What AI face recognition can — and can't — tell you about family

1. Face-recognition networks aren't kinship detectors

The neural network that powers this site (an ArcFace-family model[2] descending from FaceNet[1]) was trained to solve one problem: is this the same person as in that other photo? It achieves ~99%+ accuracy on benchmark datasets like Labeled Faces in the Wild[16] because re-identifying a person under different lighting and poses is a well-defined task.

Kinship verification is a different problem, and it's much harder. The Families in the Wild benchmark[5] evaluates models specifically trained to predict family relationships from photos — best published accuracies sit in the 75–85% range, not the 99% of face identification[6]. The model on this site was not trained on that task. It scores resemblance, which correlates with kinship loosely. Loosely is doing a lot of work in that sentence.

2. Face shape is polygenic — hundreds of genes, not one

A 2018 genome-wide study led by Claes and colleagues[12] mapped more than 200 independent genetic loci influencing global and local facial shape in humans. Features like jaw width, nasal bridge height, and interocular distance each draw from many genes, each contributing a tiny effect. Because the inheritance is so distributed, two siblings can easily end up drawing different combinations and look quite different — while unrelated people from similar ancestral backgrounds can share many of the same variants and look alike.

This is why the percentile vs. strangers number on our result cards matters more than the raw cosine similarity. It asks a more honest question: "how unusual is this resemblance, given the baseline noise of human faces?"

3. The "babies look like dad" claim is mostly wrong

A famous 1995 Nature paper by Christenfeld and Hill[7] reported that 1-year-olds resembled their fathers more than their mothers, and the finding spread widely in popular culture. Follow-up work failed to replicate it. Brédart & French (1999)[8], Bressan & Grassi (2004)[9], and Alvergne et al. (2007)[10] all found either no paternal-bias effect or evidence for a maternal-side resemblance bias at similar ages. Meta-analyses since then show the "mini-Dad" hypothesis does not survive replication.

Practical implication: a "Mini-Mom" or "Mini-Dad" result from this site reflects what an image model perceives in these particular photos, nothing about deep biology. Take the screenshot. Don't take it further than that.

4. A note on the statistic you saw on Reddit

There's a viral statistic online claiming that 10–30% of fathers are unknowingly raising another man's child. It's wrong. That number comes from paternity- testing labs, which are selection-biased to men who already have reason to suspect non-paternity — not a representative population.

Peer-reviewed population estimates are much lower:

• Anderson (2006)[13] — worldwide meta-analysis finding a median rate of ~1.9% across representative samples.
• Voracek, Haubner, & Fisher (2008)[14] — cross-temporal meta-analysis reporting median rates consistent with the 1–3% range.
• Larmuseau et al. (2013)[15] — Y-chromosome genealogical study in a Western European population found per-generation non-paternity of roughly 1%.

If you have genuine doubts about a biological relationship, an at-home DNA test costs about $30 and answers the question properly. A face-matching web app never will.

5. What this app's scores actually mean

• Cosine similarity: the dot product of two 512-dimensional unit vectors output by a MobileFaceNet recognition model[3]. It tells you how "alike" the model thinks two faces look, between roughly −0.1 (random) and ~1 (same person).
• Percentile vs. strangers: where your pair's cosine sits relative to cosines we've seen across many unrelated-stranger pairs. Higher than ~85th means "the AI sees real resemblance"; below ~40th means "stranger-level, no signal."
• Percentile vs. siblings / parent-child: comparisons against the cosine distributions we observe for real sibling and parent-child pairs. Calibrated on the Families in the Wild dataset[5]. These are the numbers the caption tiers are built on.
• Face-shape similarity: a second, independent score computed from geometric ratios of 478 landmarks (face width, eye spacing, jaw angle, etc.). Helps spot cases where the neural network is picking up on age or expression more than identity.

6. Known limitations

• Age gap. Babies and young children don't look like the adult versions of themselves. Matching a baby to an adult parent is much noisier than matching two adults. If possible, use a photo of the parent at a similar age.
• Ethnicity baselines. Face-recognition networks produce systematically higher baseline similarity between faces of similar ancestry. The percentile calibration partially corrects for this, but you should expect a bit of inflation when comparing same-ancestry unrelated pairs.
• Photo quality. Small, dark, blurry, or heavily filtered photos lose signal. Sunglasses, heavy makeup, and extreme expressions all lower the score without changing the underlying relationship.
• Single-photo noise. One photo is one observation. A different photo of the same person on the same day can produce a noticeably different score.

Further reading

Three popular-science books worth reading if this page left you hungry. Amazon affiliate links — see our disclosure.

References

1. [1] Schroff, F., Kalenichenko, D., & Philbin, J. (2015). FaceNet: A Unified Embedding for Face Recognition and Clustering. CVPR 2015.
2. [2] Deng, J., Guo, J., Xue, N., & Zafeiriou, S. (2019). ArcFace: Additive Angular Margin Loss for Deep Face Recognition. CVPR 2019.
3. [3] Chen, S., Liu, Y., Gao, X., & Han, Z. (2018). MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices. CCBR 2018.
4. [4] Guo, J., Deng, J., Lattas, A., & Zafeiriou, S. (2021). Sample and Computation Redistribution for Efficient Face Detection. arXiv:2105.04714.
5. [5] Robinson, J. P., Shao, M., Wu, Y., & Fu, Y. (2016). Families in the Wild (FIW): Large-Scale Kinship Image Database and Benchmarks. ACM Multimedia 2016.
6. [6] Qin, X., Liu, D., & Wang, D. (2019). A Literature Survey on Kinship Verification through Facial Images. Neurocomputing.
7. [7] Christenfeld, N. J. S., & Hill, E. A. (1995). Whose baby are you?. Nature 378, 669.
8. [8] Brédart, S., & French, R. M. (1999). Do babies resemble their fathers more than their mothers? A failure to replicate Christenfeld & Hill (1995). Evolution and Human Behavior 20(2), 129–135.
9. [9] Bressan, P., & Grassi, M. (2004). Parental resemblance in 1-year-olds and the Gaussian curve. Evolution and Human Behavior 25(3), 133–141.
10. [10] Alvergne, A., Faurie, C., & Raymond, M. (2007). Differential facial resemblances of young children to their parents: who do children look like more?. Evolution and Human Behavior 28(2), 135–144.
11. [11] Kaminski, G., Méary, D., Mermillod, M., & Gentaz, E. (2012). Is it a he or a she? Behavioral and computational approaches to sex categorization. Attention, Perception, & Psychophysics.
12. [12] Claes, P., Roosenboom, J., White, J. D., Swigut, T., et al. (2018). Genome-wide mapping of global-to-local genetic effects on human facial shape. Nature Genetics 50, 414–423.
13. [13] Anderson, K. G. (2006). How well does paternity confidence match actual paternity? Evidence from worldwide nonpaternity rates. Current Anthropology 47(3), 513–520.
14. [14] Voracek, M., Haubner, T., & Fisher, M. L. (2008). Recent decline in nonpaternity rates: a cross-temporal meta-analysis. Psychological Reports 103(3), 799–811.
15. [15] Larmuseau, M. H. D., Vanoverbeke, J., Van Geystelen, A., et al. (2013). Low historical rates of cuckoldry in a Western European human population traced by Y-chromosome and genealogical data. Proc. Royal Society B 280: 20132400.
16. [16] Huang, G. B., Ramesh, M., Berg, T., & Learned-Miller, E. (2007). Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments. U. Mass. Amherst Tech Report.