Custody and prisons

Algorithmic risk scores shape sentencing, parole and how people are managed in prison and on probation in England and Wales. The main tool, OASys, has run for two decades with little independent scrutiny of how it works.

Guide

What it is

  • OASys the Offender Assessment System, used by prisons and probation since 2005 to produce risk scores from historical factors and interviews.
  • Digital Categorisation Service suggests a prisoner's security category, which staff review and can change.
  • ARNS Assess Risks, Needs and Strengths, the in-house successor to OASys, now being rolled out nationally.

How it is used

More than 1,300 OASys assessments are completed every day, according to figures Statewatch obtained, and the database held over seven million risk scores at the start of 2026. Scores feed into pre-sentence reports, parole decisions and supervision. The record tracks these tools, their replacement, and what inspectors and Parliament say about them.

Key tensions

A black box at scale

Campaigners and academics have long raised concerns about bias in OASys's risk factors, its accuracy and how hard it is to challenge a score. No independent evaluation of its accuracy has been published.

Human review in substance

Tools that suggest a decision for staff to confirm are only as good as that review. Whether staff genuinely reassess, rather than accept, is what determines real oversight.

Inspecting the process, not the tool

Evidence to Parliament in April 2026 said inspectors check whether a process was completed rather than how the algorithm works, leaving scrutiny largely to the prison and probation service itself.

View all custody and prisons records Latest coverage

In depth

The full guide, about 4 minutes to read.

The system behind most sentencing and parole decisions

The most consequential algorithmic tool in the custodial system is one most of the public has never heard of. The Offender Assessment System, universally known as OASys, was developed by the Home Office following pilot studies between 1999 and 2001 and rolled out across the prison and probation services of England and Wales by 2005. It generates risk scores, drawing on both fixed historical factors and dynamic information gathered through offender interviews and self assessment questionnaires, that feed directly into decisions about sentencing, parole, and how an individual is managed in custody and under probation supervision. According to figures obtained by Statewatch through freedom of information requests, more than 1,300 OASys assessments are completed every day across the system, and the database held more than seven million risk scores as of the start of 2026.

OASys has been used for more than two decades, and campaigners and academics have raised concerns for much of that time about racial and class based bias in how its risk factors are constructed, the accuracy of its predictions, and the difficulty anyone assessed by it faces in meaningfully challenging a score whose underlying methodology is not fully public. It has been described in academic commentary as a black box system that has shaped criminal justice outcomes for over twenty years without the level of external scrutiny its influence would seem to warrant.

A second, separate system for prison security categorisation

A distinct algorithmic tool, the Digital Categorisation Service, is used specifically to inform how prisoners are security categorised, a decision that determines the conditions, restrictions and type of establishment someone is held in. Like OASys, it does not make the final decision by itself. It highlights risk information drawn from a prisoner's record and suggests an initial categorisation, which prison staff then review and may change. The distinction between a tool that suggests a categorisation for a human to confirm and one that determines the outcome outright matters enormously in practice, since it is the genuineness of that human review, rather than the mere presence of a human somewhere in the process, that determines whether meaningful oversight actually exists.

OASys is being replaced

His Majesty's Prison and Probation Service has confirmed it is developing a successor system, known as Assess Risks, Needs and Strengths, or ARNS, built in house by Ministry of Justice Digital in consultation with Capita, which provides technical support for the existing OASys system. An early prototype entered pilot testing in December 2024, and HM Inspectorate of Probation reported in April 2025 that a full national rollout was scheduled for completion by autumn 2026. The stated aims are more accurate, actuarial data informed risk assessment and faster, more consistent recording of the risk assessments, risk management plans and sentence plans that OASys currently produces.

Replacing a system that has operated for two decades is a significant undertaking in its own right, and the transition raises its own question distinct from whatever the merits or flaws of ARNS itself turn out to be: whether the criticisms levelled at OASys over accuracy, bias and contestability are genuinely addressed in the new system's design, or whether a new tool, built to do largely the same job with better underlying technology, ends up reproducing the same fundamental oversight gap under a different name.

Parliament has flagged the oversight gap directly

Written evidence submitted to a parliamentary committee in April 2026 raised concerns in strikingly direct terms about how algorithmic and AI tools in prisons and probation are currently scrutinised. It specifically named HM Inspectorate of Prisons, HM Inspectorate of Probation and the Prisons and Probation Ombudsman as bodies whose oversight of these tools needs strengthening, and criticised existing inspection practice for focusing on whether a required process was formally completed rather than examining how the underlying algorithmic assessment actually functioned, a distinction the evidence characterised as leaving genuine scrutiny of risk related tools largely confined to HMPPS itself. Given the department's own stated ambition to lead government adoption of AI more broadly, the submission argued this combination, extensive and growing algorithmic tool use alongside comparatively shallow independent scrutiny of how those tools actually work, warranted direct parliamentary attention.

What is not settled

There is no published, independent evaluation establishing how accurate OASys risk predictions have actually been over more than two decades of use, as distinct from the tool's own internal validation, nor is equivalent independent evidence yet available for its replacement, ARNS, which remains in the rollout stage. Detained and supervised individuals continue to have limited practical means of challenging a risk score that shapes decisions about their liberty, a constraint that applies with particular force in a custodial setting, where the person most affected by an algorithmic assessment typically has the least practical ability, of anyone touched by police or criminal justice AI, to contest it.

Records on the tracker

Open the full tracker

25 records on custody and prisons, most recently checked first: 18 in use now, 1 withdrawn and 6 on the horizon. Each is one police force or public body using one system.

  1. United StatesLEO Technologies

    Texas Department of Criminal Justice — Verus AI monitoring and transcription of inmate calls

    Texas

    OperationalChecked
  2. United StatesNot disclosed

    Texas Department of Criminal Justice — AI search and analytics platform to monitor border prisons and transcribe speech

    Texas

    ProposedChecked
  3. United StatesNot disclosed

    Louisiana Department of Public Safety and Corrections — AI transcription and analysis of incarcerated people's phone calls

    Louisiana

    ProposedChecked
  4. United StatesLEO Technologies2025

    Oklahoma Department of Corrections — Verus AI phone call monitoring

    Oklahoma

    OperationalChecked
  5. United StatesBancroft Global

    Orleans Parish Sheriff's Office — Bancroft Global AI tools to monitor jail calls and social media

    New Orleans, Louisiana

    ProposedChecked
  6. Spain

    Departament de Justícia i Qualitat Democràtica (Generalitat de Catalunya), prison services — RisCanvi prison risk assessment

    Catalonia

    OperationalChecked
  7. Russia2022

    Federal Penitentiary Service of Russia (FSIN) — "Digital SIZO" AI video surveillance and incident management pilot

    Moscow, Moscow remand prison SIZO-2 (Butyrka)

    TrialChecked
  8. Russia2025

    Federal Penitentiary Service of Russia (FSIN) — "Digital IK" AI video surveillance pilot with biometric identification

    Vladimir Oblast, correctional colony IK-6 (Melekhovo)

    TrialChecked
  9. United StatesLeo Technologies (LEO Technologies)

    Federal Bureau of Prisons — AI-enabled transcription, translation and analysis of inmate communications (Verus)

    Nationwide, 120 institutions

    ProposedChecked
  10. United StatesBI2018

    US Immigration and Customs Enforcement — SmartLINK facial verification for Alternatives to Detention remote check-ins

    Nationwide

    OperationalChecked
  11. United StatesSAS

    Federal Bureau of Prisons — BRAVO classification: predicting misconduct to assign security levels

    Nationwide

    OperationalChecked
  12. United StatesSAS

    Federal Bureau of Prisons — NLETS arrest classification

    Nationwide

    OperationalChecked
  13. United States

    Federal Bureau of Prisons — Static-99 Sex Offender Data System

    Nationwide

    OperationalChecked
  14. United StatesAdvanced Technologies Group

    Federal Bureau of Prisons — Truview analysis of inmate-related systems for investigations

    Nationwide

    OperationalChecked
  15. United StatesTruNarc, Smiths Detection (as listed in the inventory)

    Federal Bureau of Prisons — TruNarc and Smiths Detection narcotics testing of suspected mail

    Nationwide

    OperationalChecked
  16. United States

    Federal Bureau of Prisons — UAS threat detector near prison facilities

    Federal prison facilities

    OperationalChecked
  17. United StatesPocketalk2025

    New York City Department of Correction — Pocketalk real-time translation device for people in custody and visitors

    New York City

    TrialChecked
  18. United Kingdom2025

    HM Prison and Probation Service — Justice Transcribe (AI transcription for probation meetings)

    England and Wales

    OperationalChecked
  19. United Kingdom2019

    HM Prison and Probation Service — Violence in Prisons Estimator

    England and Wales

    OperationalChecked
  20. United Kingdom

    HM Prison and Probation Service — Remote probation check-ins with AI identity matching

    South West, North West, East of England, Kent, Surrey and Sussex

    ProposedChecked
  21. AustriaCogVis Software und Consulting GmbH (project partner)2022

    Austrian Federal Ministry of Justice (prison service) — KIIS pilot: AI sensor and smartwatch monitoring of movement in prisons

    Vienna and Lower Austria prisons

    DiscontinuedChecked
  22. Germany

    Lower Saxony Ministry of Justice (prison service) — AI camera analysis for suicide and violence prevention in a prison

    Lower Saxony (JVA Oldenburg)

    ProposedChecked
  23. United KingdomUK Ministry of Justice; technical support from Capita2001

    HM Prison and Probation Service — OASys offender risk assessment

    England and Wales

    OperationalChecked
  24. United StatesLEO Technologies (Verus)2023

    Florida Department of Corrections — AI Prison Communications Monitoring (Verus)

    Florida

    OperationalChecked
  25. United KingdomNot publicly specified2025

    HM Prison and Probation Service — AI language analysis of seized prisoner phones

    England and Wales

    TrialChecked

Latest coverage