Experts and machines recognized faces similarly but made different errors
Forensic experts, super-recognizers, and neural networks performed similarly on average, but each group made different mistakes.
The key idea
To trust an identification, it matters how often it is correct, but also how it errs and how confident it is.
The study, in plain language
The team ran laboratory tests with seven super-recognizers and compared them with novice controls and published data from forensic experts. In another stage, 37 super-recognizers, 16 forensic examiners, 19 laboratories, ten neural networks, and 106 novices took part in an international test of 20 image pairs. The three expert groups and networks performed better than novices, and expert groups had similar accuracy.
Differences appeared when looking at errors and confidence. Some super-recognizers made highly confident false identifications, while forensic experts more often gave inconclusive answers. Difficult comparisons for people were not the same as those for networks. When the team combined independent responses from different groups in a calculation, one measure of how well same-person pairs were separated from different-person pairs improved.
What they found
- The three expert groups performed similarly on the overall test, but made different kinds of errors.
- Experts gave more inconclusive responses; some super-recognizers made highly confident mistakes.
- Combining responses from different groups in a calculation improved separation of same-person and different-person pairs.
How they studied it
Researchers compared performance and confidence of super-recognizers, forensic experts, neural networks, and novices in laboratory tests and an international test of 20 face pairs. They then simulated what happened when independent responses from different groups were averaged.
What the study cannot show
- The forensic test had 20 comparisons and groups worked under different conditions. The combinations averaged independent responses and did not test interactive collaboration or every identification context.
Something to think about
When a tool suggests someone’s identity, ask what it does besides achieve an average accuracy: when does it hesitate, and what kinds of errors does it make? This study compared face tests, not real investigative cases.
The original study
Diverse types of expertise in facial recognition
Source under CC BY 4.0. English adaptation and explanation by PaperFlow, with AI assistance. The text has been reorganized for a general audience; it is not the complete paper and does not imply endorsement by its authors.
One study alone does not represent a field's consensus. Educational content does not replace professional advice. Report a possible error.
