The Dark-Skin Images Are the Infections

Somebody counted the pictures in Nelson’s Textbook of Pediatrics, which is the book American pediatricians grow up on, and then did something more interesting than counting: they looked at which conditions got which skin.

Burns, leprosy and urticaria pigmentosa were mostly shown on dark skin. Psoriasis, acne, hemangioma, molluscum, herpes and keloid were mostly shown on light skin.

Keloid. In a book that will also tell you, in the text, that keloids are commoner in people with darker skin.

The authors’ own line about this is dry and devastating: the clustering was unrelated to epidemiology.

First the plain counts, because they are worse than people assume

The reference audit is from 2018 and it looked at four books assigned at top American medical schools — Gray’s Anatomy for Students, Clinically Oriented Anatomy, the Atlas of Human Anatomy, and Bates’ Guide to Physical Examination. Four thousand one hundred and forty-six images.

By race, the books were close to the country: 62.5 percent white, 20.4 percent Black, 17.0 percent other people of color. That is roughly the United States.

By skin tone, the same images were 74.5 percent light, 21 percent medium, and 4.5 percent dark.

Those two sentences describe the same pictures. A book can hit its racial quota and still show you almost nothing but pale skin, because the people cast as Black in a line drawing are frequently drawn light, and because a photograph of a face is coded by race while a photograph of an abdomen is coded by tone. The gap between those two numbers is the whole problem in one dataset.

Since then the counting has spread out of dermatology, and the answers have been consistent to a degree that stops being interesting and starts being structural.

The clustering is the part nobody quotes

Percentages get repeated. The pattern inside them does not, and it is the finding with teeth.

Nelson’s put burns and leprosy on dark skin. A 2023 audit of one American medical school’s entire preclinical curriculum — 1,660 images, every course, not just dermatology — found that images of infectious disease were significantly more likely to be presented in darker skin. Chi-square 15.46, p below .001. Two different institutions, two different corpora, the same shape.

So a student’s visual training does not merely under-represent dark skin. It sorts it. Dark skin arrives attached to contagion and catastrophe, and the everyday inflammatory conditions that fill a real clinic — the psoriasis, the acne, the eczema — arrive pale.

That same curriculum audit contains a detail I keep thinking about. The course with the highest proportion of light skin was pulmonary medicine, at 88 percent. The course with the lowest was dermatology, at 59 percent. Dermatology is the specialty that has been publicly working on this since at least 2006, when a review of core educational resources concluded that coverage of dark skin was limited and variable. It is still the best of a bad set, and it is only just past half.

What it does to a student, which is not quite what you would guess

Here the evidence gets more honest than the argument usually is.

A Toronto group audited their own undergraduate dermatology curriculum, found fewer than 7 percent of disease images were in skin of color, and then tested their students on five common lesions shown in both light skin and skin of color.

Accuracy did not collapse. First-year students scored 77.8 percent on white skin and 85.9 percent on skin of color. Third-years scored 71.3 and 72.4. If you were expecting the underexposed images to produce a large accuracy gap, they did not, at least not on a multiple-choice test with clean photographs.

What did move was confidence. Students rated themselves significantly less sure of their diagnoses in skin of color, in both years, at p equals .0002.

I think that is a more useful finding than a big accuracy gap would have been, and I would not paper over it to make the point land harder. A doctor who is right but unsure behaves differently from a doctor who is right and certain. They hedge, they wait, they say let us watch it for a few weeks — and every delay in this field is measured in the same currency, which is how far something has got by the time somebody acts on it.

It also matters that the test was a photograph. A clean, well-lit, correctly exposed photograph, chosen by a dermatologist to be diagnostic. That is not what walks into a clinic, and it is not what a phone camera hands a doctor on a video call.

The machines learned it too

There is a version of this problem that did not exist five years ago and is now being generated at scale.

A 2025 study took four leading image generators — Adobe Firefly, ChatGPT-4o, Midjourney, Stable Diffusion — gave each the same flat prompt for the twenty commonest skin conditions, and looked at four thousand outputs.

Across all of them, 89.8 percent of the generated images showed light skin. Firefly was the outlier in the right direction, at 38.1 percent dark skin, which is roughly census-accurate. The other three came in at 6.0, 3.9 and 8.7 percent.

Then the second result, which is the one that should stop anybody planning to fill a curriculum gap this way. Two blinded dermatology residents looked at a random subset and could identify the intended condition in 15 percent of the images.

Fifteen percent. So the fastest available route to more diverse teaching images produces pictures that are mostly the wrong color and almost always the wrong disease, and it produces them faster than anyone can audit them. This is the same failure mode as the diagnostic algorithms, arriving from the other end: not a model that cannot see dark skin, but a model that will cheerfully invent it.

The fix is boring and it works

The interventions that have been tested are small, unglamorous and effective.

Medical students in Botswana worked through online perceptual-learning modules built from images of dark skin. Accuracy and fluency improved with large effect sizes, 1.5 and 3.7, at p below .0001, and the gains were still there when they were retested three to five months later. A Canadian self-learning module took a group of second-years from 51 percent correct to 87. A one-hour self-paced module at a New Jersey medical school improved both description and diagnosis.

None of these is a landmark trial. They are self-selected students, single institutions, small numbers. But the direction is consistent and the cost is close to zero, which puts the shortage of images where it belongs: not a hard scientific problem, a procurement decision.

What this changes in an appointment

Not much, directly. You cannot fix a textbook from an exam room.

What it changes is how much weight to put on the sentence that doesn’t look like — psoriasis, rosacea, eczema, whatever has been suggested and dismissed. That sentence is a comparison against a mental library, and the library was assembled from images that were 75 to 95 percent pale, sorted so that the dark-skinned examples skewed infectious.

It is entirely reasonable to ask what the condition looks like on skin your color, and to ask it as a real question rather than a challenge. It is reasonable to ask whether the differential changes if the redness is doing something other than looking red — which is the failure the severity scores have already been caught making.

And take your own photographs, in daylight, from the same angle, a week apart. A sequence you brought is worth more than one image somebody half-remembers from a book.

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.