AI improves detection but may reshape clinical reasoning

Artificial intelligence helped clinicians find more cancers and precancerous lesions in randomized trials of mammography and colonoscopy, but it did not measurably improve physicians’ reasoning through complex diagnostic cases in the one randomized trial of open-ended diagnosis reviewed. A critical narrative review published in Frontiers in Digital Health concluded that the usefulness of medical AI depends less on how accurately a model performs alone than on the task, the form of its output and the degree to which clinicians trust it.

The central distinction is between bounded perceptual work, such as flagging an abnormality in an image or during an endoscopic procedure, and open-ended reasoning, which requires clinicians to synthesize histories, examinations and test results into competing diagnoses and a management plan. In the Swedish Mammography Screening with Artificial Intelligence trial, AI-supported reading increased breast-cancer detection from 5.0 to 6.4 per 1,000 women, a relative increase of about 29%, while reducing screen-reading workload by 44.2%. The completed trial also found a non-inferior interval-cancer rate: cancers arising between screening rounds numbered 1.55 per 1,000 participants with AI support and 1.76 per 1,000 with standard double reading. Sensitivity was higher with AI support, at 80.5% versus 73.8%, while specificity was 98.5% in both groups. The analysis did not measure breast-cancer mortality or overdiagnosis, leaving those outcomes unknown.

Real-time computer-aided detection produced a similar pattern during colonoscopy. In a multicenter randomized trial involving 685 patients, the adenoma detection rate rose from 40.4% without AI assistance to 54.8% with it, corresponding to a relative risk of 1.30. The increase was concentrated among smaller adenomas, including those no more than 5 millimeters and those measuring 6 to 9 millimeters. A second randomized trial involving 1,058 patients also reported an increase, from 20.3% to 29.1%. Neither colonoscopy trial assessed whether the higher detection rate translated into lower colorectal-cancer incidence or mortality, and the differences in baseline detection rates mean that the absolute gains should not be treated as interchangeable estimates.

Open-ended diagnostic reasoning produced a notably different result. In a randomized trial, 50 attending and resident physicians in internal, family and emergency medicine worked through six complex clinical vignettes with conventional digital resources alone or with the same resources plus GPT-4. Blinded experts assessed their performance with a validated diagnostic-reasoning rubric. The median score was 76% with access to the model and 74% without it, an adjusted difference of two percentage points. The confidence interval allowed a modest benefit and therefore represented a null finding within a bounded range, not proof that the approaches were equivalent. GPT-4 achieved a median score of 92% when used alone on the same vignettes, 16 percentage points above physicians using conventional resources. The trial did not directly test why the model’s standalone performance was not reproduced in collaborative use.

That gap may reflect how the systems are used, although the review presents that explanation as a hypothesis rather than an established mechanism. During mammography and colonoscopy, the algorithm directs attention to a specific location that a clinician can inspect at modest cost. A language model may instead supply a fluent diagnostic conclusion whose verification requires nearly the same reasoning needed to reach the conclusion independently. The review proposes that calibrated use involves forming one’s own hypotheses first, treating the output as one piece of evidence and checking disagreement, whereas uncalibrated use allows the algorithm to anchor judgment. These are proposed poles on a continuum, not clinician categories directly observed in the studies.

Experimental studies indicate that incorrect advice can sharply reduce performance. In a simulated mammography reader study involving 27 radiologists, correct BI-RADS assignments fell when the system provided incorrect suggestions. Accuracy declined from 79.7% to 19.8% among inexperienced readers, from 81.3% to 24.8% among moderately experienced readers and from 82.3% to 45.5% among very experienced readers. The design compared readers with their own earlier performance and stratified results by experience, but it deliberately included incorrect advice in 12 of 40 test cases, a much higher error rate than would be expected from a deployed system. It establishes susceptibility under uncertainty rather than the likely size of harm in routine care.

Missing alerts can also alter behavior. In a small eye-tracking study, 10 breast-screening readers interpreted 60 examinations with and without AI prompts. On cases the AI failed to flag, median sensitivity fell from 71% without assistance to 39% with it, and readers’ gaze patterns changed only when prompts were incorrect. The test set was deliberately enriched for errors, so the findings remain a signal rather than a realistic effect estimate. A separate randomized trial involving 44 physicians who had completed 20 hours of AI-literacy training found that deliberately erroneous language-model advice reduced reasoning accuracy. Because every participant was trained and there was no untrained comparison group, the experiment shows that this training did not prevent the observed effect, but it does not establish that training itself is ineffective.

The review also identified a possible threat to skills retained without AI. In a retrospective pre–post study of 1,443 unassisted colonoscopies performed by 19 experienced endoscopists in four Polish centers, adenoma detection fell from 28.4% before computer-aided detection was introduced to 22.4% afterward, an absolute decrease of six percentage points and a relative reduction of about 21%. Exposure to AI remained associated with lower detection in multivariable analysis. The finding is consistent with clinicians substituting automated prompts for their own vigilance, but the design cannot exclude changes in case mix, temporal trends or seasonal variation. It also does not show that any decline is lasting or occurs in other specialties.

Problems can arise even when individual clinicians never see an obvious error. An external validation of the Epic Sepsis Model across 38,455 hospitalizations found an area under the receiver operating characteristic curve of 0.63, with 33% sensitivity, 83% specificity and a positive predictive value of 12%. About two-thirds of sepsis episodes were missed and roughly 88% of alerts were false positives. A separate audit of a population-health algorithm used expenditure as a proxy for illness, systematically under-identifying Black patients with the same clinical burden because less was historically spent on their care. In the authors’ simulation, replacing expenditure with a direct measure of illness increased the share of Black patients selected for additional support from 17.7% to 46.5%. That was a modeled counterfactual rather than the measured result of deploying a corrected system.

The evidence base remains uneven. The review was based on searches of Scopus, PubMed/MEDLINE, SciELO and LILACS in English, Spanish and Portuguese, followed by purposive selection by one reviewer and appraisal of nine anchor studies. It included randomized and prospective evidence, reader experiments, pre–post comparisons and external validation, but no meta-analysis, duplicate screening or exhaustive search. Most evidence of benefit came from large trials measuring intermediate outcomes, while evidence of harm came mainly from small or design-limited studies. The reviewer also identified no eligible study of human–AI diagnostic reasoning in Latin America within the search, although this was not a proof that no such research exists. Local validation and a broader regional review would be needed before the true regional evidence base could be characterized.

The practical implication is not that AI should be treated uniformly as either a cognitive amplifier or a source of deskilling. Evaluation must directly compare assisted and unassisted clinician performance, measure whether clinicians disagree with or override the system and track performance without assistance. The review’s proposed safeguards—including periodic audits of unassisted competence, subgroup monitoring and documentation of why clinicians accept or reject algorithmic advice—remain research proposals rather than established standards. Likewise, explainable AI is not automatically a safeguard: post-hoc explanations can be unstable or contradictory and may increase trust without helping clinicians distinguish correct output from incorrect output. Ultimately, the studies support a conditional conclusion: AI can improve clinicians’ detection when it provides a bounded, verifiable cue, but a model’s impressive standalone performance is not evidence that physicians will reason better with it, and apparent safety cannot be inferred from the scarcity of studies demonstrating harm.

Subject of Research: Effects of artificial intelligence assistance and overreliance on physicians’ clinical reasoning and diagnostic performance

Article Title: Augmentation or erosion? Artificial intelligence and the transformation of clinical reasoning: a critical narrative review

Article References: Sánchez, J. D. (2026). Augmentation or erosion? Artificial intelligence and the transformation of clinical reasoning: a critical narrative review. Frontiers in Digital Health, 8, Article 1949323. https://doi.org/10.3389/fdgth.2026.1949323

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1949323

Keywords: artificial intelligence, clinical reasoning, diagnostic accuracy, automation bias, large language models, mammography screening, computer-aided colonoscopy, medical AI, clinician deskilling, algorithmic bias, Augmentation, erosion