Only 13% of AI health-equity guidance defined equity, Georgetown review finds

Only 22 of 166 articles offering guidance on AI and health equity explicitly defined what equity meant. In comments published November 2, Georgetown researcher Caroline R. Efird explained why that gap matters: vaguely defined goals risk directing resources toward the most advantaged patients rather than those who need them most.

The underlying scoping review—a study that maps existing literature—was published October 8, 2026, in Health Affairs Scholar. Efird and colleagues examined peer-reviewed guidance on equitable AI use in healthcare settings, health systems and health services research. Medical Economics reported November 3 that the team searched four databases in January 2025, selecting articles published from 2018 through early 2025.

Most of the 22 articles that supplied a definition drew on the World Health Organization’s approach: health equity means the absence of unfair differences between groups that can be avoided or remedied. Fifteen of those articles came from U.S.-based author teams. The finding concerns how the selected literature described its goals; it does not measure whether the AI tools discussed actually delivered equitable care.

When health equity is vaguely defined, there is a risk that resources will be primarily allocated to the most advantaged group, rather than the group that needs the resources the most.

Caroline R. Efird, in Georgetown University Medical Center comments published by Medical Xpress

The literature also had a pronounced geographic imbalance. First authors at U.S. institutions wrote 116 of the 166 articles, or 70%. The United Kingdom accounted for 10 and Canada for nine. No first author was based at an institution in South America; Africa and Australia each accounted for one article. Efird warned that a U.S.-centered perspective likely has blind spots and called for more global perspectives.

Much of the guidance stayed broad rather than addressing particular medical problems. Sixty-eight percent of articles did not focus on a specific disease or condition. Cardiovascular conditions, chronic disease and dermatology or skin cancers each appeared in 5% of the articles. COVID-19 accounted for 3%, while aging, diabetes and radiology each accounted for 2%. The authors called for more scholarship connecting AI use to specific conditions and health outcomes.

Dermatology illustrates the risk the authors describe. Clinical textbooks and public image databases have historically depicted skin lesions mostly on lighter skin tones. An AI tool trained only on that material could struggle to recognize lesions on darker skin. Efird argues that specialists may be better equipped than data scientists to spot diagnostic disparities caused by training data that fails to represent diverse patients.

The authors want equity considered throughout model design, implementation and evaluation—not left to one profession or one stage of development. They recommend teams that include data scientists, equity researchers, healthcare providers, specialists and patients. Efird says that mix has the potential to reduce model bias; the review does not establish that it will.

• Grant funders and academic journals should require AI researchers to define equity explicitly and report results for subgroups and marginalized populations.
• Health systems should adopt equity-focused AI governance with continual oversight, rather than treating evaluation as a one-time step.
• National regulators should incorporate frameworks based on the WHO definition of equity into their evaluations.

These are proposed requirements, not proven remedies. Sixty-four percent of the reviewed articles were commentaries, editorials or opinion pieces. Because the review synthesized recommendations rather than outcomes from studies or trials, it could not determine whether those recommendations reduce health inequities. Its scope was also limited to English-language articles, and the authors acknowledged that more work had appeared since the January 2025 search ended.