Is AI Skin Analysis Accurate? What the Evidence Says

AI skin analysis is fairly accurate for large, visible signs such as wrinkles and sagging, and much less reliable for pores, acne, pigment spots and hydration. In published studies, selfie-grading systems matched dermatologists closely on some signs (correlations of 0.75 or higher) and weakly on pigmentation (0.40), especially on darker skin. Lighting changes scores too, so trends across weeks tell you more than any single number.

That's the short answer. Below we look at what "accurate" can even mean for a cosmetic score, what the studies found, which signs apps read best and worst, how much lighting matters, and a simple way to test any app on your own face.

What does "accurate" mean for a skin score?

For cosmetic skin scores, accuracy usually means agreement with trained human graders, because there's no lab test for "how visible are my pores." Researchers ask dermatologists to grade the same photos on a standardized scale, often a published skin-aging atlas, and then compare the AI's grades with theirs. Three different questions hide inside the word:

• Agreement: does the app rank faces the way experts do? Studies report this as a correlation (r), where 1.0 is perfect.
• Error size: how far off is a single grade? Studies report mean absolute error in grading units.
• Repeatability: does the same face get the same score twice in a row? This is the one you notice day to day, and it's the least studied for consumer apps.

A tool can do well on one and badly on another. An app can rank wrinkles in the right order across a group and still jump several points between two selfies of the same person. If you want the background on how these systems read a face in the first place, see how AI skin analysis works.

What do studies say about agreement with dermatologists?

The largest published validations come from cosmetics-industry researchers testing their own selfie-grading system, and they show good agreement on structural signs with weaker results elsewhere. Independent work is thinner. Here are the main findings:

The US study is the clearest picture of the pattern. Its AI matched 50 dermatologists strongly on five facial signs, moderately on cheek pores and weakly on pigmentation, and the authors say skin tone still needs work. They also found the dermatologists disagreed among themselves by 0.3 to 0.5 grading units, which is a useful reminder that the human "answer key" is fuzzy too.

In the Chinese study, people took selfies on their own phones and three dermatologists graded them on the spot. The AI's grades were highly correlated with the clinical grades but lower in most cases, with mean absolute errors that rarely went above 0.6 grading units.

Keep the source in mind: both studies come from teams that include L'Oréal researchers testing their own algorithm. A Korean group working on a separate model found that, on DSLR photos, it got the exact grade right 51.4% of the time and landed within one grade 93.3% of the time. Phone and tablet photos held up well, though pigmentation accuracy dropped.

Which skin signs do apps read best and worst?

Apps read large, high-contrast structures best and small, scattered or color-based features worst. Across the studies above, the rough order looks like this:

1. Most reliable: wrinkles, sagging and deep folds. These create shadows and shapes that show up in almost any photo.
2. Middle: under-eye circles, redness, shine and pores. Results depend more on light and image sharpness.
3. Least reliable: acne counts, pore size and pigment spots. A 2026 benchmark of 3,203 selfies found that standard deep learning models underperformed a trivial constant-mean predictor on these three.
4. Estimated, not measured: hydration and elasticity. A photo doesn't record water content, and the Korean study found only moderate correlations between image-based predictions and device readings for both.

So when an app says your wrinkles score improved, that's more trustworthy than a small change in your pore or spot score.

How much does lighting change a skin analysis result?

Lighting changes results enough to swamp small real changes. In a small lighting study described by dermatologist Zoe Diana Draelos in Dermatology Times, 60 women aged 25 to 45 were run through facial appearance apps under two kinds of light, and they reproducibly looked about three years younger under incandescent light than under white fluorescent or LED light, because white light sharpens edges and shadows.

Clinic cameras solve this with a booth, a headrest and fixed flash. Even then, small repeat studies of one widely used clinic system found that repeat captures varied by about 3.4% for absolute wrinkle scores and 8% to 11% for percentile scores, and its skin-age estimate ranged from 7 years under to 9 years over real age for individuals. A phone in changing light won't do better than that. Our guide to the best lighting for skin photos shows how to keep your setup steady.

Does AI skin analysis work on darker skin tones?

It works less well on darker skin for color-based signs, and the evidence base is thinner. In the US selfie study, pigmentation grading was weakest for the darkest skin tones. The lighting test above excluded Fitzpatrick types V and VI because the app makers said their apps were inaccurate for deeply pigmented skin.

The gap shows up in medical AI too. When researchers built a curated image set with diverse skin tones, state-of-the-art dermatology models performed substantially worse on dark skin tones and uncommon diseases. Dermatologists also did worse on those images, and fine-tuning on diverse images closed the gap. If you have deeper skin, weight wrinkle and texture trends over spot and tone scores, and read our piece on AI skin analysis on darker skin tones.

Are skin scan apps real or fake?

The technology is real computer vision, and the better systems agree with dermatologists on several signs. What they produce is a cosmetic estimate. They aren't measuring hydration with a probe or diagnosing anything, and you shouldn't use them for moles.

The evidence on spot-checking apps shows why. A BMJ systematic review of algorithm-based apps that assess skin cancer risk found small, poor-quality studies; one app had 0% sensitivity for melanoma in a 15-lesion study, and another reached 80% sensitivity and 78% specificity in two studies. The reviewers concluded that such apps can't be relied on to detect all skin cancers. A cosmetic selfie scanner is a different kind of product and makes no such claim, but the lesson carries over: for anything medical, see a doctor. Our comparison of AI skin analysis vs a dermatologist covers where each one fits.

How can you test a skin analysis app yourself?

You can check any app's repeatability in about ten minutes. Here's a simple protocol:

1. Three in a row. Take three scans back to back, same spot, same light, bare face. The spread between them is the app's noise floor. Treat any later change smaller than that spread as noise.
2. Change the light. Scan once facing a window and once under an overhead bulb. A large jump tells you lighting matters a lot for this app, so always scan in the same place.
3. Add makeup or SPF. Scan again with tinted sunscreen or foundation. If your texture or tone score improves, the app is partly scoring the product.
4. Change your expression. Smile and scan. Expression lines deepen when you smile, so wrinkle scores usually rise. If nothing moves at all, the app may not be looking closely.
5. Compare with your own eyes. Put two photos from a month apart side by side. If you can see clearer skin and the score dropped, trust the photos.
6. Track for six to eight weeks. Most routines need weeks to show results, so look at the trend line.

How can a selfie scan help, then?

A selfie scan is most useful as a consistent record you'd never keep by hand. Pori takes a selfie and in about ten seconds returns an overall skin score from 0 to 100, with Hydration & Moisture, Pores & Texture, Pigmentation & Tone, Anti-aging & Elasticity and Overall health, plus a skin-age estimate. It maps your face with a 468-point mesh and reads it region by region, and its progress charts and before/after comparison make the trend easier to see than a single score. Run the three-in-a-row test above on it too.

Any app score is a second opinion on appearance, not a diagnosis. Pori doesn't diagnose, treat or check moles, so see a dermatologist for anything painful, spreading or changing.