Justice Transcribe

Justice Transcribe is an AI tool the Ministry of Justice built in-house for probation staff in England and Wales. It records supervision meetings with people on probation, transcribes them, and drafts structured summaries for the case record. It makes no decisions itself.

How it is used

An officer records a supervision meeting with a person on probation. Supervision meetings are the regular meetings an officer holds with each person they supervise. Justice Transcribe transcribes the meeting, summarises it, and structures the summary for the case record. The officer reviews and accepts each summary, and owns the record.

Before the tool, officers copied handwritten notes into digital systems by hand. The ministry says the tool supports admin work. Its AI Action Plan for Justice says AI should "support, not substitute, human judgment."

The ministry's Justice AI Unit built the tool. It counts it as a benefit that summaries stay consistent as a person moves between officers.

How fast it spread

Justice Transcribe was piloted in Kent, Surrey, Sussex, and Wales. The AI Action Plan for Justice, published 31 July 2025, reported that the pilot cut note-taking time by 50 per cent. Officers rated it 4.5 out of 5.

Official transparency data give a national start date of 7 October 2025. On 23 October 2025 the government announced that more than 1,000 probation officers would be equipped with the tool. The same press release announced OpenAI's expansion into UK data hosting through the ministry's partnership.

More than 150,000 meetings had been summarised by 12 February 2026, and more than 800,000 by 2 June 2026. So meeting volume roughly quadrupled in the second four months, as the tool spread from about 1,000 officers to the whole service.

On 9 June 2026, at London Tech Week, the government said every probation officer in England and Wales had the tool. That was about a year after the pilot.

What the time savings rest on

Every time-saving figure is the ministry's own, or users' own reports. None has been measured independently.

The October 2025 announcement said AI technology across the ministry was expected to save up to 240,000 days a year. The published savings for Justice Transcribe rest on an operating assumption of about ten minutes saved per meeting. Probation Workforce Transformation, which the sources list among the bodies running the deployment, set that assumption. The department calls it illustrative and "not a formal statistical estimate." It does not account for differences in meeting length, type, practice, or summary quality.

On that assumption, the first transparency report gave an indicative 25,000 hours. The second gave about 133,333 hours over 800,000 meetings. It also gave a modelled forecast of about 450,000 hours a year, which is not a measure of time actually saved. The June 2026 announcement said Justice Transcribe alone could free up about 18,750 calendar days a year.

More than 63,000 voluntary reviews in the app average 4.7 out of 5. The ministry cautions that they "should not be interpreted as representative of all users."

What has not been published

No accuracy rate for Justice Transcribe's transcripts or summaries has been published. No rate of officer corrections or overrides has been published either. Nor has any audit of the quality of the records, or any independent evaluation of the tool.

The scale is counted to the meeting. Whether the summaries are right is not measured at all. The tool reached every probation officer before any of those figures existed.

The risk tools that draw on probation records

The same ministry runs algorithmic reoffending-risk assessment over probation records at high volume. Statewatch, a non-governmental organisation, reported in April 2025 that OASys-based risk prediction made more than 1,300 assessments a day, of prisoners and people on probation. That was 9,420 assessments in one week of January 2025. It draws on probation and prison caseload systems and the Police National Computer.

The ministry's own validation found that these predictions of reoffending were less accurate for some ethnic groups. For non-violent reoffending, that was all Black, Asian and Minority Ethnic groups. For violent reoffending, it was people of Black and Mixed ethnicity. That is a property of the risk tools, not of Justice Transcribe.

The ministry is rolling out a successor, Assess Risks, Needs and Strengths, during 2026. The AI Action Plan also sets Justice Transcribe beside a single offender identity system and AI-powered search of case materials for risk factors.

So the records Justice Transcribe helps write are among the records that high-volume risk scoring draws on. No published source documents a named data pipeline from the summaries into the risk tools. The documented connection is the shared case record.

No one has reported the tool being withdrawn, a lawsuit over it, or a ruling that it caused harm.

What outside commentary says

A peer-reviewed Probation Journal editorial by Jake Phillips, published online in December 2025, calls transcription tools the "first wave" of AI in probation. Citing policing research, it warns that "perceived efficiency gains have not materialised in practice."

The editorial flags bias and the erosion of professional judgment. It also flags model sycophancy: an AI's tendency to agree with what its user seems to want. And it raises an open question about who is accountable when an officer acts on a record an algorithm has shaped.

Mike Nellis wrote a critique for the Centre for Crime and Justice Studies, dated 26 March 2026. It records a concern of the Public Accounts Committee and the National Audit Office. They warned that the pace of new digital tools could "disrupt services, contribute to poor outcomes and staff stress."

The critique notes that "the MoJ does not have a strong history of implementing digital change programmes well." The MoJ is the Ministry of Justice. It also flags the OpenAI contract as a risk of pressure from a supplier.

What this case asks

The official account is reassuring. The tool is admin support, it decides nothing, and a person reviews and owns every record. The case file points to what that account leaves out: who else reads the record.

Officers read it to supervise people. The ministry's risk tools draw on probation records to score reoffending risk. The network assumes that includes the records Justice Transcribe helps write. A summary written to save ten minutes can become an input to a risk score. The sources describe no check between the two.

The case file also stresses the order of events. The tool reached the whole workforce before anyone published how often it is right. It argues for proving the checks before the next group of officers gets the tool, not after.

When a record serves two readers, an officer owning the record is not the same as the record being safe for a machine to read.

What this network is drawn from

This network follows the pattern the case file describes. It is not a reconstruction of the actual tool. It shows Justice Transcribe, the recordings, the officers, the case record, the OASys risk tools, the ministry's AI governance, and the recall and court-report step. People on probation are outside the network.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Closing a pathway means mistakes stop passing along it. The work along it goes on.

This case has a budget of 12 units. Understand the system, the tool that pays for ongoing study of what the deployment is doing, costs 3 units under Service Targets Only. It costs 4 under the two higher levels. While it is in use, four tools cost 1 unit less each: Review on schedule, Escalate checks, Keep prompts neutral, and Upgrade model. Its stronger setting costs 6 units and takes 2 units off each, to no less than 1.

The network starts at a tipping point: its mistakes could either die out or build on each other. In a self-correcting network, mistakes die out before they build. Seven failure pathways are open at the start. They are Recording to Justice Transcribe, Summary to the officer, Summary into the case record, and History read at next meeting. The other three are Case record to OASys, Risk score to the officer, and Case record to recall and reports.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets are not met at the start. Mark AI-written records alone meets them, for 2 units. More than a thousand different sets of tools within the budget meet them.

Service and Safety Targets and All Governance Targets ask you to close every failure pathway, among other targets. Both can be met within the budget.

Under Service and Safety Targets, 15 different sets of tools meet the targets, and the cheapest cost 10 units. Every one includes Vet connections and Escalate checks. Vet connections is the only tool on offer that closes Case record to recall and reports. It also closes Recording to Justice Transcribe and Case record to OASys. Escalate checks is the only one that closes Summary to the officer and Risk score to the officer.

Each of those sets also includes Gate record entries or Store less data, which close Summary into the case record. Each also includes Mark AI-written records or Understand the system, which close History read at next meeting.

Lingering effects is a Dynamics setting in which damage outlasts its cause. It is on by default, off under Service and Safety Targets, and always on under All Governance Targets. With it on, Vet connections changes nothing unless Understand the system is also in use.

So under All Governance Targets, two sets meet the targets. One is Understand the system, Escalate checks, Vet connections, and Store less data, for 11 units. The other adds Keep prompts neutral, for 12. Gate record entries in place of Store less data closes the same pathways, but Justice Transcribe then adds too little to the work for that level.

The three checks the network draws are needed for no target. Check with a second model adds the check between Justice Transcribe and OASys. Assign a challenger and Peer sharing rules add the accuracy check on summaries. None of them closes a failure pathway. No tool on offer adds the check of the case record against its source.

More is not better here. Using every tool on offer, each at its strongest setting, costs 39 units, more than three times the budget. It meets the targets at none of the three levels that set them, because Justice Transcribe then adds too little to the work.