Your skin has a secret, and scientists just figured it out. Or, more precisely, artificial intelligence may be able to infer race and ethnicity from health data even when nobody explicitly tells it to. That is a little like asking a toddler not to notice cookies on the counter: admirable in theory, wildly optimistic in practice.

A new Viewpoint indexed in PubMed, Impracticality of banning collection of data on ethnicity and race in artificial intelligence-enabled health care in France, tackles a thorny question: what happens when a country bans the collection of race and ethnicity data in the name of equality, but AI systems can still absorb and reproduce racial bias through the back door?

France has long followed a “color-blind” policy approach, meaning the state generally avoids collecting official data on race and ethnicity. The goal is noble: treat everyone as equal citizens, not as racial categories. But in health care, especially AI-enabled health care, the authors argue that this approach may be hiding the very inequities it hopes to prevent.

The Problem With Pretending Data Has No Accent

AI in medicine is moving fast. Algorithms can help read medical images, flag disease risk, support diagnosis, and guide treatment decisions. In the best-case version of the future, AI helps busy clinicians catch problems earlier, standardize care, and improve access for people who have historically been left waiting, misdiagnosed, or ignored.

That is the hopeful part, and it is real.

But AI does not arrive in a white coat carrying pure objectivity. It learns from data, and health data reflects the world we already live in: unequal access to care, different disease burdens, underdiagnosis, undertreatment, environmental exposures, racism, poverty, and all the bureaucratic confetti that comes with medical systems.

If a dataset underrepresents African and Afro-descendant patients, for example, an AI model trained on that dataset may perform worse for those groups. It might miss disease, misclassify risk, or recommend care based on patterns that fit the majority population better than everyone else. That is not “neutral.” That is a digital version of the same old health gap wearing a shiny new badge.

France’s Color-Blind Policy Meets a Very Nosy Algorithm

The Viewpoint focuses on France because its legal and political tradition strongly discourages collecting data on race and ethnicity. The idea is that by refusing to classify people racially, the state avoids reinforcing racial divisions.

The catch is that AI systems are not necessarily bound by the same social ideals. Research has shown that algorithms can infer race from medical images and other biological or clinical data in ways humans may not understand. So even if a dataset does not include a neat little “race” column, the model may still detect race-related patterns through proxies.

Those proxies might include geography, disease patterns, skin-related imaging features, socioeconomic markers, access histories, or other subtle signals. AI can be annoyingly good at finding patterns. Sometimes that is useful. Sometimes it is like hiring a bloodhound to organize your sock drawer: impressive, but you should ask what it is sniffing for.

This creates a paradox. If race and ethnicity data are banned, developers may be unable to measure whether an AI tool works equally well across different groups. Yet the model may still behave differently across those groups. In other words, the system can “see” race, but regulators and researchers may be prevented from checking what it is doing with that information.

That is a bad setup for patients, especially patients already more likely to be underserved.

Why This Matters for Health Equity

Health equity is not about giving everyone the same generic health care and hoping biology, history, and society politely cooperate. It is about making sure care works for people as they actually are.

For marginalized communities, biased AI could compound existing harms. A diagnostic model that performs less accurately for Black patients could delay treatment. A risk prediction tool trained mostly on majority populations could underestimate disease risk in minority groups. A medical device that has not been properly tested across skin tones, ancestry groups, or lived environments may look high-tech while quietly widening old gaps.

That last part matters. Health inequity rarely announces itself with a trumpet. It often shows up as small differences repeated thousands of times: a missed referral, a delayed scan, a lower risk score, a “come back later” when later is not safe.

The authors argue that race-conscious strategies are needed across the entire AI lifecycle: data collection, model training, validation, regulation, monitoring, and post-market evaluation. That does not mean using race crudely or treating it as biological destiny. Race is not a clean genetic variable, and pretending otherwise would be scientifically sloppy and socially dangerous.

But race and ethnicity can capture exposure to racism, migration history, structural disadvantage, ancestry-linked risks in some contexts, and unequal treatment within health systems. Used carefully, transparently, and with community oversight, these data can help reveal where tools are failing.

No data, no audit. No audit, no accountability. No accountability, and suddenly the robot has tenure.

The International AI Race Is Also a Health Equity Issue

The paper also raises a practical point: if France cannot collect or use race and ethnicity data responsibly, it may struggle to compete in the global AI-enabled medical device market. Countries developing AI tools increasingly need to show that their systems are fair, valid, and safe across diverse populations.

That is not just a business concern. It is a patient safety concern.

A medical AI tool approved without meaningful subgroup testing may look polished in a conference demo, but the real world is not a conference demo. Real patients come with different bodies, languages, histories, neighborhoods, diets, exposures, and care barriers. If AI is going to enter clinics, hospitals, and public health systems, it needs to work for the whole waiting room.

The authors are not calling for careless data collection. They are calling for a shift from color-blindness to equity-aware governance. That means collecting sensitive data only when there are clear protections, clear purposes, transparency, and safeguards against misuse.

This is where public health has something useful to offer. We already know that measuring inequality is not the same thing as endorsing inequality. Counting disparities is how we find them, name them, and reduce them. A thermometer does not cause a fever, though I admit it has terrible bedside charm.

What Better Policy Could Look Like

A more equity-focused AI policy would allow race and ethnicity data to be used under strict ethical rules. Patients should know why the data are collected, how they are protected, and how they will improve care. Communities affected by health inequities should have a voice in governance, not just be treated as rows in a spreadsheet.

Developers should test models across population groups before deployment. Regulators should ask for evidence of performance across race, ethnicity, sex, age, socioeconomic status, and other relevant factors. Health systems should keep monitoring AI tools after rollout because models can drift over time, especially when practice patterns or patient populations change.

The central idea is simple: if an algorithm may affect people differently, we need the data to find out.

That is especially true for AI-enabled medical devices, where errors can shape diagnosis and treatment. Fairness cannot be assumed. It has to be measured, corrected, and measured again. Equity work is not a one-time software patch. It is more like brushing your teeth: less glamorous than a moonshot, but skip it long enough and everyone notices.

The Bigger Lesson

This Viewpoint is about France, but the lesson travels well. Many countries are struggling with how to balance privacy, anti-discrimination principles, and the need to measure inequity. The uncomfortable truth is that refusing to collect race and ethnicity data does not make racial bias disappear. It may simply make bias harder to detect.

AI can be a powerful tool for better health care. It can help clinicians reach diagnoses faster, support overwhelmed systems, and potentially improve care in communities that have been underserved for generations. But that future will not happen automatically. Technology does not bend toward justice on its own. People have to build it that way, test it that way, and regulate it that way.

The hopeful part is that this is fixable. We can design AI systems that are more transparent, more accountable, and more useful for the people most often left out of medical innovation. But first, we have to stop confusing “not measuring race” with “solving racism.”

That distinction may not fit neatly on a bumper sticker, but it could make health care safer for millions of people.

This blog post discusses research findings and should not be taken as medical advice. If you have concerns about how AI tools may affect your care, please consult a qualified healthcare provider. Research discussed here represents ongoing scientific investigation and policy debate, and clinical validation is still in progress.

All images used in this post are decorative illustrations only and do not represent or reflect the accuracy, reality, or correctness of the referenced research.

Primary Source: Impracticality of banning collection of data on ethnicity and race in artificial intelligence-enabled health care in France. PubMed Record ID: 42062120. https://pubmed.ncbi.nlm.nih.gov/42062120/