Rosbach 2026 — Cognitive Biases in AI-Assisted Medical Decision Making

A structured review that went looking for studies measuring cognitive bias during expert–AI diagnostic collaboration, found 8 primary studies across all of medicine and exactly one in pathology, and wrote the pathology examples itself because the literature had none to offer.
Author

Rosbach E, Ammeling J, Bertram CA, Riener A, Aubreville M. Technische Hochschule Ingolstadt, Ingolstadt, Germany (Rosbach, Ammeling, Riener); University of Veterinary Medicine Vienna, Vienna, Austria (Bertram); Flensburg University of Applied Sciences, Flensburg, Germany (Aubreville). First author Emely Rosbach, ORCID 0009-0001-7526-9923

Doi

10.1177/03009858261472493

A structured review that went looking for studies measuring cognitive bias during expert–AI diagnostic collaboration, found 8 primary studies across all of medicine and exactly one in pathology, and wrote the pathology examples itself because the literature had none to offer.

Read status: abstract only. The paper is paywalled and has no open deposit anywhere. The substance of the review — twelve biases, each with a definition and a constructed pathology-specific example — is unread, and no claim below should be taken as reporting it. What is solid: the abstract’s method and counts, and the full reference list from Crossref. The full text is still wanted. Because sources/papers/ is append-only, when it arrives the reading goes on a wiki page rather than back into this note.

What it is, from the abstract

A structured literature review, search dated 19 August 2025, across ACM, IEEE and PubMed. Eligibility: studies that operationalised cognitive biases during expert–machine interaction in diagnostic decision-making, qualitative or quantitative, plus review articles citing primary studies that met the same bar.

The corpus arithmetic is the finding, and it is worth stating precisely because it is the number that justifies the paper existing:

Count
Final corpus 24 studies
— of which primary studies 8
— of which review articles 16
Additional primary studies extracted from those reviews, and analysed in their place 18
Cognitive biases identified in AI-assisted medical decision-making 12
Of the primary studies, number originating from pathology 1

So: after a three-database search of the entire medical literature, the number of primary studies that have actually measured a cognitive bias while a pathologist works with an AI is one. For each of the 12 biases the authors “derived” definitions and hypothetical pathology-specific examples — hypothetical because there was no empirical pathology work to describe. The paper calls itself a primer for the veterinary and human pathology communities, which is an accurate description of a document that had to invent its own worked examples.

What the reference list gives that the abstract does not

The Crossref deposit carries all 107 references. Two of them settle a question that the abstract leaves open, and they are the reason this note is filed alongside the other two in this batch:

The review is therefore by the same group as its own pathology evidence base, and both of this batch’s experimental papers are inside its corpus or its citation network. That is not a criticism — there was nothing else to cite — but it does mean the review’s central claim (“only one study originating from pathology”) and the authors’ own experiment are the same fact seen twice.

Which of the two is “the one pathology study” is not determinable from outside the paywall. The search covered ACM, IEEE and PubMed. The CHI’25 paper is in the ACM Digital Library and would be found; the BVM 2025 paper is a Springer Informatik aktuell book chapter and is indexed in none of the three. The likeliest reading is therefore that the single pathology primary study is the CHI’25 confirmation-bias paper, with the BVM automation-bias paper cited separately in the discussion. [unverified] — this is inference from the search strategy and the indexing, not something the paper says.

Three of the twelve biases can be named from reference titles alone — confirmation bias (ref 13, Bashkirova & Krpan 2024, the psychologist-triage study that supplies the “congruent advice increases acceptance” design), automation bias (ref 97, Schemmer et al., ECIS 2022, On the influence of explainable AI on automation bias) and anchoring bias (a CHI’23 proceedings entry, almost certainly Bach et al. on ophthalmologists and anchoring-bias mitigation, which is also ref 8 of the CHI’25 paper). The remaining nine are not recoverable from the metadata — much of the SAGE deposit is DOI-only with no titles — and are not guessed at here.

The reference list also confirms the review’s veterinary framing is real rather than decorative: it cites veterinary-specific diagnostic-error work (Orakpoghenor 2024, Problems in veterinary pathology: a focus on diagnosis) alongside the mainstream HCI corpus (Bučinča’s cognitive forcing functions, the MIDOG challenge, Aeffner’s gold-standard-paradox paper).

Why it earns the space here

Not for its findings, which are unread, but for the denominator. This wiki has repeatedly recorded that pathology AI is evaluated for accuracy and almost never for what it does to the reader. This review puts a number on that gap from the other direction: the entire indexed literature contains one primary study of a pathologist’s cognitive bias under AI assistance. Any claim that a decision-support tool “improves interobserver reproducibility” is at present resting on an evidence base with essentially no measurement of the bias it may be introducing at the same time.

The second reason is the audience. Veterinary and human pathology are treated as one community here, which matches how this group’s own material behaves — the MIDOG mitosis work, Bertram’s computer-assisted mitotic count study, and the participant panel of the companion CHI’25 experiment all span both.

Reading it critically — with the caveat that the body is unread

  1. A structured review is not a systematic review, and the paper says “structured” rather than “systematic” in its own title. No PRISMA claim is visible in the abstract, no registration is mentioned, and no risk-of-bias assessment is described. With 8 primary studies that is a defensible choice, but the label matters for how much weight the corpus count carries.
  2. The 18 extracted studies are used “in their place” — that is, primary studies pulled out of the 16 reviews and analysed instead of the reviews themselves. That avoids double-counting review-level conclusions, which is the right instinct, but it means the effective evidence base is 8 + 18 = 26 primary studies that were not all retrieved by the search strategy itself. Whether those 18 were screened against the same eligibility criteria is not stated in the abstract. [unverified]
  3. Three databases, and the choice shapes the result. ACM and IEEE cover HCI; PubMed covers medicine. Springer proceedings (BVM, MICCAI), Scopus and the psychology literature are not searched. The authors’ own BVM paper being outside its own review’s reach is a small demonstration of the gap.
  4. Hypothetical examples are the product, and they are unvalidated by construction. A derived pathology example of, say, availability bias is a plausible story, not a measurement. The review is explicit that it is a primer; a reader should not cite it for the existence of a bias in pathology.
  5. The search date is 19 August 2025 and publication is 5 August 2026 — a year of lag, in a field where the relevant HCI venues publish annually. Anything from CHI’26 is out of scope.

How it connects

Open questions

  • What are the other nine biases? Not recoverable without the full text, and they are the main reason to want it.
  • Which study is the single pathology primary study? [unverified], see above.
  • Does the review propose any mitigation, or only classify? The abstract mentions “safe and effective integration” as an aim but names no intervention. The reference list includes Bučinča’s cognitive forcing functions and Schemmer’s xAI-and-automation-bias paper, which suggests mitigation is discussed, but that is inference from citations. [unverified]
  • Is there a table mapping each bias to a pathology task? If so it is the single most reusable object in the paper for this group.