Metropolitan
Police in the UK used live facial recognition cameras at the 2024 Notting
Hill Carnival, scanning faces against a watchlist and flagging matches for
officers to assess. The technology generated arrests that would not otherwise
have occurred. It also generated false matches that subjected innocent people
to police attention. Civil liberties organisations and some elected
representatives called it an unacceptable surveillance overreach. The
Metropolitan Police argued it was a legitimate and proportionate use of an
effective crime-fighting tool. Both claims contain truth, and the tension
between them captures the central challenge of AI in law enforcement: the
same capability that makes policing more effective also makes society more
surveilled, and the trade-offs are borne unequally.

AI applications in law enforcement range from the relatively
uncontroversial to the deeply contested. Administrative automation that
reduces officer paperwork, predictive maintenance for police fleet vehicles,
and AI transcription of interviews are broadly welcomed. Facial recognition
in public spaces, predictive policing algorithms that direct patrol resources
based on crime forecasts, and automated licence plate readers that create
movement records for millions of vehicles are much more contested. The debate
about each reflects different values about the relationship between security
and liberty, and the line between legitimate policing and political
surveillance.

Facial Recognition: The Most Contested
Technology

Live facial recognition is the AI policing tool that has attracted
the most controversy, because it operates on everyone who passes through its
field of view, not just suspects or people who consent to screening. The
accuracy concerns are well-documented: multiple studies, including audits
commissioned by police forces themselves, have found higher false match rates
for women and people from ethnic minority backgrounds. At Notting Hill
Carnival, where the attendee demographic differs significantly from the
demographic profile of most facial recognition training datasets, the
accuracy disparities are particularly acute.

The legal framework for facial recognition in UK policing is
contested. The Court of Appeal found in 2020 that South Wales Police’s use of
facial recognition lacked sufficient legal basis, was not proportionate, and
failed to comply with data protection law. The government subsequently
introduced the Data Protection and Digital Information Act with provisions
intended to provide a clearer legal basis, but critics including the Liberty human rights
organisation argue the provisions are insufficiently rigorous and
insufficiently transparent about when and how facial recognition can be
used.

In the United States, a patchwork of city and state-level bans and
regulations reflects deep political disagreement about facial recognition in
policing. San Francisco, Boston, and several other cities have banned police
use of facial recognition entirely; others have expanded its deployment. Federal
legislation has been proposed but not enacted, leaving a fragmented
regulatory landscape that varies dramatically depending on where you
live.

Predictive Policing and Algorithmic Bias

Predictive policing algorithms, which use historical crime data to
forecast where crimes are likely to occur and direct patrol resources
accordingly, have been deployed and subsequently abandoned by several UK and
US police forces following evidence of racial bias. The fundamental problem
is the same as in algorithmic risk assessment for prisons: systems trained on
historical policing data learn to replicate historical policing patterns. If
certain communities have historically been over-policed relative to their
actual offending rates, predictive algorithms trained on arrest data will
direct more patrol resources to those communities, generating more arrests,
which become training data for the next generation of the algorithm, creating
a self-reinforcing cycle of disproportionate policing.

The College of
Policing in the UK has developed ethical guidance on AI use in
policing that explicitly addresses algorithmic bias, requiring forces to
conduct equality impact assessments before deploying predictive tools and to
monitor outcomes for disproportionality. Implementation of this guidance
varies considerably across the 43 forces in England and Wales.

Evidence, Accountability, and Police Legitimacy

AI is also transforming how police handle evidence, with
implications for both investigative effectiveness and accountability. AI analysis
of CCTV footage, bodycam recordings, and digital communications can process
volumes of evidence that human analysts could not review in the available
time. This is a genuine investigative benefit that has supported serious
crime investigations. It is also a capability that, if misused or
inadequately governed, could compromise the fairness of criminal
proceedings.

The risk of over-reliance on AI evidence is real. Juries and
judges who encounter AI-generated analysis of evidence may give it
unwarranted authority relative to its actual reliability. Defence lawyers who
lack the technical expertise to challenge AI evidence effectively may be at a
structural disadvantage. The principle that defendants have a right to
understand and challenge the evidence against them is genuinely threatened
when that evidence is generated by opaque algorithmic
systems.

What This Means for You

If you live in a city with an active police facial recognition
programme, your face has almost certainly been scanned without your knowledge
or consent. You have limited legal recourse in most UK jurisdictions to
prevent this. What you do have is a democratic voice: the decisions about
whether and how facial recognition is used by police are ultimately made by
elected officials and police and crime commissioners who are accountable to
the public. The evidence on accuracy, disproportionality, and legal
compliance is increasingly available and increasingly clear. Multiple
independent evaluations of facial recognition policing in the UK have found
accuracy disparities that, in the view of the Information Commissioner,
require material improvement before deployment can be considered
proportionate. The Metropolitan Police has published its own accuracy data,
which shows improvement over time but still falls short of the standards that
a human identification procedure used in evidence would be required to meet.
Engaging with this evidence, through the work of organisations including
Amnesty International, Liberty, and Big Brother Watch, as well as through
academic research published by groups at the Alan Turing Institute and the
Oxford Internet Institute. The
Alan Turing Institute’s policy work on AI in policing provides an
authoritative and balanced analysis that is accessible to non-specialists and
directly relevant to the policy questions that police and crime commissioners
and elected officials are being asked to decide, and engaging with it is part
of informed democratic participation. The forensic science dimension of AI in
policing raises specific due process concerns beyond surveillance. AI tools that
analyse DNA, fingerprints, digital device content, and voice recordings for
evidence are being used in criminal investigations with varying degrees of
validation and transparency. The Forensic Science Regulator in England and
Wales has published guidance on AI validation requirements for forensic
applications that is more rigorous than the guidance applying to operational
policing AI. The contrast between the standards applied to forensic AI, where
the scientific validity of methods can be challenged in court, and
operational policing AI such as facial recognition and predictive tools,
where the basis for decisions is often opaque and unchallengeable, reflects a
governance inconsistency that defence lawyers and civil liberties
organisations have consistently identified as problematic.

For related coverage of AI surveillance and justice, see our
analysis of AI
in prisons and AI
in border surveillance.

 The cumulative weight of this
evidence, from independent academics, civil society organisations, and
statutory bodies, makes a compelling case that the current governance
framework for AI in policing is inadequate and requires strengthening through
primary legislation that sets clear standards, requires transparency,
mandates independent audit, and creates meaningful redress for individuals
harmed by AI policing errors.

About the Author

Stuart Kerr is a technology correspondent at LiveAIWire, covering
artificial intelligence, digital innovation, and the social impact of
emerging technologies. Follow LiveAIWire for daily analysis at liveaiwire.com.