A repeatable number, an accurate facial measurement and a prediction of human ratings are different results. A tool needs evidence for each task it claims to perform. Landmark geometry alone is not a validated attractiveness score, and an appearance model can use information beyond geometric ratios.
RealSmile's published evaluation reports r ≈ 0.80 on a 1,566-face research benchmark and r ≈ 0.40 (95% CI 0.12–0.62) on 46 user phone photos, a median of 81 votes each. These are correlations, not percentages of correct scores. Our earlier 17-metric geometry composite failed to predict benchmark human ratings (r ≈ −0.33); the current Face Score is a separate appearance-based model. Read the samples, methods and limits, or see the accuracy evidence and provider-method comparison.
On desktop, the measurements run in your browser. On mobile, your photo is processed in memory by our scan server and deleted immediately after the landmarks are measured — never written to disk. Before scoring, the photo is also sent once to our server for an automated safety check, which asks outside AI services (OpenAI and Anthropic) for a verdict on it; we don't keep it. By default nothing is kept: no photo, no copy, nothing used to train anything. There are three exceptions and all three are your choice — you can save a scan photo to your own account, you can submit a photo to the community rating pool, and the paid Audit keeps a thumbnail so you can reopen your report. Each one is deletable, and deleting removes the file itself, not just the link to it.
1. What "accuracy" actually means for an AI face score
First ask what was measured. A landmark detector estimates feature positions in a 2D photo; a formula converts those positions into a distance, angle or ratio. Check the definition, units and agreement with appropriate reference measurements. A tidy calculation can still use the wrong points.
Then ask what was predicted. An appearance model trained on human-rated photos can be evaluated against ratings on independent photos. RealSmile's Face Score uses that approach; it is not a weighted sum of the geometric panel. A geometry measurement does not explain the score unless that relationship has been demonstrated.
Keep consistency and calibration separate. Repeating the same input checks repeatability. Agreement with human means checks a different target, and correlation alone does not show that the score scale or a percentile is calibrated. Ask which population the reference group represents.
2. What the perception literature actually supports
Little, Jones and DeBruine (2011) review facial traits studied in attractiveness judgments, including symmetry, averageness, sexually dimorphic shape and skin, as well as differences in preferences. That is context for research, not validation of an app's particular detector, score weights or universal numerical target.
Carré and McCormick (2008) studied FWHR and aggressive behavior in specific samples. That paper does not validate an attractiveness score or a personality judgment about an individual. Sharing a ratio name with a paper also does not prove that a consumer detector measures it with the same landmarks or accuracy.
Willis and Todorov (2006) studied judgments after brief face exposure. Rapid agreement in an impression is not proof that the inferred trait is true, nor an accuracy benchmark for a face-rating product. Product-specific claims require product-specific evaluation.
3. The five-minute verification protocol any user can run
These checks can reveal sensitivity in a tool. They cannot establish anatomical ground truth, human-rating accuracy or a clinically meaningful change. There is no universal 1–2-point or 3–5% pass/fail threshold across different metrics and scales.
Check 1: repeat the identical file. Keep the model, settings and photo bytes fixed. Record any change. Differences may involve cropping, resizing, detection, version changes or stochastic processing; a change alone does not prove undisclosed randomness.
Check 2: try a horizontal flip. A formula may be mirror-invariant on perfectly mirrored points, while a detector or crop is not. Left and right labels may also swap. Record sensitivity rather than assigning an invented numerical tolerance or declaring the tool validated.
Check 3: compare separately captured photos. Match pose, distance, expression and lighting as closely as practical. Differences can reflect capture and processing. They do not isolate a change in anatomy, and they do not by themselves prove longitudinal reliability.
For actual accuracy evidence, ask for independent reference measurements or held-out human-rated photos, relevant sample sizes, uncertainty and subgroup results. Passing a few personal checks does not replace that evaluation.
Try the free Face Score
Get a Face Score and your strongest eligible measurement. The optional one-time report includes the measurements your photo supports and a four-week photo plan; PDF length varies. Pro membership is separate and optional.
Try the free face scan →4. Five over-claims to ignore
1. Training size proves accuracy. Training scale does not establish generalization. Ask about the independent evaluation population and whether contributors or duplicate photos overlap with training.
2. Correlation means percent correct. A correlation coefficient is not a correctness percentage or a measure of absolute calibration. Samples, confidence intervals and absolute errors matter.
3. One number captures your attractiveness. A prediction applies to the photo, task and reference population evaluated. It is not a universal judgment or a prediction of one person's preferences.
4. Every AI tool reads only geometry. Some use appearance models, some use ratios, some collect human feedback and some combine methods. Provider documentation and evaluation are needed to establish which claims are supported.
5. AI must be better than a friend. One opinion, a crowd mean and a model prediction are different measurements. Compare each against the intended task; neither the word AI nor a crowd automatically makes a result representative.
5. Where the honest accuracy ceiling sits
There is no single validated numerical ceiling for every detector or rating model. Our own benchmark and phone-photo correlations differ substantially, and the phone sample is small. That is a reason to report both samples and their limits rather than promise precision for every new selfie.
Geometry can help describe what a photo shows, but successful point detection and a repeatable formula do not prove an accurate estimate of anatomy. Appearance-based scores need independent human-rating evaluation. Neither result validates the other.
| Claim | Evidence needed |
|---|---|
| Same-file repeatability | Evidence neededRepeated runs with the input, model and settings fixed; no universal percentage tolerance. |
| Landmark or ratio accuracy | Evidence neededAgreement with appropriate independent point or measurement references, including pose and capture limits. |
| FWHR matches a research measure | Evidence neededMatching anatomical definitions, detected landmarks and a measurement-error evaluation. |
| Prediction of one person’s preference | Evidence neededA group-average rating evaluation does not establish this. |
| Prediction of average human ratings | Evidence neededIndependent human-rated photos, sample details, uncertainty and absolute errors as well as correlation. |
| Scores agree across providers | Evidence neededComparable task, scale and reference population; agreement alone is not validation. |
| Feature panels agree across providers | Evidence neededComparable definitions, detected points and preprocessing; no assumed small tolerance. |
6. How to actually use an AI face attractiveness score
Use the result as feedback on a photo, with the reference population and confidence limits in view. A feature panel can suggest that a capture should be retaken, but it does not prove a grooming change will raise your score, make you more attractive or improve dating outcomes. A change across two photos may be capture variation.
The free scan returns a Face Score and the strongest eligible measurement. The optional one-time face report includes the measurements your photo supports, their limits and a four-week photo plan; PDF length varies. Inspect the sample before deciding. Pro membership is separate and optional.
For direct crowd feedback, read how Human Score works. Ratings describe that photo and the participating crowd. To compare candidate profile photos, the free dating photo ranker orders 2 to 6 of them by Face Score and expression. Neither tool promises to validate someone's worth or guarantee sales, matches or attraction.
For provider differences, see the Photofeeler comparison and QOVES comparison. Pay for a useful, clearly described deliverable rather than assuming that payment proves greater measurement accuracy.
Frequently asked questions
How accurate are AI face attractiveness tests?
Accuracy depends on the task and evaluation sample. Locating landmarks, repeating a measurement and predicting human ratings are separate questions. RealSmile reports correlation r ≈ 0.80 on a 1,566-face research benchmark and r ≈ 0.40 (95% CI 0.12–0.62) on 46 user phone photos, with a median of 81 votes each. Neither figure means that percentage of scores is correct, and the small phone sample does not establish performance for every group.
What does "accuracy" actually mean for an AI face score?
Measurement validity asks whether the detected points and calculated feature match an appropriate reference. Prediction validity asks whether a separate score agrees with independent human ratings. Repeatability asks whether the same input gives the same result. A result can be repeatable while anatomically wrong or poorly calibrated to human ratings.
Are AI face attractiveness tests scientifically valid?
Research about facial perception does not validate a particular app. Studies may support an association in a specified population while a consumer formula still fails. Our earlier 17-metric geometry composite had benchmark correlation r ≈ −0.33 and was retired as an attractiveness score. The current Face Score is a separate appearance-based predictor; its measured results and limits are published.
Why do two AI face tests give me different scores?
Tools can differ in detected points, image processing, metric definitions, training labels, models, reference populations and numerical scales. A percentile is relative to its named reference group. Matching scales or agreeing with another tool does not establish anatomical accuracy, agreement with people or calibration.
How can I verify an AI face attractiveness test in five minutes?
In five minutes you can begin consistency checks, not prove accuracy. Repeat the identical file under the same model/settings; then try a horizontal flip and a separately captured photo. Record what changes. Differences can arise from detection and processing rather than randomness. There is no universal 1–2-point or 3–5% pass threshold, and passing these checks does not replace independent reference measurements or human-rated evaluation.
Do AI face tests work the same on every face?
That must be measured rather than assumed. Pose, occlusion, lighting, age and the populations represented in training and evaluation can affect results. Overall accuracy does not establish equal performance in every subgroup. Ask for the sample, uncertainty and failure behavior, and treat an unsupported or low-quality capture as a reason to retake rather than a confident verdict.
Can an AI test predict if other people will find me attractive?
A model may correlate with average ratings in its evaluated sample, but that does not predict one person’s attraction to you or guarantee a dating outcome. Photo context and rater preferences matter. Direct human feedback also describes the particular photo and crowd that rated it, not a universal opinion.
How should I read my AI face attractiveness score?
Read the Face Score as a prediction for this photo relative to the stated reference population. Read geometry as separate measurements with their own confidence and pose limits. A feature panel is not automatically an explanation of the appearance model. Small changes across photos are not proof that your underlying appearance improved or declined.
What separates an honest AI face test from a guess?
Look for a named task and model, documented processing and units, independent evaluation data, sample sizes, uncertainty, relevant populations, clear failure handling and honest photo-privacy choices. A repeatable demo or a large landmark count is useful information but not accuracy validation. Pay only when the actual deliverable is useful to you; payment alone does not establish better accuracy.
Try the free Face Score
Get a Face Score and your strongest eligible measurement. The optional one-time report includes the measurements your photo supports and a four-week photo plan; PDF length varies. Pro membership is separate and optional.
Try the free face scan →