Facial recognition should be banned in public spaces

What's this about?

People disagree about whether police should use face-matching tools in public places. The question is whether safety gains are worth the risks.

What supporters say

• Scanning faces can let authorities track people as they go about daily life.
• Different error rates can place more harm on some groups than others.
• Rules meant to control these tools have failed in some cases.

What critics say

• Human checks and safety controls can lower the chance of harmful mistakes.
• Careful, limited use may help find people wanted for serious crimes.

How to read this

The number of points on each side does not show who is right; stronger proof matters more.

The bottom line

The evidence gives a strong case against broad, routine face scanning in public places. It is less clear whether every narrow use should face a ban, but the critics’ points have weaker support.

Facial recognition in public spaces can make ordinary movement and association identifiable, raising questions about privacy, equality and public safety. The evidence supports a strong case against routine or indiscriminate scanning, but is less conclusive about every narrowly targeted use.

The case for

The strongest argument for a ban is that public-space scanning can turn everyday life into a system of trackable records. People may be identified while attending protests, visiting public places or associating with others, often without meaningful consent. Reviews by the European Parliament and the National Academies have raised concerns about privacy, discrimination, surveillance and due process. Research on protests has also linked facial recognition to identifying participants and potentially discouraging people from joining demonstrations or political activity, even though it does not provide a population-wide estimate of how many people stay away. 1

Accuracy problems add to those concerns. The National Institute of Standards and Technology found differences in false-positive and false-negative rates among hundreds of systems, depending on the algorithm, image quality and the person’s demographic group. Later NIST findings showed that these gaps persisted across many systems, although some newer tools performed better. A separate audit found much higher error rates for darker-skinned women than for lighter-skinned men in the commercial gender-classification systems it tested, though that task is not identical to face identification (see Figure 2). 2

In a system scanning large numbers of uninvolved people, even a modest error rate could expose more people to police stops, investigations or reputational damage. The available evidence supports that risk, but does not show how often it occurs across all real-world deployments.

There is also evidence that safeguards have failed. In *Bridges v South Wales Police*, an appeals court ruled that the force’s use of automated facial recognition was unlawful because the rules were not clear enough and protections were inadequate. The Government Accountability Office has likewise reported weaknesses in federal privacy assessments, reporting and oversight. These findings do not prove that every regulated system is unlawful, but they support a precautionary approach. 3

The case against

A total ban would remove a tool that may help investigators identify wanted people in carefully defined cases. Police evaluations in London reported alerts that required human review, with confirmed matches making up only a small share of alerts. An independent review in South Wales also found possible benefits in locating wanted people. Federal guidance describes facial recognition as a source of investigative leads, while warning that a result should not by itself establish probable cause or determine someone’s identity. 4

The evidence for this benefit is limited and specific to particular deployments. Some of the evaluations were produced or commissioned by police institutions, so they show plausible usefulness more clearly than they show a broad public-safety gain.

Opponents of a ban also argue that risks can be reduced through technical testing and strict rules. Possible safeguards include small, defined watchlists; prior authorization; limits on data retention; human confirmation; transparency and independent audits. NIST’s finding that systems vary widely in performance supports comparing and testing them rather than treating all facial recognition as equally accurate or equally dangerous. Federal guidance also says that a facial-recognition result should not be used as the sole basis for identifying someone. 5

But the record does not show that these protections reliably work in large-scale public settings. The failures identified in the South Wales case and by the Government Accountability Office suggest that formal safeguards can be incomplete or poorly enforced.

The bottom line

The evidence favors restricting routine and indiscriminate public-space identification, but only weakly supports an absolute ban. The case against widespread use has stronger backing for the existence and seriousness of privacy, protest, unequal-error and oversight risks. The case for continued use rests on credible but more deployment-specific evidence that targeted systems can produce investigative leads and that controls may reduce some dangers.

A middle position is reflected in the European Union’s AI Act, which generally prohibits real-time remote biometric identification in public places for law enforcement while allowing narrow, authorized exceptions. That approach treats the technology’s rights risks as unusually high without concluding that every use is ineffective.

The central uncertainty is whether tightly limited exceptions can be enforced well enough to preserve useful investigations without recreating pervasive tracking, unequal burdens and weak accountability. On current evidence, a presumption against routine scanning—with strict exceptions for necessity, authorization, human review, retention limits, transparency and independent audits—is better supported than either unrestricted use or a categorical ban.