Chapter 11 — People

Sixty-nine per cent of the 375 DNA exonerations in the Innocence Project's fixed dataset involved eyewitness misidentification. Twenty-nine per cent involved a false confession. [ESTABLISHED] (Innocence Project)

Those two numbers describe the two human instruments in this book — the witness and the interview — and they are the least reliable instruments an investigator uses. This chapter grades them the way Chapter 3 graded the laboratory, because they deserve exactly the same scrutiny and almost never get it.

The interview

There are two dominant traditions and the evidence between them is not close.

The Reid technique, developed in the United States from the mid-twentieth century, is accusatory and confession-focused. It presumes guilt before the interview begins and proceeds through isolation, confrontation with evidence — sometimes fabricated — and theme development that minimises the moral seriousness of the offence to make admission easier.

The PEACE model was introduced in the United Kingdom in 1992 by the Home Office with the Association of Chief Police Officers and academic psychologists including Tom Williamson and Ray Bull. It is non-accusatory and information-gathering, and it is explicitly not designed to produce confessions. The name is a mnemonic: Preparation and planning, Engage and explain, Account, Closure, Evaluation. [ESTABLISHED] (CaseIQ)

What the research shows. Comparative studies of trained against untrained interviewers find PEACE-trained officers use significantly more open questions, more clarification, fewer leading questions and fewer coercive tactics — and that interviewees give significantly more information to trained interviewers. [ESTABLISHED] (systematic review, 2021)

Against Reid, the finding is not that it is unproven but that it is actively implicated. Saul Kassin and Richard Leo document that its guilt-presumptive structure elevates false-confession risk, disproportionately for juveniles and people with intellectual disability or mental illness — and 29% of DNA exonerations involved a false confession. Reid & Associates now markets a hybrid product incorporating PEACE, which tells you where the evidentiary pressure has landed. [ESTABLISHED] (review, 2023)

A comparative figure circulates — roughly 0% false confessions from innocent participants under PEACE-style methods against around 50% under accusatorial methods — which I found only in secondary summaries and could not trace to a primary study in this pass. I am reporting the claim and its status rather than the number as fact. [CLAIMED]

The technique with a measured effect

Most of what is written about interviewing is assertion. The cognitive interview is the exception: it has a meta-analysis with an effect size.

Across 42 studies, 55 comparisons and roughly 2,500 interviewees, the cognitive interview produced a strong increase in correctly recalled details, d = 0.87, against standard interviews — and a smaller but real increase in incorrect details, d = 0.28. [ESTABLISHED] (University of Portsmouth)

Report both halves. The technique gets you substantially more information and some of the additional information is wrong. That is a trade worth making — more material to corroborate against, rather than a cleaner but thinner account — but only if you know you are making it. An investigator who believes the cognitive interview produces more truth rather than more recall will over-trust the extra detail, which is the precise error the d = 0.28 figure warns about.

Who confesses to things they did not do

Gísli Gudjonsson built the standard instrument for this, the Gudjonsson Suggestibility Scale, which measures interrogative suggestibility in two components: Yield, going along with misleading questions, and Shift, changing answers under negative feedback and pressure.

The distinction that matters operationally comes from the review literature: compliance is linked to false confessions in general, while suggestibility specifically predicts the internalised kind — where the person comes to believe their own false admission. [ESTABLISHED] (Otgaar et al., 2021)

Two different failures wearing the same clothes. One person says what will end the interview. The other leaves the room having acquired a false memory. The second is far harder to unwind, and it is produced by the same pressure.

Of the false confessors in the Innocence Project dataset, 49% were 21 or younger at arrest and 9% had known mental health or capacity issues at trial. [ESTABLISHED]

The witness, and a defence of the witness

The reflexive modern position is that eyewitness testimony is worthless. The evidence says something more precise and more useful.

John Wixted's work established that initial confidence at the first identification reliably predicts accuracy — and that field studies of properly conducted lineups report very few high-confidence misidentifications of innocent suspects. His analysis of more than 900 eyewitness-misidentification exonerations found only 18 involved a confident initial misidentification, against 190 where the witness initially hedged, picked a filler, or said the suspect was not present. The first, least-contaminated test is the informative one. [ESTABLISHED]

Now the caveat, which is load-bearing and must not be dropped. A 2024 review confirms the confidence–accuracy relationship holds under pristine conditions: fair lineup construction, no contamination, prompt testing. Real investigations frequently fail those conditions — unfair lineups, delay, cross-racial identification, exposure to post-event information all degrade the relationship. The consensus is conditional, not universal. [CONTESTED] (Moore et al., 2024)

So the practical rule is not "distrust witnesses." It is: the first test is the evidence, and everything after it is contaminated by the fact that the first test happened. Confidence expressed in court, after months of questioning and feedback, carries almost no information. Confidence recorded verbatim at the first identification carries a great deal. Whether it was recorded is a procedural choice made in the first hour of a case, usually by someone junior.

And a related correction to a reform many jurisdictions already adopted. The move to sequential lineups through the 2000s and 2010s ran ahead of its evidence. The National Academy of Sciences found in 2014 that the studies underpinning sequential superiority confounded recognition memory with response criterion — sequential procedures make witnesses more conservative, which lowers false identifications and correct ones, rather than cleanly improving discrimination. The question remains open. [CONTESTED] (Judicature/Duke Law)

That is a reform adopted at scale on the strength of research that did not support it. Keep it in mind whenever this book, or anyone else, recommends a procedural change.

The instrument that failed and stayed

In 2003 the US National Academies reviewed roughly 80 polygraph studies, found 57 of them significantly flawed, and concluded the evidence for polygraph validity was scanty and scientifically weak. [ESTABLISHED] (National Academies, The Polygraph and Lie Detection)

It remains in use. That is the most instructive fact in this chapter, and it generalises beyond the polygraph: a technique's persistence is not evidence of its validity. Institutions keep tools because they are embedded in procedure, because abandoning them is an admission, and because the tool produces something — a number, a chart, a confession — that feels like progress. Every argument in this book about AI applies retroactively to the polygraph, and the polygraph is still there twenty- three years after the National Academies were finished with it.

What the machine can and cannot touch here

Almost nothing in this chapter is automatable, and that is the point of placing it last in the craft.

A model can transcribe an interview, translate it, index it, and flag leading questions for review — genuine value, and it makes interview quality auditable at a scale that was never previously possible. Auditability is not a small gain: PEACE's advantage over Reid is measurable precisely because someone measured question types across many interviews.

What a model cannot do is conduct the interview. Rapport is the mechanism by which PEACE outperforms, and it is not a text-processing operation. Nor can a model assess whether a witness's confidence was recorded under pristine conditions — that is a fact about a room, on a day, established by a person who was there.

The volume problem in Chapter 2 is a machine problem. This chapter is not. An organisation that automates its linking and leaves its interviewing untouched has improved the part that was already improvable and left the 69% and the 29% exactly where they were.

The Move — Record the first confidence, verbatim, and never ask again. At any identification or key recollection, capture the witness's own words about how certain they are, at the first asking, before any feedback, and timestamp it. Then do not re-ask in a way that invites a revised answer. This single record is the only version of witness confidence the research supports relying on, and the opportunity to create it exists exactly once, early, and is usually missed.

What would change my mind: if replication showed the confidence–accuracy relationship holding under ordinary, non-pristine investigative conditions, the emphasis on the first test would be excessive caution. The 2024 review points the other way, and the sequential-lineup episode is a standing reminder of what happens when a procedure is adopted before that question is settled.