safety
Humans and Face AI Share a Biased Sense of Similarity
A comparison with 4,000 participants found accuracy shifted with the viewer's race, the face's race, and individual recognition skill.
Summary
A comparison with 4,000 participants found accuracy shifted with the viewer's race, the face's race, and individual recognition skill.
University of Notre Dame researchers compared commercial and open-source face-recognition systems with judgments from 4,000 people. Humans and algorithms often agreed about which faces looked similar, but accuracy depended heavily on the race of the participant, the race of the face being viewed, and the person's natural face-recognition ability. The work does not make proprietary systems transparent or establish fairness in deployment. It shows that benchmark averages can conceal structured differences in how both people and machines make identity-adjacent comparisons.
Why it matters
A comparison with 4,000 participants found accuracy shifted with the viewer's race, the face's race, and individual recognition skill.
Limits and context
- The work does not make proprietary systems transparent or establish fairness in deployment.
Key claims
A comparison with 4,000 participants found accuracy shifted with the viewer's race, the face's race, and individual recognition skill.
Qualification: The work does not make proprietary systems transparent or establish fairness in deployment.
Evidence: source-2026-07-16-011
Sources
- University of Notre Dame via Newswise: AI and human face recognition accuracyUniversity of Notre Dame via Newswise · official announcement
Corrections
No corrections have been recorded for this story.