According to news on June 7, a number of police forces in England and Wales have been asked to stop using AI tools to prepare court statements and criminal justice documents because errors in the relevant output will affect the reliability of subsequent legal procedures. According to a report by the Financial Times on June 5, the request was made by Alex Murray, the head of AI work in the British police.
Murray is also the director of the National Crime Agency and head of AI affairs at the National Police Chiefs Council (NPCC). The UK just announced in January this year that it will invest 115 million pounds to build a national police AI center to promote the more unified use of AI by the 43 police forces in England and Wales.
Court materials cannot rely solely on “pre-submission verification”
What is being halted this time are materials that would enter the criminal justice process, including court statements and related case documents. For police officers, AI-written reports seem to save time; but for witnesses, defendants and prosecutors, the problem is that these words may become the basis for judging whether the facts and procedures of the case are in compliance.
If AI writes non-existent facts, confusing timelines or erroneous summaries into materials, and subsequent personnel use this material to organize the case, prepare evidence statements, or submit it to the court, the error is no longer just a "wrong sentence", but may spread along the case handling chain. The so-called "checking AI output" in this kind of scenario is also difficult to solve with one prompt: Should the checker check every fact and every quote, or just judge whether the text is smooth? If the answer is to do it all again, the time saved by the AI will be offset.
The British Crown Prosecution Service (CPS) has issued an AI ethics statement, emphasizing that AI tools should be used under human supervision, with special emphasis on fairness, transparency and accountability. However, once front-line police documents enter criminal proceedings, the boundaries of responsibility are more sensitive than in ordinary office settings: clear records need to be kept of who wrote the materials, who checked them, and who is responsible for the mistakes.
The fictional football match in January this year has sounded the alarm
The British police have become cautious about the risks of AI documents, which is not without precedent. In January this year, former West Midlands Police Chief Constable Craig Guildford apologized to the UK Parliament's Home Affairs Select Committee for an intelligence material containing an AI error.
This material mentioned that fans of the Israeli Maccabi Tel Aviv Football Club (Maccabi Tel Aviv) had conflicts with the Premier League West Ham United, but the relevant game did not exist. The Guardian reported that Guildford admitted that the error stemmed from fictional content generated by police using Microsoft Copilot.
This incident exposed that when AI-generated content is put into formal law enforcement or public decision-making documents, it is difficult for external readers to judge from the text which content has been originally verified and which is just a secondary representation after the model is generated. When it comes to court materials, this opacity will directly affect the rights of the parties and the credibility of the case.
AI police promoters take the lead in withdrawing the most sensitive links
The British government announced the Police AI Center in January this year with the goal of accelerating the establishment of a unified testing, training and management mechanism. The NPCC said at the time that the £115 million investment would help the police solve crimes faster and was expected to free up about 6 million police hours per year, equivalent to the workload of about 3,000 full-time police officers.
Therefore, this suspension is more like drawing a boundary for AI policing: AI can first be used for back-end work that is low-risk, traceable, and easy to review, such as data desensitization, preliminary screening of non-emergency alarms, and internal information sorting; however, court testimony and criminal justice documents are high-risk links, and cannot only be covered by "human beings are ultimately responsible."
If you are a witness, reporter, suspect or defendant in a case, the statement materials you face should come from someone who can be held accountable, not a set of AI texts whose generation process is difficult to restore afterwards. It is of course important for the police to save time in writing reports, but not at the expense of reducing the credibility of judicial materials.
The UK's suspension this time will not be an isolated case. In the past two years, there have been cases in the United States, Canada, Australia and other places where lawyers cited artificial AI cases and courts issued guidelines for the use of AI. Now the risk has advanced from lawyers submitting documents to the police's underlying case files. The real difficulty in AI entering the judicial system lies not in the quality of text generation, but in whether every statement that may affect the rights of others can be verified, traced and held accountable.