Following a patrol duty, a police officer hits a single button; shortly thereafter, the officer receives an edited version of the recording captured by the body-worn camera. This document, signed and dated by the officer, becomes part of the case file. In addition, the initial draft produced automatically by the machine disappears. This has been reported in a 2025 investigation conducted by the Electronic Frontier Foundation concerning the Draft One software developed by Axon.

Although it seems like a minor aspect of the workflow, its effects extend far beyond what might be initially assumed. Indeed, the creation of an accurate account plays a critical role within criminal investigations because a police report influences several processes, including charging decisions, plea bargaining, and the cross-examination phase, possibly many years following the event. Moreover, without knowing who composed it, determining the validity of evidence presented during prosecutions is complicated.


We begin with the back office

Since 2016, when the media organisation ProPublica discovered racial biases within the COMPAS predictive model, public discussion regarding applications of artificial intelligence within criminal justice had revolved around decision-making programmes. The findings revealed significant disproportion between black and white defendants' assessments concerning the likelihood of reoffending. The issue remains relevant today. Nonetheless, this particular iteration is distinct. This round involves deploying AI solutions to handle the mundane, time-consuming aspects of the criminal court system's workflow.

According to the report published by Sir Brian Leveson in February 2026, in which he reviewed the state of England and Wales' criminal courts, the introduction of technologies such as automatic transcriptions, summaries generated via AI of the prosecutor's case, and assistance from machine learning software in analysing vast amounts of data in disclosure processes received approval. At the same time, Leveson drew strict boundaries. He stressed the importance of maintaining "human decision-making" and expressed reservations towards algorithms predicting court decisions' outcomes.

In July 2025, the Department of Justice unveiled its Artificial Intelligence Action Plan for Justice. The document was revised after only one year in September 2026 with the inclusion of a new strategic priority: "Identifying Emerging Risks Posed by AI". Such a development appears significant. In fact, a mere twelve months later, the agency's focus seems to shift towards protection rather than implementation.

It comes as no surprise. Courts have too few hours within the day; police departments lack sufficient officers; prosecutors must deal with overwhelming amounts of data extracted from smartphones and camera footage. There appears to be plenty of room for improvement here. For instance, the software capable of transcribing hours-long audio recordings and analysing large datasets of information may seem particularly attractive.


And yet, the claim about efficiency does not seem to hold up either

There is no substantial proof of this benefit, despite what may be said elsewhere. According to the results of the randomised control test, registered in advance, conducted among 85 officers who had to compile 755 police report documents within a yearlong experiment in Manchester, NH, the use of the algorithmic aid failed to help reduce the time needed for writing. Additionally, after analysing more than 6,000 police report documents compiled over the course of one year, researchers drew similar conclusions.

Concerning quality, the results are hardly comforting either. In another study led by Ian Adams, the University of South Carolina criminologist, participants - 92 experienced police supervisors - had to evaluate 80 documents compiled both manually and using the mentioned algorithms' aid. While the latter were ranked significantly lower in terms of accuracy, evaluators failed to recognise the difference between the two categories above chance levels. Although this paper is still available only as a preprint, without peer review, Adams concluded unequivocally: "There isn’t one independent evaluation that would confirm vendors’ statements." 

None of this implies that such technologies lack value completely. It seems clear that when compiling massive volumes of evidence or conducting searches, machines perform those tasks much faster than human investigators. The problem lies elsewhere: in accepting efficiency claims while they remain untested in applications where mistakes might deprive citizens of freedom.


When the record itself has been created

The more serious issue is systemic. The generative model will create coherent and believable text, and coherence is persuasive. An officer reviewing a nicely written draft towards the end of a shift may take the wording at face value. But when a lawyer challenges that wording years later, there will be no way for him to prove that the wording was originally created by the software and not the witness.

Legislators have begun to react. The California law SB 524, which took effect on 1 January 2026, mandates all police reports produced in whole or in part with the use of artificial intelligence to come with a clear disclaimer, mention the name of the program used and keep the AI-generated draft in archives along with the official report for the same period of time and the audit trail of its users. It is a very narrow regulation, but it directly addresses the problem of provenance highlighted by the EFF.

Lawyers cannot be exempt from it as well. The database collected by Damien Charlotin, a researcher at HEC Paris, counted 1,240 cases of AI-created fake citations in US court filings by July 2026, up from 11 in 2023. The head of the defence team apologised in August 2025 in Melbourne when the submissions in the case of teenage murder contained fabricated quotes and citations to judgements that do not exist, postponing the hearings by a day.


Regulation is here, albeit slowly

The EU AI Act, which places AI utilised in law enforcement in the high-risk category, requires risk management, documentation, and human oversight. However, this has not been enforced, as per a provisional deal made in May 2026, which has pushed back the compliance date for Annex III high-risk systems from August 2026 to December 2027. This gives vendors and law enforcement agencies an additional 16 months to comply with the Act's regulation.

The United Kingdom lacks such regulation altogether, depending instead on guidelines, procurement policies, and judicial oversight to handle these technologies. This is flexible enough but puts the onus on individual law enforcement agencies and courts.


What decision-makers should take from this

The implication for justice ministries, police authorities, and the vendors who sell to them is clear: the question of whether or not “should AI make decisions” is outdated. AI is influencing the inputs to decisions which is the area in which accountability is the most easily lost.

Three concrete commitments should follow. Any document generated by AI technology that might potentially serve as evidence must carry a disclosure and have its original draft kept, like California already does now. The productivity claims should be tested independent of procurement. And the defence counsel must have access to the same provenance data as the prosecution, otherwise, the adversarial process becomes inherently biased.

The use of AI may significantly speed up an overburdened criminal justice process. But in criminal justice, it is crucial to be able to ask who said what and why. Technology, which answers the first question quicker but leaves the second unanswered, is not a reform. It’s an unnecessary risk wrapped in a user interface.