How Accurate Is AI Skin Analysis on Dark Skin?

Why skin tone affects accuracy

AI reads patterns it has seen before. If a model trained mostly on lighter skin, it learned less about how concerns appear on deeper tones. Pigmentation, redness and post inflammatory marks look different across the Fitzpatrick range, so a model that never learned those differences reads them poorly.

This is not a small edge case. Deeper skin tones make up most of the world's population, and concerns like hyperpigmentation are among the most common reasons people buy skincare.

The data problem

Historically, many public skin image datasets over represented lighter skin. Models built on them inherited that gap and performed worse on Fitzpatrick types IV, V and VI. The result was weaker reads, missed concerns and recommendations that did not fit.

Fixing this is not a matter of a filter or a tweak. It requires training and testing on genuinely diverse skin, then measuring the result honestly.

How accuracy is measured

A credible tool reports accuracy per skin tone, not as a single blended number. Ask for performance broken down across the full Fitzpatrick range. A high average can still hide poor results on deeper tones if lighter tones dominate the test set.

Good measurement also covers real conditions: varied lighting, everyday phone cameras and a spread of concerns. Lab numbers on perfect photos do not prove real world accuracy.

What good looks like

Strong performance on deeper skin shows up as:

• Consistent reads of pigmentation and marks across skin tones.
• Recommendations that fit real concerns rather than defaulting to generic advice.
• Transparency about how the model was tested and where it is uncertain.

Oyster is built around this. Its engine is measured across the full Fitzpatrick range and weighted toward deeper skin tones, an approach it calls the melanin moat.

Why this is a commercial issue, not only a fairness one

Accuracy on deeper skin is often framed as a fairness question, and it is one. It is also a plain business issue. If a large share of a retailer's customers have deeper skin, a tool that reads them poorly gives them weak recommendations, so they buy less and trust the brand less.

Getting this right lifts results across the whole customer base. A shopper whose pigmentation and concerns are read correctly gets a recommendation that fits, buys with confidence and returns. The case for accuracy on every skin tone is both ethical and commercial, and the two point the same way.

Questions to ask a vendor

Before trusting any skin AI, ask:

• How does accuracy break down across each Fitzpatrick type?
• What skin tones were in your training and test data?
• How does the system handle a low quality photo?
• Can you show results on the customers we actually serve?

Clear answers signal a tool you can rely on for every shopper.