AI Diagnoses Eczema From Photos. Poorly On Dark Skin. The Training Data Is Embarrassing.
A scoping review found that AI tools for atopic dermatitis mostly rely on convolutional neural networks analyzing clinical or dermoscopic images. Diagnostic applications dominated the literature, with predictive models also common. But only 27% of studies reported skin color, fewer than half included basic demographics like age or sex, and two teledermatology studies entirely excluded Fitzpatrick types V and VI.
This illustrates what I would call the Representation Cascade. A model trained on incomplete demographics doesn't just underperform on unrepresented groups. It actively encodes their exclusion as a statistical pattern. When darker skin tones are absent from training data, the model learns that lighter skin is the default condition for disease. That isn't a bug. It is a structural assumption baked into the math.
The authors of the scoping review, published via Hospital Healthcare, evaluated AI-based diagnostic, monitoring, and management tools for atopic dermatitis. They found reduced diagnostic performance in darker skin tones and inadequate demographic reporting across the field.
- Open Google Scholar at scholar.google.com and search for 'AI dermatology Fitzpatrick skin types.' Notice how many results mention skin tone classification. The number is smaller than you'd expect.
- Visit an open-source skin condition tool such as the one at modelderm.com and upload a photo of a skin area. Read what the tool says about its training data or limitations.
- Search 'Fitzpatrick scale chart' on any search engine and review the six skin type categories. Ask yourself which types you have seen represented in AI health marketing imagery. The gap is the lesson here.