What Bias Means Here
In face recognition, bias means error rates that differ consistently between groups of people. Not random mistakes, but patterns: a system that wrongly matches Black women more often than white men, or fails to recognise older people more often than younger ones.
It can show up as:
- More false matches for some groups (two different people judged the same).
- More missed matches (the same person not recognised).
- Worse ranked results in searches, where the right person appears lower down the list.
For the measured numbers, see accuracy across demographics.
Real Cases
Bias stops being abstract when a face search leads to an arrest.
- Robert Williams, Detroit, 2020. Arrested in front of his family for a shoplifting case after a face recognition search matched his old licence photo to blurry security footage. He was held for about 30 hours. Detroit later settled his lawsuit and agreed to new rules, including that face recognition alone can't justify an arrest.
- Nijeer Parks, New Jersey, 2019. Spent about ten days in jail over a crime in a town he said he had never visited.
- Porcha Woodruff, Detroit, 2023. Arrested while eight months pregnant for a robbery and carjacking she had nothing to do with.
In these cases, the technology was only part of the failure. Police treated a search result as evidence instead of a lead, and didn't do the checks that would have cleared these people quickly. But the people wrongly flagged were overwhelmingly Black, which matches the patterns researchers had warned about.
Read more: Face Recognition Accuracy Across Demographics: What the Research Shows
Where Bias Comes From
| Source | How it causes bias |
|---|---|
| Training data | Groups that appear less often get less practice, so the model separates their faces less precisely |
| Labelling | Mistakes in identity labels can be more common for under-represented groups |
| Photo quality | Camera exposure and lighting historically suited lighter skin, leaving some faces darker and less detailed |
| Thresholds | One cut-off for everyone produces more false matches for groups whose scores run higher |
| Testing | Benchmarks that under-represent some groups hide the problem |
| Deployment | Low-quality probe images and over-trust in results turn small biases into real harms |
Get new guides by email
How Bias Is Reduced
On the technical side
- Better data. Larger and more balanced datasets, and sampling that gives every group comparable training time.
- Fairness-aware training. Loss functions and penalties that discourage large gaps between groups.
- Per-group testing. Measuring error rates separately for every group before release, as NIST now does routinely.
- Threshold calibration. Setting cut-offs so false match rates are similar across groups.
On the human side
- Treating a match as a lead, never as proof.
- Requiring trained reviewers and independent evidence before any action.
- Banning or restricting high-risk uses. Several US cities have banned police use of face recognition, and the EU's AI Act restricts real-time remote face recognition in public spaces.
- Publishing accuracy results and audits.
No technique removes bias completely, and some trade a bit of overall accuracy for more even results. That trade-off is a decision about values as much as engineering.
Read more: Privacy and Face Recognition: What You Should Know
Where Ollie Stands
Ollie is a lookalike game, not an identification tool. It never tells you who a stranger is, celebrity results come from a fixed set of public figures, and nothing is stored about the faces you upload. That design keeps the stakes low.
The same technical biases can still make results better for some people than others. Ollie's model was trained on a research dataset that isn't perfectly balanced, and the celebrity index adds a separate list of stars from Asia, Latin America, Africa and the Middle East so more people have close matches. Scores come from the network alone, with no skin-tone filters or manual weighting. For how the system works end to end, see how Ollie works.