Perfect Corp publishes a 95% test-retest figure and Haut.AI an ICC of 0.97 to 0.98: both say a scan repeats itself. Which?'s 30-person test asked whether it matches your skin, and about 37% of results did not.
SlipperSkin finding
In Which?'s test published 15 October 2026, the Boots skin scanner repeated itself within a point in the same light, yet missed or contradicted acne for 10 of 14 testers; of 6 numbered maker claims SlipperSkin read on 23 October 2026, 3 measure repeatability or correlation and none is given per skin tone.
Consistent, yes; accurate about your skin, often not. The short version:
Mostly whether the scan agrees with itself. A repeatability figure (test-retest, or an intraclass correlation, ICC) says the same face in the same conditions gets the same score twice. That is worth having, and it is what makes a scanner useful for tracking change. It says nothing about whether the score is right. Haut.AI, whose engine runs Neutrogena Skin360, publishes ICC values of 0.97 to 0.98 across clinic cameras and at-home selfies on its clinical software page, a strong repeatability result honestly labelled as one. Perfect Corp's YouCam skin analysis page carries a 95% test-retest reliability rate next to an 80% correlation with dermatologist evaluations, and in one paragraph presents the 95% as proof of accuracy. The accuracy question is the one on the right of this figure.
Correlation and agreement figures are closer to accuracy but still need reading. An 80% correlation says scores rise and fall with a dermatologist's grades across a group; it does not say how often one person's score is right. Perfect Corp's case study with Dr Feldman and Dr Taylor reports 69% clinician agreement across 14 skin conditions against 56% for a $50,000 clinical imaging system, which is a fair and useful comparison, though the page gives no sample size or skin-tone breakdown. L'Oreal Paris says Skin Genius is "up to 95% accurate compared with a live dermatologist consultation", without the method. Revieve, the white-label engine behind at least seven of the 17 tools Which? found, describes "dermatologist-grade precision" on its FAQ and gives no figure.
| Maker or tool | Published claim | What kind of number it is | Figure per skin tone? | Recommends | Your selfie |
|---|---|---|---|---|---|
| Haut.AI (engine behind Neutrogena Skin360) | "98% diagnostic accuracy across diverse skin types" | Accuracy, reference not stated | No | Client brand's products | Not stated on the pages read |
| Haut.AI clinical software | ICC 0.97 to 0.98 across capture settings | Repeatability | No | n/a (research tool) | Not stated on the pages read |
| Perfect Corp (YouCam web tool) | "95% test-retest reliability rate" | Repeatability, described as accuracy | No | Suggestions; no own line | API: generally 30 days |
| Perfect Corp (YouCam web tool) | "80% correlation with dermatologist evaluations" | Correlation | No | Suggestions; no own line | API: generally 30 days |
| Perfect Corp case study | 69% clinician agreement vs 56% for a $50,000 system | Agreement | No | n/a | n/a |
| L'Oreal Paris Skin Genius | "Up to 95% accurate compared with a live dermatologist consultation" | Agreement, method not given | No | Own: L'Oreal Paris | Deleted within seconds; results up to 3 years |
| Revieve (white-label engine) | "dermatologist-grade precision" | No number | No | Client retailer's catalogue | Deleted immediately; no model training |
| La Roche-Posay Spotscan+ | Clinical study of 1012 acne patients | Study cited, no figure on page | No | Own: up to 4 La Roche-Posay | Deleted after analysis |
Of the 8 claims above, 6 carry a number and none publishes a figure broken down by skin tone. Several say their training data covers every Fitzpatrick type, which is a statement about the data, not a measured result.
That the scan agreed with testers' own concerns about 63% of the time. Which?'s "Can you trust AI skin analysers?", by James Aitchison, put 30 people through the Boots Skincare Diagnostic Tool in early September 2026: over 120 scans, four each, two back to back in good light, one with deliberately wrong questionnaire answers and one in dim or patchy light. Boots ran it on the Revieve engine. The misses cluster in exactly the local, visible concerns people scan for.
Boots did not answer Which?'s questions on how products were chosen, and told it the tool was no longer available online; Which? notes several high-profile tools have recently been pulled from retailer sites. Dr Emma Wedgeworth, the consultant dermatologist Which? consulted, notes that radiance and skin typing "are not recognised in clinical dermatology".
Far more than a repeat scan does. For about half of Which?'s testers, moving the light swung the overall score as high as +39 or as low as -25 points, and light from the left rather than the right changed results. In the same light, the average change was under one point. Drawn on one scale, the photo conditions dwarf everything else.
The evidence says not reliably, and no maker we read publishes a per-skin-tone accuracy figure. In Which?'s test the tool struggled to detect hyperpigmentation and dark spots on testers with Fitzpatrick type IV and V skin, and gave them higher average health scores than lighter-skinned testers: a miss that looks like a compliment. The best-documented measurement of the underlying problem comes from medical imaging: Daneshjou and colleagues' 2022 Science Advances study scored three dermatology models on 656 pathology-confirmed images balanced across skin tones.
Because the questionnaire drives them more than the photo. 25 of Which?'s 30 testers got a completely different morning product list just by changing their pre-scan answers, while the overall score barely moved. A tester with rosacea was offered products containing retinol and salicylic acid. Across the group, L'Oreal was recommended 214 times and Cosrx 154, 1.4 to 1, while Olay and La Roche-Posay barely appeared. Brand tools are more straightforward about this: Skin Genius chooses from more than 30 L'Oreal Paris products, and La Roche-Posay Spotscan+ recommends up to four of its own and tells anyone at acne grade 3 or above to see a dermatologist. Ingredient apps have the same problem one step later; see whether ingredient apps agree on the products a scanner recommends.
1. Same light every time: face a window, nothing behind you, no overhead spot. Lighting moved Which?'s scores by up to 64 points. 2. Scan twice. If the two scores differ by 5 or more, the photo, not your skin, is moving the number. 3. Check the concern you know you have (acne, dark spots, redness). If the scan misses it, treat the rest of the report with the same caution. 4. Answer the questionnaire honestly: the product list follows your answers more than your photo. 5. A skin condition, a changing mole or anything sore goes to a dermatologist, not an app.
We re-read every maker page linked here on 23 October 2026 and quote only what each page says now; a claim we could not re-read today was left out. We classified each published number by what it compares: a scan with itself (repeatability), scores with a dermatologist's across a group (correlation), or a verdict with a clinician's (agreement). Independent figures come from Which?'s test, read from the article itself, and the peer-reviewed Daneshjou 2022 study. SlipperSkin did not scan anyone; the counts are Which?'s, and every derived figure on this page is computed from them. Nothing here is medical advice, and no consumer skin scanner is a diagnostic device.
As an Amazon Associate I earn from qualifying purchases.
Consistent more than accurate. In Which?'s test published 15 October 2026, repeat scans in the same light moved under a point on average, but the Boots tool missed or contradicted acne for 10 of 14 testers and flagged redness for 0 of 7; about 63% of results matched the tester's skin.
That the tool gives the same face nearly the same result when it is scanned again. It is a repeatability figure. It does not say whether the result agrees with a dermatologist, which is the accuracy question. Perfect Corp publishes it alongside an 80% correlation with dermatologist evaluations.
L'Oreal Paris says Skin Genius is "up to 95% accurate compared with a live dermatologist consultation", read on its page 23 October 2026. The page does not give the study, sample or method, so the figure cannot be checked from the page.
Not reliably, on the evidence available. Which? found the Boots tool missed hyperpigmentation on Fitzpatrick IV and V testers and gave them higher scores. In medical imaging, Daneshjou 2022 found three dermatology models scored lower on Fitzpatrick V to VI than I to II. No maker we read publishes accuracy by skin tone.
Mostly light and angle. For about half of Which?'s testers, changing the light swung scores by up to +39 or -25 points; in the same light the average change was under one point. Scan facing the same window at the same time of day.
No. These are cosmetic tools, and Which? found calling them diagnostic sets unrealistic expectations. Its consultant dermatologist advises anyone with a skin condition to seek professional help. La Roche-Posay's acne tool itself tells users at grade 3 or above to see a dermatologist.
AI skin scanners are good at repeating themselves and much less good at seeing what you see. The makers' headline numbers are mostly about the first. Use a scanner to track your own skin in steady light, read its product list as a catalogue, and take anything that worries you to a dermatologist.