Augmentation or erosion? Artificial intelligence and the transformation of clinical reasoning: a critical narrative review
- 1Facultad de Ciencias de la Salud y Bienestar Humano, Universidad Tecnológica Indoamérica, Quito, Ecuador
Abstract
Background: Artificial intelligence (AI) systems are used to support diagnosis, but a model’s accuracy does not establish that clinicians reason more accurately when they use it. This review appraises how AI decision support and large language models affect physicians’ diagnostic performance and the risks of relying on them.
Methods: This critical narrative review follows the Scale for the Assessment of Narrative Review Articles. Scopus, PubMed/MEDLINE, SciELO and LILACS were searched in July and August 2026 in English, Spanish and Portuguese. Eligible studies, published from 2015 onward, compared AI-assisted with unassisted clinicians, measured reliance on AI or unassisted performance after exposure to it, or externally validated deployed models. A single author selected sources purposively. Nine primary studies were appraised with an adapted Critical Appraisal Skills Programme framework.
Results: In randomized trials of perceptual tasks, AI support improved detection. In the MASAI mammography screening trial, cancer detection rose from 5.0 to 6.4 per 1,000 women, reading workload fell by 44%, and the interval-cancer rate was non-inferior. In colonoscopy, computer-aided detection raised the adenoma detection rate from 40.4% to 54.8%. In the only randomized trial of open-ended diagnostic reasoning, access to GPT-4 did not improve physicians’ scores (76% vs. 74%; P=.60), although the model alone scored 92%. Evidence of harm came from smaller or non-randomized studies: reader accuracy fell sharply when AI suggestions were incorrect or the AI missed a cancer; physicians trained in AI literacy were still misled by erroneous model advice; unassisted adenoma detection fell from 28.4% to 22.4% after routine AI use in one pre–post study; and a deployed sepsis model discriminated poorly on external validation (area under the curve 0.63). No eligible study from Latin America was identified.
Conclusions: In the studies reviewed, AI assistance improved clinician performance in bounded detection tasks but not in open-ended diagnostic reasoning, where evidence is limited to one vignette-based trial. The evidence that reliance on AI can impair performance comes from single or design-limited studies and does not establish how large or lasting such effects are. Benefit therefore cannot be inferred from model accuracy alone; assisted and unassisted clinician performance should be measured directly.
Related
- Clinical Reasoning, Pathophysiology, and Artificial Intelligence in Critical Care: A Conceptual Framework for Integrating Decision-Making
- Toward an epistemic grounding of clinical reasoning: artificial intelligence and psychological diagnosis.
- Artificial intelligence in biomedical visualization and education: a narrative review
- Reference Standards in Artificial Intelligence Detection of Apical Periodontitis: A Narrative Review of What Reported Accuracy Measures
- Bias Mitigation Across the Healthcare Artificial Intelligence Lifecycle: A Structured Narrative Review
BibTeX
@article{sanchez2026augmentation,
title = {{Augmentation or erosion? Artificial intelligence and the transformation of clinical reasoning: a critical narrative review}},
author = {José Daniel Sánchez},
journal = {Frontiers in Digital Health},
year = {2026},
month = oct,
doi = {10.3389/fdgth.2026.1949323},
url = {https://doi.org/10.3389/fdgth.2026.1949323},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Augmentation or erosion? Artificial intelligence and the transformation of clinical reasoning: a critical narrative review
AU - José Daniel Sánchez
PY - 2026
DA - 2026/10/21
JO - Frontiers in Digital Health
DO - 10.3389/fdgth.2026.1949323
UR - https://doi.org/10.3389/fdgth.2026.1949323
AB - Background: Artificial intelligence (AI) systems are used to support diagnosis, but a model’s accuracy does not establish that clinicians reason more accurately when they use it. This review appraises how AI decision support and large language models affect physicians’ diagnostic performance and the risks of relying on them. Methods: This critical narrative review follows the Scale for the Assessment of Narrative Review Articles. Scopus, PubMed/MEDLINE, SciELO and LILACS were searched in July and August 2026 in English, Spanish and Portuguese. Eligible studies, published from 2015 onward, compared AI-assisted with unassisted clinicians, measured reliance on AI or unassisted performance after exposure to it, or externally validated deployed models. A single author selected sources purposively. Nine primary studies were appraised with an adapted Critical Appraisal Skills Programme framework. Results: In randomized trials of perceptual tasks, AI support improved detection. In the MASAI mammography screening trial, cancer detection rose from 5.0 to 6.4 per 1,000 women, reading workload fell by 44%, and the interval-cancer rate was non-inferior. In colonoscopy, computer-aided detection raised the adenoma detection rate from 40.4% to 54.8%. In the only randomized trial of open-ended diagnostic reasoning, access to GPT-4 did not improve physicians’ scores (76% vs. 74%; P=.60), although the model alone scored 92%. Evidence of harm came from smaller or non-randomized studies: reader accuracy fell sharply when AI suggestions were incorrect or the AI missed a cancer; physicians trained in AI literacy were still misled by erroneous model advice; unassisted adenoma detection fell from 28.4% to 22.4% after routine AI use in one pre–post study; and a deployed sepsis model discriminated poorly on external validation (area under the curve 0.63). No eligible study from Latin America was identified. Conclusions: In the studies reviewed, AI assistance improved clinician performance in bounded detection tasks but not in open-ended diagnostic reasoning, where evidence is limited to one vignette-based trial. The evidence that reliance on AI can impair performance comes from single or design-limited studies and does not establish how large or lasting such effects are. Benefit therefore cannot be inferred from model accuracy alone; assisted and unassisted clinician performance should be measured directly.
SN - 2673-253X
ER -