The suspect in position three

In 1984 Gary L. Wells put 192 eyewitnesses in front of a lineup that demonstrably did not contain the culprit — a blank lineup. Only afterwards were they shown the actual lineup. Those who had picked nobody out of the blank lineup went on to choose the wrong person markedly less often than the remaining witnesses did. Correct identifications did not drop appreciably because of the intermediate step. The detour had therefore separated two kinds of witness who until then had looked the same.

From this Wells drew the conclusion that has organised the field ever since: someone standing in front of a lineup does not compare face with memory but face with face. The person chosen is whoever most resembles the inner image — relative to the other five. The strategy works as long as the culprit is present. It fails silently the moment he is absent, because the most similar of six is always somebody. Whether he is present is something the witness never knows. An instruction alone does not fully remedy this, as Wells already noted in 1984.

The question nobody asks

Six faces stand side by side. One belongs to the suspect, five are fillers. The witness knows the police have found someone. Without ever being spoken, the arrangement contains a question: which of them is it? The scientifically correct question is a different one: is the culprit present at all?

Whether a lineup presented sequentially cures the relative choice is an open question. The National Academies report noted in 2014 that the superiority of either procedure remains unresolved, and urged caution about switching procedures. The rest is not in dispute: fillers must match the original description, and the suspect must not be the only one carrying a conspicuous feature. A lineup with five clean-shaven men and the bearded suspect is not a memory test but a cue.

What the procedure can control and what it cannot

The conceptual framework for this comes from Wells in 1978. He separated two classes of influence. Estimator variables are the circumstances of the offence: lighting, distance, duration, stress, the question of whether witness and culprit belong to the same population group. After the fact they can at best be estimated, not altered. System variables are everything the criminal justice system itself controls: the composition of the lineup, the instruction, the behaviour of the officer, the timing of the interview.

The distinction sounds technical and has far-reaching consequences. It shifts the question from “how good is this witness’s memory?” to “how good was the procedure by which we questioned it?”. Everything research can recommend concerns the second question.

Nine recommendations, four of them from 1998

In 2020 six experts working for the American Psychology-Law Society produced a revision of the influential 1998 recommendations; feedback from researchers and the courts was gathered over eighteen months. Nine recommendations resulted. Four came unchanged from the 1998 paper: the selection of fillers, the instruction given before the lineup, double-blind administration, and the collection of a confidence statement.

Double-blind here means what it means in an experiment: the officer conducting it should not know who the suspect is. Ignorance plus standardised instructions prevents the small signals that in any study go by the name of demand characteristics — a longer look, a repetition, audible relief after a choice. And the witness must hear explicitly that the person sought may be absent and that the investigation will continue if she recognises nobody.

Five recommendations were added in 2020. Police should interview the witness before the lineup rather than only afterwards. There should be evidence-based suspicion before anyone is placed in a lineup at all. The entire procedure should be video recorded. Repeated identification attempts with the same witness and the same suspect are to be avoided. And showups — one person, presented without comparison — should be avoided where possible and, where unavoidable, conducted better.

What confidence is worth

This is the point at which the field revised its own received wisdom. For decades courts and textbooks alike held that a witness’s stated confidence said next to nothing about whether her identification was correct. Wixted and Wells showed in 2017 that this pessimism was reasonable but tied to the very procedures that had produced it.

Under what they call “pristine” conditions — a first, uncontaminated test, a fair lineup, no influence from the administrator, a confidence statement taken immediately — confidence and accuracy are closely related in adults. Identifications made with high confidence are then remarkably often correct; identifications made hesitantly the first time round are error-prone. Both hold only under those conditions. If the lineup is unfairly composed, even the high-confidence identification can no longer be relied on.

What follows is not that witnesses should be believed more, but that they must be asked earlier. Feedback such as “good, you picked our suspect” raises confidence in one’s own memory without altering its original accuracy. The Oregon Supreme Court put the heart of it into one sentence in 2012, which the research has cited ever since:

How far the gap between appearance and substance can open is shown by the analysis of 161 convictions overturned by DNA. Every one of those witnesses had testified with confidence at trial.

When a procedure manufactures its own evidence

Repeated identifications are problematic for the same reason. A witness who first sees a photograph and later picks the same person out of a lineup may be remembering the first photograph rather than the offence. Reconstructive memory works with the material available to it, and in doing so does not reliably distinguish between images that were experienced and images seen afterwards. The second identification then does not confirm the first; it repeats it.

That is exactly the point at which a lineup stops being a measurement. It is a psychological experiment with a single trial. The experimenter holds a hypothesis, the participant does not. The dependent variable is her choice. Design the procedure badly and you do not merely measure badly — you alter the object you set out to measure, and then produce the confirmation for it.

What remains

Research does not say that eyewitnesses are unreliable. It says something more uncomfortable: the reliability of an identification is not a property of the witness but in large part a product of the conditions under which it was collected — and those lie with the authorities. A fair arrangement, a blind administrator, an early interview, a recorded procedure, no second lineup with the same person, and a confidence statement documented immediately: these preserve what the witness actually remembers. Without them the damage cannot be undone, because the memory itself has been overwritten.

In everyday life the pattern shows up in one detail: when somebody’s confidence rises only after a confirmation, the confirmation is the source and not the memory. That holds inside the interview room and outside it. Anyone who wants to know what a person really remembers has to ask before the first feedback — after that, they are measuring their own feedback along with it.