AI health equity guidance remains vague as the literature grows
Only 22 of 166 peer-reviewed articles on AI and health equity defined health equity explicitly, leaving a large body of guidance open to interpretation at a moment when automated tools are moving into clinical workflows. The Georgetown scoping review, published October 8 in Health Affairs Scholar, read like a warning about language as much as technology: if equity is not defined, implementation can drift toward the patients already easiest to serve. In practice, that can mean cleaner data, smoother adoption, and better reimbursement pathways for the same groups that already receive the most reliable care.
The review examined commentaries, editorials, literature reviews and other articles offering guidance on equitable AI use in health care settings, health systems and health services research. Published between 2018 and early 2025, the corpus was dominated by broad conceptual discussions rather than disease-specific operational guidance. Most of the articles that did define equity leaned on the World Health Organization framework, which describes health equity as the absence of unfair, avoidable or remediable differences among groups of people.
A single sentence in a governance document can decide who gets counted and who gets missed.
That matters because AI systems are only as fair as the data, targets and assumptions used to build them. If the intended goal is never pinned down, models can be optimized for average performance while still underperforming in communities already carrying the heaviest burden of disease. Caroline R. Efird, Ph.D., an assistant professor in the Department of Health Management and Policy in the School of Health at Georgetown University, warned that vague definitions create a real danger of directing resources toward the most advantaged group rather than the group that needs them most.
A U.S.-weighted evidence base shapes the guidance
The review found a strong U.S. imprint across the literature. First authors at U.S. institutions wrote 116 of the 166 articles, or 70%. The United Kingdom contributed 10 and Canada nine, while no first author was based at an institution in South America. Africa and Australia each accounted for a single article.
That distribution does more than map authorship. It shapes which populations are visible, which regulatory assumptions dominate, and which inequities are treated as central. A U.S.-centric perspective on AI in health care can easily miss structural issues that appear differently in other systems, particularly where access barriers, data infrastructure and baseline health disparities do not mirror those in the United States.
The review also found that 68% of the articles addressed AI and health equity in general terms rather than in the context of a particular disease or condition. When topics were specific, cardiovascular conditions, chronic disease and dermatologic conditions including skin cancers each made up 5%. COVID-19 accounted for 3%, while aging, diabetes and radiology each accounted for 2%.
Those numbers reveal a striking imbalance. The everyday use cases where bias can quietly enter practice — triage, imaging, lesion recognition, chronic disease management and risk stratification — remain underdeveloped in the literature compared with the broad moral arguments about fairness. Specialty-specific expertise is essential here because the clinical consequences of biased training data differ by field. In dermatology, for example, training sets built mostly from lighter skin tones can leave an algorithm less able to recognize lesions on darker skin. In radiology and cardiology, the failure mode may be subtler: disparate sensitivity or calibration that changes who gets flagged for follow-up and who does not.
Human judgment still sits at the center of safe AI use
Across the reviewed articles, a consistent message emerged: AI should support clinical judgment rather than replace it. Direct patient-provider interaction was described as essential even when algorithms are embedded in decision support systems. That stance reflects a basic feature of medicine that software cannot reproduce. A model can rank probabilities, but it cannot fully interpret a patient’s story, weigh social context, detect when symptoms do not fit the statistical pattern, or understand why a recommendation may not be feasible in a given household.
The guidance also stressed that clinicians need training in the basic principles of AI and its limits if they are expected to use it responsibly. That is not a minor educational add-on. Without an understanding of how a model is trained, where it may fail and what kinds of bias can enter through data selection or label construction, physicians may place too much confidence in outputs that appear objective but are actually shaped by narrow inputs.
Health system policy came through as another recurring theme. The reviewed articles emphasized institutional AI policies designed to prevent harm, including race-conscious and gender-conscious algorithms in the United States where patterns of disease and access are tied to those variables. Efird argued that diverse, interdisciplinary teams are essential when algorithms are developed and implemented in clinical settings. Data scientists, health equity researchers, clinicians, specialists and patients each bring different forms of expertise, and those perspectives are necessary if a model is going to reflect real-world care rather than an abstract dataset.
Regulation is moving faster than the evidence base
The review placed the literature against a shifting policy backdrop. The European Union’s AI Act, passed in 2024, requires AI software used for medical purposes to include risk-mitigation systems, high-quality data sets, clear user information and human oversight. In the United States, the framework described in the review pointed in a different direction, with federal policy since early 2025 prioritizing deregulation. The authors contrasted that approach with a January 2025 executive order from President Donald Trump that revoked the Biden administration’s 2023 order on safe and trustworthy AI.
A separate legal obligation remains in force under the 2024 rule implementing Section 1557 of the Patient Protection and Affordable Care Act. Covered health care entities must make reasonable efforts to identify patient care decision support tools that use race, color, national origin, sex, age or disability as input variables and mitigate the risk of discrimination from each one. Those duties took effect May 1, 2025, yet the review did not address the rule. The absence is telling, because equity discussions can become detached from the compliance structures that will ultimately shape how AI is deployed in hospitals and clinics.
For physicians, the liability stakes are still unsettled. Deepika Srivastava, chief operating officer of The Doctors Company, has said physicians continue to carry the largest share of legal risk when an AI tool contributes to patient harm because malpractice law has not resolved how responsibility is divided among clinicians, health systems and developers. In the clinic, that uncertainty increases the importance of local governance, documentation and a clear understanding of what an algorithm can and cannot do.
Definitions, subgroup reporting and governance gaps
The authors called for grant funders and academic journals to require explicit definitions of equity and to demand subgroup reporting for marginalized populations. They also urged health systems to adopt equity-focused AI governance with continuous oversight and recommended that national regulators incorporate the WHO definition into evaluation frameworks.
The limitations are straightforward and consequential. Sixty-four percent of the reviewed articles were commentaries, editorials or opinion pieces, so the evidence base is built more on argument than on measured outcomes. The review also included only English-language articles. Because the synthesis focused on recommendations rather than intervention results, it cannot show whether any of the proposed governance models actually reduce inequities in care. The next test is not whether institutions can write about equitable AI in polished terms. It is whether those systems can prove, in specific clinical settings and across specific patient groups, that the tools perform fairly when the consequences are real.