Rosbach 2025 — Confirmation Bias and Time Pressure in AI-Assisted Tumour Cell Percentage Estimation

Twenty-eight pathologists estimated tumour cell percentage twice, two weeks apart, the second time with an AI: when the AI’s number sat close to their own wrong first answer they moved further towards it than their own prior judgement warranted, and time pressure made that effect smaller rather than larger.

Author

Rosbach E, Ammeling J, Krügel S, Kießig A, Fritz A, Ganz J, Puget C, Donovan T, Klang A, Köller MC, Bolfa P, Tecilla M, Denk D, Kiupel M, Paraschou G, Kok MK, Haake AFH, de Krijger RR, Kasantikul T, Sonnen AF-P, Dorrestein GM, Smedley RC, Stathonikos N, Uhl M, Bertram CA, Riener A, Aubreville M. Twenty-seven authors across Technische Hochschule Ingolstadt, Katholische Universität Eichstätt, Freie Universität Berlin, Animal Medical Center New York, University of Veterinary Medicine Vienna, Medical University of Vienna, LMU Munich, Ross University School of Veterinary Medicine, University of Milan, Michigan State University, Universiti Putra Malaysia, UMC Utrecht / Princess Máxima Center, NOIVBD Vessem, and Flensburg University of Applied Sciences

Doi

10.1145/3706598.3713319

Read status: full text, from the arXiv v1 preprint rather than the CHI ’25 version of record. The preprint is 16 pages; the ACM record gives pages 1–18. Treat every number below as the preprint’s. Read alongside Rosbach 2025 — Automation Bias in AI-Assisted Tumour Cell Percentage Estimation under Time Pressure — that paper reports the same experiment, same 28 participants, same 20 patches, same two-week washout, same ten-second timer. They are two analyses of one dataset, not two studies.

Why this one matters here

It is, as far as Rosbach 2026 — Cognitive Biases in AI-Assisted Medical Decision Making could establish, the only empirical measurement of a cognitive bias in a pathologist working with an AI. Everything else this wiki records about decision support — that computer-assisted mitotic counting improves reproducibility, that a triage model is 80% precise — is a statement about accuracy. This is a statement about what the tool does to the reader, and it is the first one with a coefficient attached.

The second reason is the task. Tumour cell percentage (TCP) — the proportion of neoplastic cells among all cells — is not an exotic research endpoint. It gates molecular testing: each assay has a tumour-DNA threshold, and a sample estimated above a threshold it is really below produces a false negative that is then trusted. The paper picks it precisely because it is quick, routine, quantitative, unstandardised and famously irreproducible — which makes it the ideal substrate for a bias, and also makes the finding directly relevant to any department sending blocks for sequencing.

The design

A 2×2 within-subject online experiment: AI (yes/no) × time pressure (yes/no).

Participants: recruited by invitation from an existing professional network, unpaid, each independently verified by the authors. 31 completed round one, 28 completed both — 25 pathologists, 2 residents, 1 non-physician pathology staff member. 64% male, 32% female; 15 of 28 (53.5%) with more than 15 years of experience. Both veterinary and human pathologists (the author list itself spans both, across nine countries).
Task: estimate TCP on 20 H&E patches (plus 3 for training), zoomable in a web interface, drawn from BreCaHad, Frei 2023 and BreastPathQ. Deliberately not size-normalised or stain-normalised — the authors wanted variable difficulty, including necrosis, immune infiltrate, staining artefacts and low-cellularity slides.
Two rounds separated by a two-week washout. Round one: unaided, establishing each participant’s own baseline per patch. Round two: identical images, now with an AI prediction.
The AI is real, not simulated — an FCOS object detector trained on BreCaHad, with five BreCaHad test patches held into the experiment. About half the study slides got accurate detections; the rest showed high false-positive and false-negative rates from covariate shift across the three source datasets. The authors argue, correctly, that a real model’s error pattern is what makes participant reactions meaningful.
Explanation: prototype-based xAI — the highest-confidence tumour and non-tumour detections shown as exemplars — plus a toggleable overlay of colour-coded cell detections.
Time pressure: a ten-second countdown per slide, sized from a keystroke-level-model estimate (7 s) adjusted upward after pilot feedback. At three-quarters elapsed a lock overlay blurs the image and the AI panel; submission after expiry is still permitted so every slide gets rated.
Order controlled by random assignment to one of four groups plus a balanced 10×10 Latin square within each image set; confidence rated 1–5 after every estimate.

What was measured, and how confirmation bias was operationalised

The definition is borrowed from Rosenbacke — false confirmation: an initially wrong human judgement reinforced by wrong machine advice. That is a sharper target than “agreement with AI” and it is what makes the analysis worth reading.

Two dependent quantities, both in percentage points: - Pred_AI − Est_B — how far the AI’s number sits from the participant’s own baseline estimate. - Est_AI − Pred_AI — how far the final, AI-aided answer ends up from the AI’s number.

If people simply ignore the AI, the second grows with the first at slope ≈ 1. If they are pulled towards it more strongly when it agrees with them, the slope falls below 1 and the intercept sits near zero. Linear mixed-effects models (lme4 1.1.35.5, R 4.4.0, lmerTest 3.1.3, Satterthwaite), participant as random intercept.

Step 1 restricted to cases where the AI-aided answer was wrong by more than 5 percentage points (the upper two tertiles of |Est_AI − GT|) and where baseline, final and AI estimates all erred in the same direction.

Results

Step 1 — Table 2 of the appendix. DV is Est_AI − Pred_AI:

Model	Condition	n	Slope on `Pred_AI − Est_B`	SE	p	Intercept	p
1	all	216	0.61	0.03	<.001	−0.72	0.53
1.1	no time pressure	105	0.65	0.04	<.001	−1.51	0.21
1.2	time pressure	111	0.56	0.05	<.001	0.32	0.85

The intercept is statistically indistinguishable from zero in all three: when the AI’s number equals your own, your final answer is the AI’s number.

Step 2 — Table 5, and the more interesting half. DV is the final estimate itself, with the baseline estimate and the AI prediction as competing predictors — a simulated weighted average. The subgroups split on the median distance between AI and baseline, 15 percentage points:

	Congruent AI (Model 2, n=78)			Incongruent AI (Model 3, n=139)
Predictor	coeff	SE	p	coeff	SE	p
Own baseline estimate	0.25	0.15	0.09	0.47	0.05	<.001
AI prediction	0.60	0.17	<.001	0.43	0.06	<.001

Read the p-values, not only the coefficients. When the AI agreed with them, the pathologists’ own prior estimate no longer significantly predicted their final answer (p = 0.09) — only the AI did. When the AI disagreed, their own estimate was the stronger of the two and both were significant. The paper reports the coefficient comparison; it does not draw out that the congruent-condition baseline coefficient fails significance, which is the sharpest single sentence available from its own appendix.

Confidence and reliance (descriptive): with congruent AI, mean confidence 3.87 and mean JAS 0.55; with incongruent AI, 3.24 and 0.49. JAS is the judge-advisor statistic, 0 = ignore the advisor, 1 = adopt it wholly, 0.5 = equal weight. So congruent advice made them both more confident and more compliant.

Time pressure — and the result that contradicts the hypothesis. Confirmation bias got weaker under the timer (0.65 → 0.56). But overall reliance on the AI rose: JAS 0.49 → 0.55, paired one-tailed t-test t(27) = −2.80, p = 0.005, and mean confidence fell slightly (3.65 → 3.63). The authors’ reading, which is reasonable and which they flag as interpretation rather than measurement, is that automation bias eclipses confirmation bias under time stress: with no time to reflect, people stop weighing whether the advice agrees with them and simply take it. That handoff between two biases is the finding this batch is worth reading for, and it is completed by Rosbach 2025 — Automation Bias in AI-Assisted Tumour Cell Percentage Estimation under Time Pressure, which measures the automation-bias half on the same data.

Hypotheses H1a, H1b and H1c accepted; H2 rejected.

Reading it critically

The step-2 subgroups are not built by the same rule, and the paper’s central comparison rests on comparing them. The congruent group (n=78) required the AI prediction to deviate from ground truth by >5% and all three deviations to share a sign. For the incongruent group (n=139) the Methods state plainly that “no additional filtering was applied based on the AI prediction or the direction of deviation.” So Model 2 describes false-confirmation cases specifically while Model 3 describes all distant-advice cases. The asymmetry is disclosed, not hidden — but its consequence, that the two coefficient sets are not like-for-like, is not discussed. This is the paper’s weakest structural point.
Appendix Table 6 is not mentioned in the main text and it complicates the headline. Stratifying by how wrong the AI-aided estimate turned out to be, the same positive slope appears in every stratum: correct (|Est_AI − GT| ≤ 5, N=203) 0.52; minor error (5–12, N=172) 0.64; severe error (12–79, N=185) 0.49, all p < .001. The relationship is therefore not specific to erroneous outcomes — it is present, at similar strength, when the final answer was right. That is consistent with ordinary sensible weighting of a second opinion, and it weakens the inference that what is being measured is specifically a bias rather than advice-taking. The main analysis’s restriction to errors is what makes it a false-confirmation study; Table 6 shows what the unrestricted picture looks like, and the paper does not reconcile the two.
n = 28, one professional network, volunteers. The authors say so, and note it is typical for this HCI literature. It still means the 0.61 has a wide implicit uncertainty, and the step-2 models rest on 78 and 139 observations from those same 28 people.
A ten-second countdown is not clinical time pressure, and the paper says so. Real pressure in a laboratory is specimen volume across a session, not a per-slide clock. A timer that blurs the image at 7.5 s manufactures urgency of a quite different kind — closer to a reaction-time task than to a workload. This caveat matters for anyone tempted to read the JAS result as “AI reliance rises in busy departments”; that is an extrapolation, not a finding.
The AI was accurate on about half the slides, and the error mix was a property of dataset shift rather than a designed manipulation. That is more realistic than fabricated advice, but it means the exposure each participant received to congruent-wrong advice was not controlled — it fell out of which patches happened to break the detector.
Explanations were always present, so this measures confirmation bias with prototype-based xAI, not bias from a bare number. The authors flag that xAI is known to increase trust and may therefore have inflated alignment. There is no no-explanation arm.
No clinical context, no alternative views, isolated patches. The authors list this as a realism limitation; it also means the task is closer to a psychophysics judgement than to sign-out.
Ground truth is dataset annotation, reviewed by at least one pathologist per the source datasets. TCP ground truth is itself irreproducible — which is the premise of the study — so “error” here is deviation from one annotation, not from truth.

What it would change here

Any reader study this group runs on an AI-assisted scoring task should collect the unaided baseline first, with a washout, rather than a single AI-assisted round. Without the baseline none of the analysis above is possible, and a study that only measures agreement-with-AI cannot distinguish a corrected error from a reinforced one.
Record confidence per case. It is one extra click and it carried a real result here.
The JAS is a cheap, interpretable summary of reliance that this wiki had no instrument for.
The finding argues for showing the model’s answer after the reader commits to their own, not beside it — an interface decision, not a modelling one.

How it connects

Rosbach 2025 — Automation Bias in AI-Assisted Tumour Cell Percentage Estimation under Time Pressure — the same experiment analysed for the other bias; read the two together or the 7% automation-bias rate and the 0.61 confirmation-bias slope will look like independent replications when they are one dataset.
Rosbach 2026 — Cognitive Biases in AI-Assisted Medical Decision Making — the review by the same group that places this in a literature of eight primary studies, and almost certainly counts this one as its single pathology study.
Cognitive Bias in AI-Assisted Diagnosis — the concept page holding the confirmation/automation handoff this paper half-measures.
Interobserver Agreement — TCP’s irreproducibility is the study’s premise and the reason the task was chosen; this is what a second-opinion effect looks like when the second opinion is a model.
Biomarker Cut Points — TCP matters because it is thresholded for molecular assay eligibility; the companion paper uses the 25% direct-sequencing threshold explicitly.
Cell Detection and Counting — the AI here is an FCOS detector counting neoplastic and non-neoplastic nuclei, exactly the route this wiki records for Ki-67-style indices.
Laboratory Workload Measurement — the paper’s own caveat that clinical time pressure is specimen volume rather than a per-case clock is a workload statement, and it bounds what the timer result can be used for.

Open questions

Does the CHI ’25 camera-ready match the preprint? It is two pages longer. Every figure here is the preprint’s. [unverified]
Is there a de-anonymised code repository post-acceptance? Only the anonymous.4open.science review link is in the preprint. [unverified]
Does the effect survive the asymmetric-filter objection in point 1 — i.e. what do Models 2 and 3 look like with the same inclusion rule applied to both? Answerable from the released data if it exists.
Would showing the AI’s answer only after the reader commits remove the effect? The design already contains the unaided condition needed to test it, but the experiment does not compare simultaneous versus sequential presentation.