The Algorithm Hasn't Met Every Kind of Skin

Personalisation is one of the most appealing promises of AI in skincare. A tool that adapts to your skin rather than offering the same generic advice to everyone sounds like progress.

But personalisation raises an important question: whose skin was represented when the system was built and tested?

Machine-learning systems learn patterns from data. If the examples used to develop or evaluate a system don't adequately represent the people who will use it, its performance may not transfer equally to everyone.

In dermatology, this is not merely a theoretical concern.

A 2024 systematic review and meta-analysis of AI skin-cancer classification research discussed the limited diversity of some datasets and the risk that underrepresentation of particular skin types could contribute to healthcare disparities. It also highlighted differences between research settings and real-world clinical use.

These findings do not mean every AI system performs poorly on darker skin. Nor do they establish that every model has the same limitations. They show why performance needs to be evaluated rather than assumed.

It is also important to separate skin tone from other characteristics. Skin tone is not the same thing as skin type, sensitivity, acne history or an individual's medical history. A system that recognises a visible feature in one image has not necessarily learned how that feature behaves across different people and conditions.

The same caution applies to commercial skincare tools.

A company may describe its product as personalised, but that word alone doesn't tell you which tasks the model has been tested for, how diverse its evaluation data was, or whether its performance is known across different skin tones and camera conditions.

Consumers deserve clearer answers.

Developers should evaluate systems on representative data, report limitations honestly and test performance beyond a single curated dataset. They should also consider privacy: a facial photograph is personal data, and people should understand why it is collected, how it is processed and how long it is retained.

For users, the practical lesson isn't to reject every AI tool. It's to ask better questions.

Does the tool explain what it can and cannot assess? Does it distinguish a visible observation from a diagnosis? Does it describe its evidence and privacy practices? Is there a sensible route to professional care when the result is uncertain?

These questions matter more than a sleek interface or a confident numerical score.

Technology can make skincare information more accessible, but accessibility isn't achieved simply by putting an upload button in front of everyone. A tool must work responsibly for the people it claims to serve.