Latent Fingerprint Examination: The Discipline That Argues With Itself

Ask a fingerprint examiner how many matching features it takes to declare an identification, and the honest answer is that there is no number. The United States abandoned any minimum point standard in 1973, and the international Ne'urim declaration of 1995 confirmed it: no scientific basis exists for requiring a set count of corresponding ridge features to establish an identification. What replaced the number was the examiner's training and experience. As Simon Cole traces it, the discipline's own founders knew the criteria were, in Stoney's words, subjective and ill-defined, the products of probabilistic intuitions widely shared among examiners, not of scientific research.

That missing threshold is not a historical footnote. When Bradford Ulery and colleagues at the FBI and Noblis actually measured where examiners draw the line, they found no shared line at all. More than a third of examiners would individualize on eight or fewer corresponding minutiae; others held out for as many as fourteen. Handing examiners a twelve-point standard did not even change how often they individualized, sixty-nine per cent for the twelve-point group against sixty-two for the rest. What the jury hears as a fact of nature is, in the authors' words, the examiner's determination, not an objective metric.

None of this means fingerprints are worthless, and an honest witness should not pretend it does. A clear latent with ample ridge detail, compared to a good exemplar, is powerful evidence, and examiners are genuinely skilled at it. The point is narrower and harder to wave away. The identification rests on a premise that was assumed rather than proved, and it is delivered without the one thing that would make it objective: a stated, validated amount of agreement that separates a match from a non-match. When an examiner testifies that two prints match, the jury should know there is no measured line the examiner crossed to say so. There is a judgement, made by a person, that this was enough.

In court the discipline's method has a name: ACE-V, for Analysis, Comparison, Evaluation and Verification. It sounds like a protocol, and it is offered as the reason the work is scientific. Look closely and it is a description of the order in which an examiner does things, not a procedure with rules, thresholds, or a measured error rate. The statistician Sandy Zabell called it an acronym, not a methodology. Joseph Kadane, reviewing it, wrote that it is more an outline than a scientific method: two analysts can perform different activities, and arrive at different conclusions, while both can claim to be using the ACE-V methodology.

Ralph and Lyn Haber put the challenge at its sharpest. To validate ACE-V you would run skilled examiners through many latent-and-exemplar pairs of known ground truth, documenting each step, and measure how often the method's conclusions match the truth. That experiment, they note, has never been offered as evidence in any of the Daubert hearings, nor has one ever been published. Resemblance to the scientific method is not evidence of accuracy. It is, in their words, an analogy and nothing more.

There is a constructive alternative, and it matters because it shows the categorical claim is a choice rather than a necessity. Christophe Champod, Cedric Neumann and colleagues have built likelihood-ratio models that ask a different question: not is this the source, yes or no, but how much more probable is this degree of agreement if the prints share a source than if they do not. On that scale, exclusion and individualization are simply the two endpoints, a likelihood ratio of zero and of infinity, with everything real sitting in between as a weight of evidence. The honest reading of most casework is a strength, not a certainty. ACE-V, as practised, collapses that continuum into three boxes and hands the jury the most extreme one.

For most of its history the field claimed an error rate of zero. In 2011 it finally measured itself. Bradford Ulery and colleagues ran 169 experienced examiners through thousands of comparisons of known ground truth, the study the discipline had never done, funded by the FBI's own laboratory. The results are now the single most important fingerprint numbers in any courtroom. Six false positives occurred, an overall false-positive rate of about one in a thousand. But false negatives, missed identifications, ran at seven and a half per cent, and eighty-five per cent of examiners made at least one. The reassuring headline and the uncomfortable one live in the same study: false positives are rare, and yet error is entirely ordinary.

A year later the same group asked whether examiners even agree with themselves. They gave 72 of the original examiners the same prints again, months later, without saying they had seen them before. Examiners repeated only about ninety per cent of their own individualizations and exclusions. Roughly one decision in ten changed. If the same expert, on the same prints, reaches a different answer on a different day, then the confident, unqualified match delivered from the stand describes the examiner's state on one occasion, not a fixed property of the prints.

Koehler closes the last escape hatch. Examiners sometimes concede a human error rate but insist the method itself has an error rate of zero, that mistakes are the analyst's, not ACE-V's. The distinction does not survive contact with how the work is done. The method has no existence apart from the person applying judgement at every step, choosing features, deciding sufficiency, calling the conclusion. As Koehler puts it, the method literally is the people who employ it. There is no clean machine underneath the examiner to which a zero could honestly attach.

The most consequential argument in fingerprinting right now is not about ridges. It is about arithmetic, specifically what to do with inconclusive decisions. The famous one-in-a-thousand false-positive rate depends on leaving inconclusives out of the sum. When the President's Council of Advisors on Science and Technology reworked the Miami-Dade study, counting differently, it reported that false positives could occur as often as one in eighteen. Same discipline, same kind of data, a rate that moves by nearly two orders of magnitude depending on a bookkeeping choice most juries never hear discussed.

Why the choice matters so much is easy to see. If inconclusive decisions never count against an examiner, then, as Itiel Dror and Nicholas Scurich point out, an examiner could reach an inconclusive on every comparison and post a perfect score. So the denominator is doing enormous work. Reworkings of the Miami-Dade figures by Jonathan Koehler and others land anywhere from under one per cent to over four, and Madeline Ausdemore's careful re-analysis argues the highest figures overstate it. The point for a witness is not which number wins. It is that the error rate is not a fact read off the data. It is a number that depends on decisions the examiner's own field has not settled.

One more wrinkle belongs here, because most modern latents arrive through a database. When a print is searched against an automated system, the computer returns a ranked list of candidates, and Itiel Dror and Kasey Wertheim showed that false identifications cluster at the top of that list, occurring even when the true match sits lower down. The ranking is a suggestion, and the suggestion biases the human. Ralph Haber, cataloguing the ways automated search introduces error, put it memorably: the technology tail is wagging the forensic dog. The candidate the machine ranked first is not the same thing as the person who left the mark.

Set outright error aside and a quieter problem remains: examiners disagree, with each other and with themselves. The most revealing finding concerns what examiners do after they see the suspect's print. Ulery and colleagues had examiners mark the features of a latent during analysis, then compared that markup to what they relied on during comparison. On more than ninety per cent of individualizations, examiners added or deleted minutiae once the exemplar was in front of them, and every examiner revised at least some markups. Even when an identification rested on eight or fewer corresponding features, in most cases some of those features had not been marked until after the suspect's print appeared.

That is the textbook shape of circular reasoning: the answer you are testing against reshapes the evidence you test it with. It is not a hypothetical. In the Madrid misidentification, the initial examiner reinterpreted five of the original seven analysis points once Mayfield's exemplar was in view. And it changes conclusions wholesale. When examiners were required to mark features against the exemplar, prints they would otherwise have called not good enough to identify became identifications far more often, twenty-six per cent of the time against under two per cent when no markup was required. The exemplar does not just confirm the latent. It can manufacture the correspondence.

This is why the field-average error rate, whichever one you accept from the last two sections, does not tell you about the examiner in the witness box. The studies measure a population; the conclusion in this case came from one person, with one private threshold, on one day, using features they may have marked only after they saw whom they were looking for. Philip Kellman's work shows the same latent can be misjudged by nine examiners in ten when it is genuinely difficult, so error is a property of the print pair as much as the person. A defensible identification survives all of this. But the jury deserves to know that same print, two examiners can mean two different answers, and that same print, same examiner can mean two different answers on two different days.

In 2006 Itiel Dror ran an experiment that should be on the syllabus of every forensic discipline. He took five experienced examiners and gave each a pair of prints they themselves had identified as a match in real casework years earlier. This time he told them the prints were the ones the FBI had wrongly matched to Brandon Mayfield. Four of the five reversed their own earlier conclusion, now calling the prints a non-match or insufficient. The ridges had not changed. Only the context had. A follow-up with subtler cues, a suspect confessed, a suspect was in custody, flipped two-thirds of examiners on at least one decision. The prints were fixed; the answers moved with the story attached.

The mechanism is not stupidity or corruption. It is how expert perception works, and it enters everywhere. A suggested target print changes which features an examiner marks in the latent. Being told a suspect confessed, or that a DNA test already implicated him, shifts the fingerprint call, as Stevenage and colleagues showed for a DNA result specifically. Bias even reaches the earliest, most innocent-looking decision, whether a mark is suitable to compare at all: Fraser-Mackenzie and colleagues found a non-matching exemplar made examiners markedly more willing to call a borderline latent suitable. The context does not wait until the end to tip the scale. It is present from the first glance.

Because the effect is real, the field has built a defence against it, which means a witness can be asked whether they used it. Linear Sequential Unmasking, developed by Dror, William Thompson and colleagues and later expanded for casework by Quigley-McBride and colleagues, sets out the discipline. Analyse and document the crime-scene mark in isolation first, before you are shown the suspect's exemplar, and limit what you are allowed to revise afterward, so the reference cannot quietly reshape the latent. The logic runs from evidence to suspect, never the other way. If an examiner worked the latent side by side with the suspect's print, was told the back-story before they finished, or verified a colleague's call knowing the answer, they skipped the safeguard the field itself designed. The question is not whether they are honest. It is whether the procedure protected the honesty.

Everything in the previous sections converges on a single word. When an examiner testifies to an identification, or an individualization, or in the newer official language a source identification, the claim is that this mark was left by this person and by no one else on Earth. Christophe Champod's objection is arithmetic, not rhetorical. To exclude everyone alive, the evidence would need a likelihood ratio somewhere above ten to the thirteenth power, and, he writes, that is out of reach of current systematic research. The best validated fingerprint statistics reach something like one in a billion. Between one in a billion and everyone on the planet lies a claim the science cannot cover, and the word quietly makes it anyway.

The field knows this, and its reforms have mostly renamed the claim rather than retired it. After the National Academy of Sciences said in 2009 that no discipline but DNA could support individualization, the profession dropped the phrase to the exclusion of all others, relabelled individualization as identification, and added softer words like practical impossibility. Simon Cole's verdict is that these changes are logically empty. As the critique of the Department of Justice's approved language puts it, there is no logical difference between saying the defendant is the source of the print and saying it to the exclusion of all others; the official conclusion still rounds an admitted, non-zero probability down to a statement of certainty for the jury to hear.

The abstractions have names. In 2004 a bomb on a Madrid commuter train left a latent print on a bag of detonators, and the FBI matched it, through its database, to Brandon Mayfield, an Oregon lawyer and Muslim convert who had not left the country. The examiner called it a hundred per cent positive identification, and two more experienced examiners verified it. The Spanish police disputed the match within weeks; the Bureau held to its certainty until Spain identified the true source, an Algerian named Ouhnane Daoud. The Inspector General's review found exactly the failure modes in this reading: the power of the database hit created a mind-set that the prints matched, the examiner reasoned backward from Mayfield's exemplar in what it called circular reasoning, charting features into the latent that were not there, and the verifications were not blind, each examiner already knowing an identification had been made.

Mayfield is not alone. In Scotland, Shirley McKie, a detective constable, was identified by four examiners of the Scottish Criminal Record Office as having left her print inside a murder scene she said she never entered, and was prosecuted for perjury when she denied it; she was acquitted after two American examiners testified she was excluded, and the 790-page public inquiry that followed concluded the identification was wrong and that fingerprint evidence should be recognised as opinion evidence, not fact. In Massachusetts, Stephan Cowans was convicted partly on a fingerprint and served six years before DNA cleared him, the first such exoneration in which a fingerprint identification helped convict an innocent man. The database had not flagged him even though his prints were on file; the print only matched after other evidence made him a suspect and an examiner compared it to his known print.