Inside
prisons in the United States, the United Kingdom, and dozens of other
countries, AI systems are being used to predict which prisoners are most
likely to commit violence, to monitor inmate communications for security
threats, and to generate risk scores that inform parole and sentencing
decisions. These are not experimental pilots. They are operational
deployments affecting hundreds of thousands of people who are, by definition,
among the most legally vulnerable and institutionally controlled members of
society. The accountability for how these systems work, and what they get
wrong, is far weaker than the power they exercise.

The appeal to prison administrators is straightforward: prisons
are dangerous environments with significant resource constraints, and AI
promises to make them safer and more efficiently managed. Predictive systems
that identify violence risk could allow staff to intervene before incidents
occur. Communication monitoring systems could detect the planning of escapes
or criminal coordination. Risk assessment tools could help parole boards make
more consistent, evidence-based release decisions. Each of these applications
has genuine potential merit. Each also carries the potential for serious harm
to individuals when the systems are inaccurate, biased, or applied without
adequate human oversight.

Risk Assessment and the Parole Question

The use of algorithmic risk assessment in parole and sentencing
has been the subject of the most sustained public scrutiny, following
investigative journalism and academic research exposing the limitations and
biases of widely deployed tools including COMPAS (Correctional Offender Management
Profiling for Alternative Sanctions). A 2016 investigation by ProPublica
found that COMPAS assigned higher risk scores to Black defendants than to
white defendants with similar criminal histories, raising fundamental
questions about whether actuarial risk assessment tools reproduce and amplify
historical patterns of racial inequality in the criminal justice
system.

The developers of COMPAS and similar tools dispute specific bias
claims while acknowledging that their systems reflect historical data. This
defence raises a deeper problem: if historical criminal justice data reflects
racially discriminatory policing, charging, and sentencing practices, then a
machine learning system trained on that data will learn to replicate those
discriminatory patterns regardless of whether race is included as an explicit
variable. Proxy variables, including neighbourhood, employment history, and
family criminal justice involvement, can substitute for race in ways that
produce equivalent discriminatory outcomes while appearing
race-neutral.

Several US states have mandated the use of algorithmic risk
assessment in parole decisions; others have moved to restrict it following
legal challenges. The American
Civil Liberties Union has litigated multiple cases challenging the
use of opaque algorithmic tools in criminal sentencing, arguing that
defendants have a constitutional right to challenge evidence used against
them, including algorithmic risk scores whose methodology is often shielded
as proprietary.

Surveillance Inside Prison Walls

Within prisons, AI surveillance systems go considerably beyond
risk scoring. Facial recognition cameras track prisoner movement continuously
in facilities across the US and UK. Voice analysis systems monitor inmate
phone calls for keywords and sentiment indicators associated with security
threats. Natural language processing tools scan written communications for
prohibited content. Biometric systems identify individuals through gait,
fingerprint, and iris recognition at internal checkpoints.

The justification for this level of surveillance is institutional
security. The concern raised by prison reform advocates and human rights
organisations is that comprehensive AI surveillance of a captive population,
with no possibility of exit or meaningful consent, creates conditions of
total informational control that go beyond what is necessary or proportionate
for safety purposes. The HM
Inspectorate of Prisons has noted in recent reports that the
expansion of surveillance technology in UK prisons has outpaced the
development of governance frameworks and independent oversight
mechanisms.

Rehabilitation, Discrimination, and Second
Chances

Beyond security, AI is being deployed in prison programmes
designed to support rehabilitation and reduce reoffending. Personalised
education programmes, mental health screening tools, and employment matching
systems are all being used in prison settings with genuine rehabilitative
intent. These applications face a different set of concerns from security
surveillance, but concerns nonetheless. If AI systems used in rehabilitation
programmes reproduce the same demographic biases documented in risk
assessment tools, they may systematically assign lower-quality educational
opportunities or employment prospects to prisoners from already disadvantaged
backgrounds, compounding inequalities rather than reducing
them.

The post-release implications of prison AI data are also
underexplored. Risk scores generated during incarceration can persist in
criminal justice databases and influence employment screening, housing applications,
and future interactions with the criminal justice system. People who have
served their sentences and are attempting to rebuild their lives may find
that algorithmic records from their incarceration follow them in ways that
were not intended or disclosed when the data was collected.

What This Means for You

Most people will not personally experience the AI systems deployed
in prisons. But the criminal justice system is a central institution of
democratic society, and the standards of fairness, accountability, and human
dignity it maintains are a measure of the values that society holds. AI
systems that make consequential decisions about liberty, rehabilitation, and
release without adequate transparency or accountability represent a failure
of those values, regardless of whether the individuals affected are
sympathetic figures.

The broader principle at stake is that algorithmic decision-making
in high-stakes contexts requires higher standards of accountability, not
lower ones, than human decision-making. The current practice in many
jurisdictions, where AI tools are deployed in criminal justice contexts with
less scrutiny than would be applied to a new pharmaceutical treatment,
inverts this principle in ways that should concern anyone who cares about the
rule of law. The international dimension of AI in criminal justice is also
significant. Countries with weaker rule-of-law frameworks and less
independent judiciary oversight are deploying AI criminal justice tools with
even less accountability than in the UK and US contexts discussed here. The
risk that AI criminal justice systems provide a veneer of objectivity and
scientific authority to what are essentially politically motivated detention
and risk assessment decisions is not hypothetical. Several authoritarian
governments have explicitly framed AI criminal justice tools as modern,
evidence-based alternatives to traditional justice processes that are subject
to independent review. but enforcement against governments deploying AI
criminal justice tools in authoritarian contexts is essentially absent. The
development of minimum international standards for AI in criminal justice,
including mandatory transparency requirements, independent accuracy auditing,
and enforceable rights to explanation and challenge for individuals subject
to algorithmic criminal justice decisions, is urgent. The UN Special
Rapporteur on Torture has specifically noted that AI-enabled predictive
incarceration and risk-based detention, where people are held based on
algorithmic assessment of future risk rather than past acts, raises serious
concerns under international law that have not been adequately addressed by
existing human rights frameworks but is enforced only weakly in practice, and
the development of international standards for AI in criminal justice is a
priority that the UN Human Rights Council has identified but not yet
adequately resourced.

The accountability mechanisms available in different jurisdictions
vary considerably. In England and Wales, the Biometrics and Surveillance
Camera Commissioner provides independent oversight of police use of biometric
and surveillance technologies, with powers to publish recommendations and
require responses from police forces. This mechanism is more robust than most
equivalent oversight bodies internationally, but its resources are limited
relative to the scale of AI deployment in policing it is asked to monitor.
Strengthening and properly resourcing independent oversight is a precondition
for public trust in AI-assisted policing, and that trust is itself a
prerequisite for the police legitimacy that effective community policing
depends upon.

For related analysis of AI in law enforcement and surveillance,
see our coverage of AI
in policing and AI
in state control systems.

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.