Field Notes

concept

Investigative confirmation bias

StatusWorkingUpdatedStands on3 sourcesSupportsnotes

Confirmation bias is a tendency to selectively search for and emphasize information consistent with a preferred hypothesis, while opposing information is ignored or downgraded.1 In a criminal case the preferred hypothesis is normally a particular person’s guilt, and the actors who adopt it early also decide what gets collected next: which questions are asked, which devices are searched, which alternative suspects are pursued, and what the file eventually contains.

The mechanism is worth naming separately from the mistakes it produces because it is not one actor’s lapse of care. It operates at successive decision points, each narrowing what the next can see, so a case file can end up internally coherent because every stage built on the first hypothesis rather than tested against it. A later reader sees a record that agrees with itself and cannot tell agreement from corroboration.

It is not simply carelessness

Treating the bias as unprofessional sloppiness both misdescribes the evidence and points at the wrong remedy. Lidén’s experimental work concluded that the confirmatory reasoning her studies found can be considered rational or irrational depending on which rationality is applied, probabilistic or judicial.2 An investigator who treats an arrest decision as raising the probability that this suspect is the offender is doing something defensible taken alone. The damage comes from that updated belief silently governing which questions get asked next, because the answers then arrive already filtered and are read as independent support.

This is why exhortations to remain objective work poorly as a control. The actors in the documented cases were not indifferent to truth; they were following a hypothesis that had become the frame through which further evidence was legible at all.

The driver changes by stage, so the remedy does too

The finding that matters most for reform is that confirmation bias is not one uniform effect applied evenly across a criminal procedure. Lidén ran separate experiments on Swedish police officers, prosecutors, and judges, and found the bias appearing at different points, for different reasons, with different available countermeasures.3

Stage What was measured Apparent driver Implied control
Police questioning More guilt-presumptive questions to apprehended than to non-apprehended suspects Cognitive load Reduce the load under which questioning is planned
Prosecutorial review No bias before the charging decision; afterwards, less interest in further investigation and more guilt-confirming proposals Not established Unclear; the trigger is the decision itself, not new evidence
Pre-trial detention Judges who had ordered detention later rated the prosecution evidence as stronger and were likelier to convict Social and commitment factors Separate the judge who detains from the judge who tries

The prosecutorial result is the most awkward of the three, and it should not be smoothed over: the bias appeared only after charging, which means the decision itself, rather than any new evidence, changed how the same material was assessed. Lidén reports the driving forces there as needing further examination.

The practical consequence is that training people to notice their own bias is the weakest available response, since only one of the three stages has a driver that awareness plausibly addresses. The detention finding in particular supports a structural fix rather than an attitudinal one, and it is the direct warrant for separating detention and trial decision-makers, a reform Rättssäkerhet in Swedish criminal cases recommends.

Technical stages are not exempt

Forensic work is often assumed to sit outside this problem because its outputs are machine-produced. The assumption does not survive contact with how the work is tasked. Cognitive and human factors in digital forensics maps sources of error running through the digital-forensics process that are contextual and organizational rather than technical, arising from what the examiner is told, what they expect, and how their role is positioned.

Mobile-device extraction and evidentiary selection gives the bias unusual leverage: an extraction can yield an enormous archive, and an examiner who knows the allegation before choosing search terms, date ranges, and applications decides which fraction of it becomes evidence. Every exported artifact can be authentic while the selection that produced them was governed by the hypothesis they are offered to support.

Why the record hardens

Confirmation bias would matter less if later review reliably caught it. The documented pattern is that review inherits the frame rather than re-examining it. Felaktigt dömda found the same chain across eleven reopened convictions: an early hypothesis, questioning built on it, missing alternative suspects, and an incomplete file that reached prosecutors and courts looking coherent.

Three features make the bias unusually durable here. The original conditions are unrecoverable, so a contaminated first account cannot be re-collected. Review runs through the institutions whose earlier work is in question. And the reopening threshold is applied to a record those same institutions assembled. The last two are the general defect in a Correction channel, arriving here with an investigative frame already in the file: the channel that would detect the bias is fed by the process the bias shaped. Lidén’s archive study estimated that far from all wrongfully convicted people who appeal or petition for a new trial are acquitted, a conclusion that survived a wide range of assumptions about unknown parameters.4

Relationship to neighbouring notes

Coercive interrogation and false confessions develops the interrogation stage in detail, where presumed guilt combines with legal deception, information contamination, and pressure to produce a specific output. This note covers the mechanism across the wider chain, including stages that produce no confession at all.

Automation bias is the same shape with a machine supplying the frame. Where confirmation bias starts from a hypothesis a person adopted, automation bias starts from an output a system produced, and both end with a review step that ratifies what it was supposed to test. The two literatures converge on the remedy as well: instructing the reviewer to be careful is the weakest control either one identifies.

Witness reports are not ground truth addresses a distinct failure that this one amplifies. Witness error arises inside a person; investigative confirmation bias determines which witnesses are asked, how, and how often, and can convert several accounts into one contaminated chain.

  1. Lidén 2018, abstract: “a tendency to selectively search for and emphasize information that is consistent with a preferred hypothesis, whereas opposing information is ignored or downgraded.” ↩

  2. The thesis states that the confirmatory reasoning in Studies I–III “can be considered rational or irrational, following different types of rationality, like probabilistic or judicial rationality.” ↩

  3. Studies I–III, experimental work with Swedish police officers, prosecutors, and judges, reported in Lidén 2018. ↩

  4. Study IV, an archive study of appellate and Supreme Court appeals and petitions for new trials. ↩

Built on 3 sources (2 archived here, 1 external).

Working out connections…