Artificial
intelligence is now being used to monitor the people who monitor us. In
police departments across the United States, the United Kingdom, and Canada,
machine learning systems are being deployed not only to predict where crime
might occur but to track officer behaviour, flag unusual patterns in
complaint data, and generate early-warning alerts when an individual
officer’s conduct profile suggests elevated risk. The Electronic
Frontier Foundation’s 2024 review of AI and policing documented how
these systems have moved from pilot programmes to operational deployment in
major departments, often without the transparency or community oversight that
would accompany comparable changes in other areas of public administration.
The question of who polices the police now has a partial algorithmic answer.
Whether it is an adequate one is a different question
entirely.

The promise of algorithmic oversight is straightforward to
articulate: impartial, consistent, and immune to the social pressures that
have historically allowed misconduct to go unreported or unpunished within
police institutions. A system that processes every complaint, every
use-of-force report, and every stop-and-search record against a baseline
model of expected officer behaviour should, in theory, identify patterns that
supervisors motivated to protect colleagues might miss. In cities including
Los Angeles, Chicago, and London, machine learning models have been piloted
to assess officer conduct and provide early-warning signals to internal
affairs divisions before individual incidents escalate to formal complaint or
legal action. In two Arizona police departments, an AI-driven body-worn
camera review platform called Truleo was deployed through randomised
controlled trials in early 2024, assessing officer behaviour for
professionalism indicators and generating automated feedback.

The Data Problem That Technology Cannot Resolve

The practical record of these systems reveals a fundamental
limitation that the technology cannot resolve on its own. Algorithmic oversight
depends entirely on the quality and completeness of the underlying data it
processes. In most police departments, that data originates from internal
reporting systems that are already vulnerable to underreporting and
suppression. Misconduct that is not reported does not appear in the dataset.
Misconduct that is reported but classified downward by supervisors generates
a weaker signal than the underlying behaviour warrants. A system trained on
data shaped by these dynamics will reproduce them in its outputs, and in some
cases amplify them.

The NAACP’s
analysis of predictive policing tools reaches a conclusion directly
relevant to internal oversight systems: models built on historical arrest and
complaint data inherit the biases embedded in that data, including patterns
of over-policing in particular communities that inflate the apparent risk
signals in those areas. Officers who police already over-policed communities
may generate more complaint data than those whose conduct in quieter areas
has produced fewer records. The system flags according to the data it has,
not according to the conduct that actually occurred. Governing magazine’s
analysis of AI-powered predictive policing found that most American police
departments lack clear policies on algorithmic decision-making and provide
little or no public disclosure about how their predictive models are
developed, trained, or monitored for accuracy and bias.

Transparency and Accountability Gaps

The EFF’s review documented accountability gaps arising from
surveillance technology deployed without adequate local oversight, finding
that data pipelines established for one stated purpose are routinely
repurposed in ways that were not disclosed when the original systems were
approved. Body-worn camera review platforms using AI to assess officer
behaviour raise procedural questions that most departments have not resolved:
who has access to the algorithmic assessments, who can contest them, and what
procedural rights do officers have when algorithmic outputs influence
disciplinary decisions? Officers subject to adverse algorithmic assessments
may have limited ability to challenge the data or methodology on which those
assessments are based.

The legal framework governing wrongful outcomes from AI-assisted
law enforcement is still being constructed through case law. A 2025
settlement in Reid v Jefferson Parish involved a facial recognition
misidentification that resulted in wrongful arrest, producing a
200,000-dollar settlement. The Clearview AI class action resulted in a
51.75-million-dollar settlement in 2025. These cases establish that
AI-assisted law enforcement errors carry legal consequences, but they do not
yet create clear standards for what due process rights apply when algorithmic
tools shape the decisions affecting individuals. As our analysis of AI’s
hidden infrastructure and accountability gaps found, the gap
between stated commitment to responsible AI and operational governance of how
systems behave in practice is where risk accumulates, and in law enforcement
it falls most heavily on those with the least power to contest
it.

What Effective Accountability Requires

The governance requirements for AI used in law enforcement
oversight are substantially more demanding than for AI in consumer
applications, because the consequences of error fall on people with the least
power to contest them. San Jose, California has attempted to address this gap
by developing a public process requiring transparency about which AI systems
are in use, what data they rely on, and what oversight mechanisms govern
their outputs. The NAACP’s recommendations call for independent oversight
bodies, mandatory disclosure of data sources and methodologies, and community
involvement in deployment decisions. These are reasonable minimum standards,
and they remain the exception rather than the rule in how AI oversight tools
are currently being adopted.

Algorithmic accountability in policing, like other forms of police
accountability, requires governance infrastructure that is independent of the
institutions being overseen. Oversight that is internal to police
departments, or that relies on the same data systems whose integrity is in
question, cannot deliver the accountability that algorithmic tools promise in
theory. The technology is available. The governance frameworks that would
make it effective are not, and the pace of deployment is not waiting for them
to be built.

The Officer Rights Question

The deployment of AI oversight tools in law enforcement raises
procedural questions that employment law and civil service frameworks have
not yet resolved. Officers subject to adverse algorithmic assessments have
limited ability, in most current deployments, to understand the data or
methodology on which those assessments are based, to contest their accuracy,
or to require independent review before the assessment influences a
disciplinary decision. The opacity that makes algorithmic systems a
governance concern when applied to the public applies equally when those
systems are applied to the officers whose conduct they are meant to evaluate.
Due process requirements that protect individuals from arbitrary
institutional action apply in both directions.

The governance framework that would make algorithmic police
oversight both effective and legitimate requires transparency about what
systems are deployed, independent oversight of how those systems perform,
meaningful procedural rights for individuals affected by algorithmic
assessments, and community involvement in decisions about deployment and
governance. None of these requirements is technically demanding. All of them
require institutional and political commitments that have proved difficult to
sustain in the face of institutional resistance from the departments being
overseen. As our analysis of AI’s
hidden governance gaps found, the distance between where AI
decisions are made and where accountability for those decisions falls is a
structural feature of current deployments, not a temporary oversight. Closing
that distance in law enforcement, where the consequences of inadequate
governance are most severe, remains one of the most urgent challenges in AI
policy. Our analysis of AI
and unequal access to justice and services finds the same pattern
of governance frameworks lagging deployment across multiple
sectors.

About the Author

Stuart Kerr is the Technology Correspondent for LiveAIWire. He
writes about artificial intelligence, emerging technology, and the forces
reshaping work, business, and society.