The AI Made the Gap Worse

In 2020, Sjoding and colleagues published a short paper in the New England Journal of Medicine that should have changed how medicine thinks about its instruments. Pulse oximetry — the clip on the finger, in every hospital, driving oxygen and ventilation decisions — was found to miss low blood oxygen in Black patients at roughly three times the rate it missed it in white patients.

The device works by shining light through tissue and reading what comes back. Melanin absorbs light. The calibration was done on a population that mostly did not have much of it.

Nobody designed that. A device was built, validated on the people who were available, deployed universally, and for decades it quietly under-detected a life-threatening condition in one group.

Dermatology has a version of this, and it has several.

What the diagnostic accuracy actually is

Groh and colleagues, Nature Medicine, 2024. Three hundred and sixty-four images, 46 diseases, assessed by 389 board-certified dermatologists and 459 primary care physicians.

Baseline, with no assistance: dermatologists reached 38 percent top-three accuracy and 27 percent top-one. Primary care physicians, 19 percent and 14 percent.

Those are lower numbers than most people expect, and they are on a deliberately hard image set rather than on patients in a room. The relevant comparison is not the absolute figure, it is the split.

Accuracy on dark skin tones ran 4 to 5 percentage points below accuracy on light skin tones, for both dermatologists and generalists.

Then they added AI decision support, which is supposed to be the answer to exactly this. Dermatologist top-one accuracy rose from 27 to 36 percent, a 33 percent relative improvement. Primary care accuracy rose from 14 to 22 percent, a 69 percent relative improvement. Both groups got meaningfully better.

And the paper’s own caution: fair AI decision support exacerbated the gap in the diagnostic accuracy of generalists across skin tones.

Everyone improved. The distance between light and dark widened for the doctors most patients actually see first.

What the tool was trained on

A review of skin type diversity across the major dermatology image datasets puts hard numbers on why.

Fitzpatrick17k: 16,577 usable images. Fitzpatrick type VI — the deepest skin — makes up 3.97 percent.

PAD-UFES-20: 2,299 images, 1,494 with Fitzpatrick labels. Type VI: 0.07 percent.

Nought point nought seven. That is one image in roughly fourteen hundred.

The Diverse Dermatology Images set was built explicitly as a corrective and is more balanced, though it reports skin type in three aggregate bands rather than across the full six-point scale, which limits what can be analyzed.

And a methodological problem that sits over all of them: none of the three links skin-type metadata to patient ethnicity. So it is not possible to separate a model’s skin-tone bias from its ancestry bias, which means every fairness claim built on these datasets is measuring something less specific than it sounds.

And the textbooks, which are the original training set

There is a whole sub-literature of content analyses counting how many images of dark skin appear in dermatology textbooks, image atlases, curricula, journal case reports and Google image results. The papers exist — in JAAD, in the International Journal of Women’s Dermatology, in JMIR Dermatology, in Archives of Dermatological Research as recently as 2026. Several of them I could not get into, so their percentages stay out of this paragraph and sit instead in the piece linked at the foot of this article, where they are attached to their sources.

What the existence of that many independent content analyses tells you is that the field knows, has known for years, and has published on it repeatedly.

This is the root of the chain. A physician learns what a disease looks like from images. If four images in a hundred show the deepest skin — which is what Fitzpatrick17k contains — then four images in a hundred is the pattern she learns. She then trains the next cohort, writes the next atlas, and labels the dataset that trains the model.

What a patient can do about a system-level problem

Very little about the system. Two things about the appointment.

Bring your own photographs, taken in daylight, at the worst point of whatever it is. A condition that flares on a Tuesday and settles by the Friday appointment is a condition being diagnosed at the wrong moment, and this is doubly true when the visual signs are already harder to read.

And describe rather than display. How long. What it feels like — burning, itching, stinging, tightness. What sets it off. Whether it leaves a mark and how long the mark lasts. Symptoms transmit through the melanin filter. Signs do not, reliably.

That is not a substitute for a system that can see you. It is triage, performed by the patient, in the ten minutes she has been allocated.

A task force reported on one of these instruments — the definition of a Stage 1 pressure injury, which is a color — in 1995, and the definition still leads with redness. That one is here, along with the devices proposed to replace the eye and what their validation cohorts actually contain.

The first instrument on this list is also the one most people have in a drawer. Pulse oximeters read slightly high on darker skin, and part of the reason turns out to be the light source rather than the skin — swap the LEDs for laser diodes and the melanin dependence goes. And before any instrument is involved there is the library a clinician learned from: counted repeatedly, with the figures and their sources set out there.

This article is educational and is not a substitute for care from someone who can look at your skin. Anything new, changing, painful or spreading should be assessed rather than researched.