demonstration that it does not. In their shop display, people examined identical pairs of nylon stockings laid out in a row and chose the rightmost far more often than the leftmost — a position effect. Asked why, they cited knit, sheerness, elasticity. Told that position might have mattered, they denied it, sometimes with mild concern for the questioner's sanity. The influence was real, the reasons were sincere, and the two were unrelated.

Push further: could the system be caught explaining a choice it did not even make? Johansson, Hall and colleagues managed exactly that in 2005. Participants picked the more attractive of two faces; using a sleight of hand, the experimenters handed back the rejected card and asked why it had been chosen. Most such swaps went undetected, and participants then argued for the face they had just rejected — "she's radiant, I'd have preferred her in a bar" — sometimes citing features present only in the card they had actually picked. Confabulation of that kind is not rare pathology; it is the ordinary output of a system asked to explain with insufficient data.

Now the mechanism question. If the reasons are not retrieved, where do they come from? The best-supported answer is that they are constructed from what is available: the option chosen, the situation, and a stock of culturally shared theories about what causes what. If a plausible reason is available, it is offered. Which predicts something testable and slightly eerie — an observer who knows only the situation, and nothing of your inner life, should be able to predict your explanation about as well as you produce it. That prediction is roughly what has been found for the classic effects.

Gazzaniga's work with split-brain patients shows the same machinery in stark form. Instruct the mute right hemisphere to stand up and the person stands; ask why, and the speaking left hemisphere — which never received the instruction — answers at once: "I wanted a drink." He called it the interpreter: a subsystem whose job is to produce a coherent account of behaviour from whatever information reaches it, and which does not report the shortfall when the information is missing.

Where should we stop? Not at "introspection is worthless." It is reliable about some things — the current contents of experience, the fact that you are in pain, an argument you deliberately worked through on paper. The limit bites on causes, especially the causes of preferences, judgements and evaluations. And the strength of the 1977 conclusions has been debated since; the safer statement is that introspective access to causes is partial and unmarked, not that it is absent.

The most common misreading is that this shows people are unaware of their choices. They are not. You know perfectly well what you chose, and often that you had reasons. What is unavailable is which of the many candidate influences actually moved the outcome — and the position of a stocking on a table is exactly the kind of influence no one's theory of themselves includes.

A second misreading is that confabulation is lying. It is not, and the distinction matters practically. A liar knows the truth. A confabulator has a gap, and the gap is filled before it can be noticed, which is precisely why confidence is no evidence at all. Confidence tracks the fluency of the constructed account, not its accuracy.

The practical upshot is uncomfortable for anything built on self-report. Asking why customers bought, why an interviewer preferred one candidate, or why a policy felt fair returns theories, not measurements. The fix is not better questioning but a different design: vary one factor across otherwise identical conditions and observe the behaviour, which is what every experiment described above did.

• Choice blindness beyond faces — taste, smell, and political attitudes, and how far the effect stretches.
• Whether training or incentives improve introspective accuracy about causes, or only about willingness to say "I don't know."
• What this implies for interviews, user research, and any method whose raw material is "why did you?"

Every term the collection defines is gathered in the glossary.