More accuracy isn't more equity: health AI needs to measure whom it serves, not just how often it's right

By Beinsure · November 3, 2026.

Only 22 of the 166 papers on AI and health equity reviewed by Georgetown University researchers explicitly define what they mean by "health equity," according to the review itself. The analysis covers publications from 2018 to early 2025. In addition, 70% of first authors (116 papers) work at US institutions, and none in South America.

Sixty-eight percent of the texts discuss equity in general, without focusing on a specific disease. Sixty-four percent are commentaries, editorials or opinion pieces. According to lead author Caroline R. Efird, a vague definition of equity increases the risk that resources end up with the most advantaged group rather than the one that needs them most.

The article recalls documented cases in which accuracy is not enough. In a 2019 Science study, a US patient-management algorithm assigned the same risk score to Black patients who were substantially sicker than white patients. The reason: it predicted medical spending, not clinical need, and spending also reflects unequal access to care.

Another example is a 2024 Nature Medicine experiment with 389 dermatologists and 459 primary care physicians from 39 countries. AI assistance improved overall diagnostic accuracy but widened the gap between light and dark skin among general practitioners. Note: it was conducted with images and limited clinical context in a simulated setting, not with real patient outcomes.

The article adds the disability angle. Systems trained on data that mostly represent people without disabilities may perform worse for those who speak, move or communicate differently. It also cites STAT's "Denied by AI" investigation, which according to the outlet linked algorithmic predictions to cuts in rehabilitation coverage in Medicare Advantage. That investigation also reported pressure from UnitedHealth on medical staff; these are STAT's claims, reported here as such, not as proven facts.

On the regulatory front, the text contrasts the European AI Act, from 2024, which requires risk management, data quality and human oversight in high-risk uses, with a US shift toward deregulation since 2025, according to the researchers. In the US there are more than 1,600 AI-enabled medical devices authorized by the FDA. John Whyte, CEO of the American Medical Association, warned that autonomous medical models are "not ready" and that much of the evidence comes from simulations.

**Our take.** Our thesis is simple: in medicine, the metric we choose is a values decision disguised as a technical one. The case of spending as a proxy for illness illustrates this well. The model was "right" by its own objective, but the objective was poorly chosen. That fewer than one in six studies define equity shows the field is still debating ends before means.

The dermatology experiment points to something more uncomfortable: the unit of evaluation should be the clinician-plus-algorithm pair, not the model in isolation. A system can excel on a benchmark and worsen inequalities in the clinic. Broadly, this fits what we see in other regulatory frameworks for clinical AI: the defensible value lies in verification, traceability and external validation, not just in accuracy rates.

Some nuance is also warranted. The review is a literature review, mostly of opinion pieces, and does not by itself measure real harm. And the article comes from an insurance outlet, which views the issue through the lens of risk and liability. Even so, the cases it cites are verifiable and point in the same direction.

In the short term, the transition will be uneven: those with better-represented data will benefit sooner, and those without will be left behind or receive lower-quality care. It is a real risk, one managed through representative data, clinical specialists involved in design, subgroup audits and human oversight. In the long term, the same technology can expand access to diagnosis for underserved populations and bring us closer to eradicating diseases. But that promise will only belong to everyone if equity is defined and measured from the start, rather than added as a patch afterward.