Methods & limits
What ClaimCheck measures, how the numbers are produced, and what they do not mean.
The number we report
YouCam Skin Analysis gives each concern a raw score from 1 to 100, higher meaning healthier-looking skin. We use severity = 100 − raw and report a change as the share of the starting severity that is gone. "−30% wrinkles" therefore means wrinkle severity went from, say, 60 to 42 on the same crop of the same face. Every image is scored on an identical crop, so the before and after numbers are comparable. Scoring the same image twice gave identical results.
Choosing the simulation strength
YouCam Skin Simulation takes a strength between 0 and 1 for each concern. How much severity a given strength removes depends on the face, so we did a small calibration first: on two AI-generated faces we simulated one concern at a time at strengths 0.1 and 0.2 and measured the result.
| Face | Concern | Starting severity | Removed at 0.1 | Removed at 0.2 |
|---|---|---|---|---|
| A | wrinkles | 66.9 | 16.0% | 29.8% |
| A | dark spots | 26.4 | 22.5% | 32.9% |
| A | texture | 36.4 | 16.5% | 32.0% |
| B | wrinkles | 18.9 | 11.3% | 20.9% |
| B | dark spots | 14.7 | 34.2% | 47.2% |
| B | texture | 13.6 | 27.4% | 47.0% |
The response is close to a straight line below 0.2 and flattens above about 0.3 (on face A, wrinkle severity removed was 55% at 0.3, 72% at 0.6 and 76% at 1.0 when all concerns were simulated together). The average slope per concern is the starting guess: strength = claimed share ÷ slope (1.28 for wrinkles, 2.0 for dark spots, 1.98 for texture). After rendering, the image is measured again. If the measured change misses the claim by more than 4 percentage points, the strength is corrected once by proportion and the closer of the two renders is shown.
The ad-style retouch
The same concern simulated at full strength (1.0). We use it as a stand-in for the heavy retouching common in advertising imagery and measure it the same way. It is a reference point, not a claim about any specific advert.
Reading the claim
A fixed set of rules (no language model) extracts the concern, the percentage, the time frame, the study size and the share of people who saw an improvement. Percentages about people ("92% of women agreed") are never read as a size of change. Claims about pores, redness, dark circles, radiance, firmness or hydration are recognised but not rendered: in our tests the pore score did not move steadily with the simulation, and the others were not calibrated, so a number would be misleading.
What this is not
- Not a prediction. The middle image shows what the claimed size of change looks like on your face. Whether a product achieves it for you is a different question, and the study's own small print usually says that a share of people saw no change.
- Not a clinical measurement. The scores come from a photo-based model. Clinical studies use their own instruments and scales, so "−31%" on a pack and "−31%" here are related ideas, not the same measurement.
- Not medical advice.
- Photo-to-photo variation was not measured. The same image always scores the same, but two photos of the same face in different light will differ. We have not yet measured by how much.
- Small calibration. Two faces. The per-photo correction step is there because of this.
Units and privacy
A live check makes at most 4 analyses and 3 simulations (up to 48 YouCam units). The demo runs on a fixed budget of free units, limits each visitor to 2 live checks a day, and switches to recorded runs when the budget is low. The recorded runs were produced by the same code path and cost nothing to view.
Your photo is sent to YouCam for processing through our server, which holds the API key. We do not store photos or results; YouCam's own retention policy applies to files sent to its API. The sample faces are AI-generated and do not show real people. Product names on the site are fictional.
Source
Code, tests and the calibration data: see the repository linked from the project page.