The Biggest Study: NIST 2019
In December 2019, the US National Institute of Standards and Technology published NISTIR 8280, the largest study of demographic differences in face recognition to date. It tested 189 algorithms from 99 developers on over 18 million photos of more than 8 million people, from US mugshots, visa applications and border crossings.
Its main findings:
- False matches by race. For many algorithms, false positives (wrongly matching two different people) were 10 to 100 times more common for West African, East African and East Asian faces than for Eastern European faces.
- US mugshots. The highest false positive rates were for American Indian faces, with elevated rates for African American and Asian faces.
- Sex. False positives were higher for women than men across most algorithms.
- Age. Higher error rates for children and older adults than for middle-aged adults.
- Where algorithms came from mattered. Several algorithms developed in China did not show the elevated false positives on East Asian faces that many Western algorithms did.
- The best were much better. The most accurate algorithms showed far smaller differences.
Two Kinds of Error
It helps to separate the two ways a system can fail:
| Error | What happens | Real-world harm |
|---|---|---|
| False match (false positive) | Two different people are judged to be the same | Wrongful police suspicion, someone else unlocking your phone |
| False non-match (false negative) | The same person isn't recognised | Being stopped at a gate, failing to verify an ID |
NIST found that demographic gaps were usually larger for false matches than for false non-matches, and false matches are the error most likely to harm someone in a police search.
Read more: Bias in Face Recognition: What It Is and What to Do About It
Gender Shades: A Related Warning
A year earlier, in 2018, Joy Buolamwini and Timnit Gebru published "Gender Shades". They tested commercial systems from IBM, Microsoft and Face++ on a different task: guessing whether a face was male or female. Error rates were as low as 0.8% for lighter-skinned men and as high as 34.7% for darker-skinned women.
That study was about gender classification, not identity matching, but it drew wide attention to the problem, and the companies involved improved their systems within months of the publication.
Get new guides by email
Why the Gaps Exist
- Training data. Most large face datasets were scraped from the web and over-represent some groups. A model gets more practice telling apart the faces it sees most.
- Photo conditions. Camera exposure settings and many lighting setups were historically tuned for lighter skin, which can leave darker faces underexposed and lower in detail.
- Thresholds. Systems usually use one similarity threshold for everyone. If scores for some groups naturally run higher, that group sees more false matches at the same threshold.
- Test data. Popular benchmarks like LFW are skewed, so gaps can go unnoticed until someone tests specifically.
The human version of the first point is the own-race effect: people are also better with the faces they see most.
Read more: Face Recognition Benchmarks Explained: LFW, IJB-C, and Beyond
What This Means for Ollie
A celebrity lookalike search has lower stakes than a police search: nobody is identified, and a less-than-perfect match costs a laugh, not a liberty. But the same biases can make results worse for some people, so it's worth being open about.
- The model. Ollie's network was trained on about 5.8 million photos of 85,742 people. Like most large face datasets, it is not perfectly balanced.
- The index. Ollie's celebrity list adds a separate ranking of stars from East and South Asia, Southeast Asia, Latin America, Africa and the Middle East, using page views on their own languages' Wikipedias, so more people have close matches to find. About three in five people in the index are men.
- No hidden adjustments. Scores come from the network alone, with no skin-tone filters or manual weighting.
If your results seem consistently off, try a couple of photos in soft, even light, which helps every face. If that doesn't help, the Ollie team would like to hear about it through the contact page.