How accurate is AI facial recognition technology for law enforcement use

Variable Accuracy, With Documented Demographic Disparities

AI facial recognition accuracy varies considerably depending on the specific system and conditions involved, and importantly, independent testing has documented that accuracy for many systems has historically differed meaningfully across demographic groups, a well-documented pattern central to ongoing controversy about the technology’s use in law enforcement.

What Independent Federal Testing Has Found

The National Institute of Standards and Technology has conducted extensive, independent testing of numerous facial recognition algorithms from different vendors, and this testing has documented that many, though not all, tested systems showed meaningfully lower accuracy for certain demographic groups compared to others, providing rigorous, independently verified evidence of this disparity rather than relying only on anecdotal reports.

Which Groups Have Been Documented as More Affected by Lower Accuracy

Across a number of tested systems, documented accuracy disparities have shown lower performance particularly for women and for people with darker skin tones compared to accuracy rates for other demographic groups, a pattern that has been independently replicated across multiple studies and testing efforts rather than being an isolated finding.

Why These Disparities Likely Occur

While specific technical causes can vary by system, these disparities have generally been attributed at least partly to the composition of the data used to train these algorithms, with some historical training datasets containing less representative coverage of certain demographic groups, potentially contributing to less accurate performance for those groups.

Why This Matters So Much in a Law Enforcement Context

In a law enforcement context specifically, accuracy disparities carry serious real-world consequences, since a facial recognition system that performs less accurately for certain demographic groups could disproportionately contribute to wrongful identification or wrongful arrest for individuals within those groups, a concern that has moved from theoretical to documented reality in specific, publicly reported cases.

Documented Cases of Wrongful Identification

There have been publicly documented cases where individuals were wrongfully identified, and in some cases wrongfully arrested, based partly on facial recognition matches that were later determined to be incorrect, providing concrete, real-world examples underlying broader concerns about relying on this technology, particularly for systems with documented accuracy disparities, without additional robust verification steps.

Bottom Line

AI facial recognition accuracy varies considerably by specific system, and independent federal testing has documented meaningful accuracy disparities across demographic groups for many systems, with women and people with darker skin tones historically experiencing lower accuracy in a number of tested systems — a well-documented pattern that has contributed to real, publicly reported cases of wrongful identification in law enforcement use.

Go deeper

• What Is Facial Recognition AI and How Widely Is It Used?
• What Real-World Harms Have Resulted From Biased AI Systems?
• How accurate are AI fraud detection systems in insurance?