Physicians Struggle to Detect AI Errors in Clinical Decision-Making, Study Finds
Artificial intelligence (AI) holds promise for transforming healthcare by assisting physicians in patient classification and treatment decisions. However, new research from the University of the Basque Country in Spain reveals a troubling trend: physicians often over-trust AI recommendations, even when presented with contradictory clinical outcomes. This study, published in the open-access journal PLOS Digital Health, highlights the difficulty doctors face in spotting and correcting AI errors — an issue that could impact patient safety as AI tools become more widespread.
The research involved 223 physicians who participated anonymously in controlled online experiments. They were asked to imagine treating patients with a rare illness using an experimental treatment guided by AI-generated recommendations. Although the physicians were informed that the AI indicated which patients might benefit more or less from the treatment, the underlying experimental design purposely did not align with these suggestions. In one scenario, the treatment proved moderately effective for all patients regardless of AI advice; in the other, it was completely ineffective across the board.
Despite clear data showing uniform treatment outcomes, physicians consistently rated the AI system as reliable and failed to adjust their clinical decisions accordingly. In particular, many did not recognize the lack of treatment efficacy in the second experiment, revealing a deep-rooted bias toward AI guidance. This suggests that even trained medical professionals can struggle to critically evaluate AI recommendations when evidence seems to challenge algorithmic authority.
Lead author Aranzazu Vinas emphasized the implications: “Physicians mostly trusted the AI’s classifications and had trouble learning from feedback that contradicted the algorithm. This highlights potential risks in clinical practice when doctors uncritically rely on AI outputs.” Co-author Helena Matute noted that the common assumption of humans as ultimate controllers of algorithms may be overly optimistic, as the study demonstrates difficulties in processing contradictory evidence.
These findings raise important considerations for the integration of AI in healthcare workflows. While AI can process vast amounts of data and generate predictive models, its fallibility requires vigilant oversight. The tendency of physicians to place undue trust in imperfect algorithms underscores the need for improved training and protocols that foster critical appraisal of AI tools. Enhanced transparency in AI decision processes and better feedback mechanisms could empower doctors to better detect errors and mitigate patient risk.
Co-author Fernando Blanco concluded, “Investigating how humans interact with AI, especially the errors they make, is crucial. Understanding these dynamics will help us develop strategies that maximize the benefits of human-AI collaboration while minimizing errors.” As AI continues to permeate clinical settings, these insights urge a cautious and evidence-informed approach.
This research underscores a paradox of human-AI partnership in medicine: the promise of enhanced care through algorithms coexists with cognitive biases that may hinder appropriate skepticism. Future efforts must bridge this gap to ensure AI serves as a genuine aid rather than an unchecked authority in patient care.
Subject of Research: People
Article Title: Doctors vs. Algorithms: Physicians, too, struggle to learn from evidence that contradicts AI suggestions
News Publication Date: 9-Jul-2026
Web References: https://doi.org/10.1371/journal.pdig.0001490
References: Vinas A, Blanco F, Matute H (2026) PLOS Digital Health 5(7): e0001490
