By
Stuart Kerr, Technology Correspondent,
LiveAIWire
Greater knowledge about AI is correlated with decreased trust in police facial recognition technology. That finding, from a peer-reviewed study published in Computers in Human Behavior in January 2026, challenges the assumption that public scepticism toward AI in law enforcement stems from ignorance that better communication will overcome. [FLAG: the original draft cited “3,500 participants across ten countries” — the actual study is a scenario-based survey of 507 participants, with no ten-country breakdown mentioned in the abstract. Corrected below to the verified figure.] The study, a scenario-based survey of 507 participants, found that the strongest predictor of acceptance of AI-driven facial recognition is trust in law enforcement institutions generally, not familiarity with the technology.
People who understand how facial recognition works, and who also understand its error rates across demographic groups, tend to be less comfortable with it, not more. The implication for law enforcement agencies pushing for AI adoption is significant: the path to public acceptance runs through institutional accountability, not through technology literacy campaigns.
The evidence base for the technology’s actual performance reinforces this concern. A 2018 study found commercial facial recognition systems showed error rates of 0.8 percent for light-skinned men and 34.7 percent for darker-skinned women, a 40-fold disparity. A 2019 NIST study of 189 facial recognition algorithms from 99 developers found that African American and Asian faces were 10 to 100 times more likely to be misidentified than white male faces.
At least eight Americans have been wrongfully arrested following facial recognition misidentifications, with police in documented cases treating software suggestions as definitive identifications rather than investigative leads. These are not isolated incidents attributable to outdated technology. They reflect the operational practice of treating probabilistic AI output as sufficient basis for high-consequence decisions about specific individuals.
What AI Is Actually Being Used For in Law Enforcement
The law enforcement AI landscape in 2026 is broader and more varied than the facial recognition debate captures. Predictive policing algorithms use historical crime data, geographic patterns, and demographic information to forecast where crime is likely to occur and allocate patrol resources accordingly. Body-worn camera footage analysis applies AI to automatically flag use-of-force incidents, transcribe interactions, and generate incident reports, reducing paperwork burden while creating searchable records.
License plate recognition systems log vehicle movements continuously in deployed areas. Social media monitoring tools scrape public and, in some jurisdictions, non-public posts for threat indicators. DNA analysis tools use AI to identify partial or mixed samples. Each of these applications has legitimate use cases and specific accountability concerns that the current governance infrastructure addresses inconsistently at best.
The EU AI Act categorises real-time biometric surveillance in public spaces as a prohibited practice with limited law enforcement exceptions requiring judicial authorisation. This creates a meaningful legal distinction between using facial recognition to identify a specific suspect after a crime, which is more permissible, and running continuous facial recognition on everyone in a public space to flag individuals matching a watchlist, which is much more restricted.
In practice, the line between retrospective identification and real-time surveillance is technically fluid in ways that the legal distinction does not fully capture, and enforcement will require case-by-case scrutiny rather than simple category assignment.
The Accountability Gap
The fundamental accountability challenge with AI in law enforcement is the automation bias documented across AI applications: officers who receive an AI-generated risk score, identification, or recommendation face significant psychological pressure to act on it rather than applying the additional scrutiny that the AI’s error rate warrants. When eight people are wrongfully arrested after facial recognition misidentifications, each individual officer likely believed they were acting on reliable information.
The accountability question is not whether those officers acted in bad faith but whether the systems, institutional practices, and evidentiary standards they were operating within were adequate for a technology with a 34.7 percent error rate on some demographic groups.
The NIST AI Risk Management Framework‘s approach to high-risk AI systems requires explicit documentation of error rates, bias testing across demographic groups, human oversight mechanisms, and defined escalation paths for cases where human judgment should override AI output. These requirements are more onerous than current law enforcement practice in most jurisdictions, but they are the standard that the technology’s actual performance record justifies demanding.
The connection between the digital resistance movement responding to AI surveillance and the law enforcement deployment context is direct: the tools being built to challenge surveillance in public spaces are precisely the accountability infrastructure that the performance data suggests is needed. And the bias mitigation challenges documented across AI systems apply with the highest stakes in law enforcement contexts, where false positives are not customer service errors but potentially wrongful arrests of people who bear no resemblance to the suspect other than demographic category.
Understanding the compliance architecture that high-risk AI systems require but rarely have, the same governance frameworks now being tested against open-source AI models, is directly relevant here. Is AI the future of law enforcement? The technology will be deployed regardless. Whether that deployment is accompanied by the accountability infrastructure its error rates require is the question that actually matters, and the current answer is inconsistently at best.
What Responsible Deployment Actually Requires
The evidence base does not support a conclusion that AI should be excluded from law enforcement. It supports a conclusion that current deployment practices are inadequate given what the technology can and cannot do reliably. Facial recognition as an investigative lead, clearly labelled as probabilistic output with known error rates, reviewed by human investigators before any enforcement action, with mandatory documentation of its use in any case where it was consulted, is a meaningfully different deployment model from the documented cases where it was used as definitive identification by officers who did not understand its limitations.
Predictive policing tools that forecast high-crime areas can improve resource allocation if used as patrol guidance, explicitly acknowledged as statistical prediction rather than individual assessment, and audited regularly for whether they are reducing crime or merely redistributing policing attention onto communities already overpoliced.
The same tools used to generate individual risk scores for people who have not been charged with anything, used in bail, sentencing, or parole decisions without adequate transparency about their error rates or demographic performance, are doing something categorically different with the same underlying technology. The accountability infrastructure determines which of these deployment models occurs, not the technology itself.
Several jurisdictions are moving toward the evidence-based accountability framework the research supports. Portland, Oregon, banned government use of facial recognition in 2020. New York City requires algorithmic impact assessments for automated systems affecting residents. The EU AI Act classifies AI in criminal justice as high-risk, requiring extensive documentation and human oversight for every deployment.
The challenge in the United States is the patchwork of municipal and state rules that produces deployment practices ranging from best-in-class to essentially unregulated within the same country. For the communities most affected by the technology’s error rates, which the demographic disparities in facial recognition data make clear are not randomly distributed, the variation in local governance is the variation in how consequential those errors are.
The question of whether AI can be deployed responsibly in law enforcement is ultimately an institutional and governance question rather than a technical one. The technology can be used in ways that meet a defensible accountability standard. The current evidence is that it frequently is not. Changing that requires explicit standards, mandatory disclosure of error rates and demographic performance, meaningful human oversight before enforcement action is taken on AI output, and accountability mechanisms when AI errors contribute to wrongful outcomes. None of those requirements exceed what the technology’s own performance record makes necessary.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.