Only 13% of 166 articles about artificial intelligence and health equity clearly defined what “health equity” means, according to a review led by Georgetown University researcher Caroline R. Efird. That gap matters: if health systems do not say whose health they aim to improve, AI tools could direct help toward people who already have the most advantages.
AI is increasingly used in health care, but it can also produce unfair results. Algorithms learn from data, and if the data or design leaves out some groups, the tools may work less well for them. Efird, an assistant professor in Georgetown’s Department of Health Management and Policy, warned that such bias can widen health disparities. She said weak transparency and oversight pose particular risks for marginalized populations. The review was published Oct. 8, 2026, in Health Affairs Scholar.
The researchers examined 166 articles on AI and health equity. Along with the finding that only 13% gave a specific definition of health equity, they found that 70% of the publications had author teams based in the United States. A largely U.S.-based body of research may miss important perspectives from other parts of the world, the authors said. They called for more global input on how AI can support fair health outcomes.
The review also found that nearly 30% of the articles included a figure, framework or chart to help explain how AI could be used equitably in research and care. Most—68%—discussed AI and health equity in broad terms rather than focusing on a particular health condition. Among papers that did name a condition, 5% addressed cardiovascular health, 5% chronic diseases generally and 5% dermatology or skin cancers. Smaller shares focused on COVID-19 (3%), aging (2%), diabetes (2%) and radiology (2%).
Those details point to two needs: clearer goals and more work tied to real health conditions. If “equity” stays vague, Efird said, resources may flow to groups that are already better served rather than to those with the greatest need. Research focused on specialties such as radiology, dermatology and cardiology could help identify where AI systems might miss patients or make less accurate decisions. Medical specialists may notice risks in training data that does not represent a wide range of patients.
The review’s authors say safeguards should begin before an AI tool is used and continue through its design, implementation and evaluation. They also call for teams that bring together data scientists, health equity researchers, health care providers, specialists and patients. Including people with different skills and experiences could help uncover bias and make tools more useful across communities.
The researchers’ goal is not simply to add AI to health care, but to make sure it improves care without deepening existing gaps. More inclusive research methods, clearer definitions and a wider range of voices could help health systems judge whether these tools serve all patients. As AI use grows, the choices made now about whose data and expertise count will shape who benefits.