Bias in large AI models for healthcare

Estimated reading time: 5 minutes · Last updated: 17 October 2026

The scoping review, as first reported by Nature, synthesizes evidence from 92 studies on medical bias in large AI models for medicine and healthcare. It covers general-purpose and medical-specific models and finds persistent problems across clinical decision support, patient communication, documentation and medical education. The review documents measurable subgroup gaps — for example, representation and diagnostic errors in dermatology and cardiology — and identifies three core failings: inconsistent fairness definitions, insufficient representative benchmarks and limited real-world validation. It calls for clinically grounded, reproducible evaluation pipelines, representative datasets and deployment-aware validation frameworks.

What the review measured and why it matters

The review assembled literature from five databases—ACM Digital Library, IEEE Xplore, Web of Science, PubMed and arXiv—covering papers posted or published between 1 January 2023 and 30 April 2026 and screened 161 records for full-text eligibility before retaining 92 studies. Its authors coded recurring concepts across datasets, model classes and clinical scenarios to build a taxonomy of medical bias that maps sources to manifestations and detection strategies.

This mapping matters because the review shows bias is not a single failure mode but an intersectional challenge that arises at the data, model and evaluation stages. For clinicians and health systems, the practical impact is concrete: subgroup errors or representational gaps can change diagnoses, triage and treatment recommendations across patient groups and specialties.

Where bias appears: scenarios and specialties

The review groups findings around four practical settings—supporting clinical decisions, communicating with patients, creating medical records and educating health professionals—and also compares results by clinical specialty. Across these slices, race, sex and skin tone emerge as dominant axes of disparity. Examples include a dermatology sample of 4,000 AI-generated images that were 89.8% light skin and 10.2% dark skin, and GPT-4 changing a cardiology diagnosis after only a patient sex swap, with notable female underperformance in several tests.

Specialty-level patterns differ: pulmonary imaging studies reported subgroup performance drops of roughly 6.8–11.6% across sex and race for specific tasks; an ophthalmology benchmark showed demographic-parity improvements when FairCLIP methods were applied, reducing race-related parity differences from 15.35% to 6.07% in a zero-shot setting. These concrete figures show aggregate accuracy can mask important subgroup gaps.

How bias is detected in practice

Researchers use three complementary detection strategies: answer-consistency checks that compare outputs on counterfactual inputs; statistical group-wise fairness metrics such as Demographic Parity, Equal Opportunity and Equalized Odds; and structured human expert assessment. The review documents large-scale experiments—Omar et al. ran 1,728,000 model runs across 1,000 emergency cases, 31 sociodemographic variations and nine LLMs—to surface systematic recommendation differences tied to identity cues.

Counterfactual methods include synthetic patient generation and mutation-based variations, but the authors caution that mutated cases can lose clinical plausibility and may miss intersectional effects. Where counterfactual and group statistics are infeasible, human raters remain essential: the review reports its own screening produced 157 consistent labels, four disagreements, a 97.5% agreement rate and a Cohen’s kappa of 0.949.

Mitigation strategies and remaining gaps

Mitigations fall into pre-processing (data augmentation, reweighting), in-processing (fine-tuning, loss modifications) and post-processing (output adjustments, decoding constraints). The review highlights examples such as FairCLIP and domain-specific benchmarks (FairMedFM, Harvard-FairVLMed) but notes in-practice limitations: in-processing changes are often costly or infeasible for closed-source, large-scale models and many mitigation techniques trade off fairness against utility.

Crucially, the authors identify three persistent gaps: a lack of unified, clinically grounded fairness definitions; insufficient representative benchmarks and datasets; and limited real-world validation and evidence about fairness–performance trade-offs. They argue these gaps must be closed before large AI models can be deployed safely at scale in healthcare.

How progress could unfold

The bottom line

The review compiles quantifiable evidence that medical bias in large AI models is widespread and varied across scenarios and specialties. It shows that bias arises from data composition, model design and evaluation choices and that current detection and mitigation methods—answer-consistency checks, classic fairness metrics and expert review—are necessary but insufficient. To reduce patient harm, the authors argue for clinical benchmarks, representative datasets and real-world validation that test subgroup outcomes, not just aggregate accuracy. Health systems piloting these models should require such evidence before adoption.

What to watch

• watch for publication of reproducible, specialty-specific fairness benchmarks; no date has been set.
• watch for wider deployment studies that validate model behaviour in clinical workflows; no date has been set.

This article is information, not medical advice. Anyone acting on it should speak to a qualified professional.