Questions
Everything most people wonder about this toy.
Is this a paternity test?
No. Face resemblance is a noisy signal that correlates only loosely with genetic relatedness. If you have real doubts about a biological relationship, a $30 at-home DNA test gives a definitive answer. We don't.
Can this tell me if someone is biologically related?
No. Face-recognition networks learn to tell whether two photos show the same person — they aren't trained to detect kinship. Academic kinship-verification models that ARE trained for that task max out around 75–85% accuracy on ideal benchmark images, which is nowhere near sufficient for a paternity claim. DNA testing is.
How accurate is AI face recognition?
At identity — extremely accurate. Modern models hit 99%+ on benchmark datasets like Labeled Faces in the Wild. At kinship — much less accurate, because the network was trained to match the same person, not to detect family. That's why we contextualise every score against real reference distributions.
Are my photos uploaded anywhere?
No. All face detection, alignment, embedding, and scoring happens inside your browser tab using WebGPU or WebAssembly. The photos themselves never leave your device, and we store nothing across sessions.
Is there a free face comparison app that doesn't need an account?
Yes — this one. No sign-up, no email, no photo uploads. You can close the tab and every trace is gone. Ads pay the hosting bill.
Why do siblings sometimes score low?
Siblings share about 50% of their DNA on average, but face shape is controlled by hundreds of genes, each with a tiny effect. It's entirely normal for siblings to draw different combinations and look quite different — and just as normal for cousins or even strangers of the same ancestry to look alike.
Do siblings actually look alike?
On average, yes — more than unrelated pairs of similar age and ancestry. But the variance is huge. Published research on perceived sibling resemblance shows rater agreement is only moderate; AI face-recognition similarity between siblings averages higher than strangers but far below twins. One low score between your kids means nothing.
Why do unrelated people sometimes score high?
Face-recognition networks encode coarse features like face shape, skin tone, and age. Two unrelated people who happen to share those features will score moderately high even though they aren't related. The percentile-vs-strangers number in our result card is specifically there to flag when this is happening.
Does my baby look more like mom or dad?
Popular belief (and one famous 1995 study) said newborns look more like dad, but every replication since has failed to show a paternal-bias effect — and some found a maternal bias at similar ages. Your baby probably looks like a blend with a lot of random developmental luck. We have a whole page explaining the research at /the-science.
At what age do facial features become set?
Face shape changes dramatically through adolescence and continues to drift into early adulthood. Baby faces are noticeably different from what the same person will look like as an adult — which is why comparing a toddler to a parent scores lower than comparing two adults, even when the family link is strong.
Can identical twins have different AI face scores?
Identical twins will score extremely high against each other, but not always 100%. Different lighting, expressions, ages at the time of photo, or even a beard vs clean-shaven can move the score. A 0.92 cosine between identical twins is normal, not a concern.
Do adopted kids look like their adoptive parents?
There's some psychological research showing people often perceive resemblance between non-biological family members, likely shaped by shared expressions and lifestyle. Our AI scores are based on pure facial features, so adopted kids will typically score at stranger-level vs adoptive parents — which is normal and says nothing about how close the family is.
Does AI face recognition work on babies?
Sort of. The underlying networks were trained mostly on adult faces, so baby-to-adult comparisons are noisier than adult-to-adult. A low score for a baby-vs-parent pair is uninformative; a high score still means something. When you can, use a photo of the parent at a similar age.
Does ethnicity affect the score?
Yes, somewhat. Face-recognition models produce systematically higher baseline similarity between faces of similar ancestry. The percentile calibration partially corrects for this, which is why the percentile-vs-strangers matters more than the raw cosine number.
Why does the first analysis take a while?
We download about 17 MB of neural-network weights the first time you run an analysis. Your browser caches them, so subsequent runs are near-instant.
What browser do I need?
Any modern browser works (Chrome, Firefox, Safari, Edge). For best performance, use a WebGPU-capable browser — recent Chrome or Edge on desktop. Safari and older browsers will fall back to a slower WebAssembly path, but everything still works.
Why do people say kids look more like their fathers?
A 1995 Nature paper by Christenfeld and Hill reported that finding, and it spread widely in popular culture. Follow-up studies have repeatedly failed to replicate it. As of today, there is no good evidence that babies look more like their fathers. See /the-science for the citations.
Can I upload a celebrity photo to compare to someone?
Right now the site only compares faces you upload. A celebrity-lookalike mode is on the roadmap and will ship with a curated set of public-domain reference faces so we don't need to redistribute copyrighted photos.
How common is it for a child to not be biologically related to the presumed father?
Peer-reviewed population studies put it at about 1–3% in Western populations — far lower than the 10–30% number that circulates online, which comes from biased samples of men who already suspected something. Anderson's 2006 meta-analysis remains the gold standard. Citations on /the-science.
Can I use this commercially?
The site is free to use. The underlying face models are distributed under their original licenses from the InsightFace project, which permit non-commercial use.