Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index; 5 of the 14 sources its claims cite were checked for support.

Automation Bias in Large Language Model Assisted Diagnostic Reasoning Among AI-Trained Physicians

Open access
Qazi, Ihsan Ayyub, Ali, Ayesha, Khawaja, Asad Ullah, Akhtar, Muhammad Junaid, Sheikh, Ali Zafar, Alizai, Muhammad Hamad · 2025 · medRxiv

Abstract

Whether AI-trained physicians are vulnerable to automation bias when LLM use is voluntary remains unknown. Erroneous LLM recommendations significantly degrade performance...

Overview

CoChat Summary

TL;DR

In a single-blind randomized trial, 44 AI-trained Pakistani physicians diagnosed six clinical vignettes with voluntary access to ChatGPT-4o recommendations; those offered deliberately flawed advice in 3 of 6 vignettes scored substantially lower on diagnostic reasoning than those given error-free advice. The authors conclude that erroneous LLM output induces automation bias even among physicians with dedicated AI-literacy training.

Key findings

  • Primary outcome (diagnostic reasoning accuracy): control 84.9% (SD 19.7%) vs treatment 73.3% (SD 30.5%); adjusted mean difference −14.0 percentage points (95% CI −18.9 to −9.1; P < .0001). (The abstract reports the same point estimate with a CI of −8.3 to −19.7.)
  • Secondary outcome (top-choice diagnosis accuracy per case): 90.5% (SD 28.9) control vs 76.1% (SD 42.5) treatment; adjusted difference −18.3 percentage points (95% CI −26.6 to −10.0; P < .0001).
  • AI consultation was voluntary and similar across arms: 68.9% treatment vs 66.7% control (P = .69).
  • Subgroup analyses: larger declines among physicians at/above the median 10 years of practice (−16.6 pp, 95% CI −23.1 to −10.1; P < .0001) vs −9.1 pp (95% CI −18.1 to −0.1; P = .0474) as reported; frequent LLM users (≥ weekly) −11.0 pp (95% CI −18.5 to −3.6; P = .0037) vs −10.7 pp among less frequent users (95% CI −24.5 to 3.1; P = .1285); male physicians −25.8 pp (95% CI −33.8 to −17.7; P < .0001) vs −2.1 pp among female physicians (95% CI −9.8 to 5.5; P = .5839).
  • Grading reliability: Krippendorff's α = 0.93 across three blinded graders; Cronbach's α = 0.80.

Methods

  • Single-blind RCT (ClinicalTrials.gov NCT06963957), conducted June 20–August 15, 2025, at Lahore University of Management Sciences and via video conferencing; 44 physicians (22 per arm) registered with the Pakistan Medical and Dental Council, holding MBBS degrees, all having completed a 20-hour AI-literacy training. Participants were blinded to the study aims; randomization was known only to the study administrator.
  • Participants diagnosed 6 vignettes (internal medicine, cardiology, neurology, pediatrics, infectious disease, emergency medicine) in 75 minutes, with voluntary opt-in access to pre-generated ChatGPT-4o recommendations; control received unmodified output, treatment received clinically significant embedded errors in 3 of 6 vignettes. Google "AI Overviews" were blocked via browser extension; conventional online medical resources were permitted. 264 cases completed (132 per arm).
  • Primary outcome scored by three blinded physicians using an expert-validated rubric (differential diagnosis accuracy, supporting/opposing evidence, recommended next steps); analysis by intention-to-treat at the case level using linear mixed-effects models with random effects for participant and case; target sample was 50 (25 per arm), with post-hoc power ≥80% given the observed 14.0 pp effect.

Limitations

  • Stated by the authors: vignettes may not capture real-world complexity (context, time pressure, multimorbidity); errors were physician-introduced and may be less subtle than real AI errors; findings are limited to ChatGPT-4o; the single-session design cannot show whether bias persists or fades with repeated use; mitigation strategies were not explored; broader healthcare professionals (nurses, physician assistants) were not studied.
  • Evident from the text: enrollment fell short of the 50-participant target (44 enrolled, 88% of target); the subgroup analyses were not adjusted for multiple comparisons; the subgroup reporting is internally inconsistent, describing the −16.6 pp group as "at or above the median years of practice" while calling the −9.1 pp group "more experienced physicians"; and the trial was conducted at a single country's institutions with a preprint that has not been peer reviewed.

Model-written summary of a shortened version of the paper’s full text (it is longer than we read). Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1101/2025.08.23.25334280

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
From the PDF

Verification

Claims the paper makes

3 Valid · 2 Not valid · 3 Not checked · 31 Own results and interpretations

7 more claims

Create a free account to see all 39 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 17
  • Probably not a substantial claim. 18

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 31 references

Research works printed but not linked: 2

  1. American Medical Association. (2025, May 8). AI in medical education. https://www.ama-assn.org/education/changemeded-initiative/ai-medical-education
  2. American Medical Association. (2024, February 26). Future of health: The emerging landscape of augmented intelligence (AI) in health care. American Medical Association. Retrieved September 3, 2025, from https://www.ama-assn.org/system/files/future-health-augmented-intelligence-health-care.pdf

Cited by

See all 16

Metadata

Authors
Qazi, I. A., Ali, A., Khawaja, A. U., Akhtar, M. J., Sheikh, A. Z., & Alizai, M. H.
Affiliations
  • Lahore University of Management Sciences (PK)
  • King Edward Medical University (PK)
  • Lahore General Hospital (PK)
Corresponding author
Qazi, Ihsan Ayyub
Venue
medRxiv
Publisher
openRxiv
Published
2025-08-26
OpenAlex
W4413640513
Cited by
16
References
31
Citations by year
2026: 14 · 2025: 2
Type
posted content
Language
en
Open access
Locations
medRxiv (acceptedVersion, cc-by-nc-nd)
Funders
Lahore University of Management Sciences
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Health Sciences › Medicine › Family Practice › Clinical Reasoning and Diagnostic Skills
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
automation bias, large language models, AI-assisted diagnosis, GPT-4o, human-AI collaboration, AI literacy training, differential diagnosis, patient safety, human oversight of AI, clinical vignettes, prompt engineering, critical evaluation of AI outputs, Pakistan, South Asia

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Automation Bias in Large Language Model Assisted Diagnostic Reasoning Among AI-Trained Physicians. https://app.cochat.ai/papers/10.1101/2025.08.23.25334280
The paper
Qazi, I. A., Ali, A., Khawaja, A. U., Akhtar, M. J., Sheikh, A. Z., & Alizai, M. H. (2025). Automation Bias in Large Language Model Assisted Diagnostic Reasoning Among AI-Trained Physicians. medRxiv. https://doi.org/10.1101/2025.08.23.25334280

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.