Police Law Blog European Decisions Statutory Materials

AI-generated statements condemned by the Court of Appeal

R v Jerome Gibson [2026] EWCA Crim 1302, 22 September 2026, judgment here.

The Court of Appeal has issued a stark warning about the use of artificial intelligence (‘AI’) in criminal proceedings after discovering that a police officer used Microsoft Copilot to generate a victim personal statement (‘VPS’) designed to maximise emotional impact and so influence the sentencing of an offender.

Given the unrelenting march of AI-generated content already appearing in everything from police misconduct proceedings, inquests, and applications for judicial review,  this case nevertheless provides a stark early warning about the misuse of AI in any judicial forum.

The background

In criminal proceedings the accused had pleaded guilty to four offences of (i) GBH contrary to section 20 of the Offences Against the Person Act 1861; (ii) threatening with a bladed article, contrary to section 52(1) and (6) of the Offensive Weapons Act 2019, and (iii) two offences of damaging property, contrary to section 1(1) of the Criminal Damage Act 1971.  He had been sentenced to 44 months’ imprisonment.

When the offender sought permission to appeal his sentence, one of his grounds of appeal was that at the sentencing hearing the Crown had relied on a VPS which had been prepared by a police officer using a form of AI.

It seems that the police officer had uploaded the victim’s interview transcript into Microsoft Copilot with the following direction: “From this victim interview I want you to create the most impactful victim personal statement possible, with a focus on the knife offence but featuring the abuse as a whole also. I want you to create it so that it meets the higher harm category for sentencing.”

The officer then quoted from the sentencing guideline and further instructed the AI that its goal was “to achieve the highest sentence and the most tear‑jerking VPS to make a judge or reader weep when reading.”

Although the statement drew on what the victim had said in interview, the court remarked that it read very much as an exercise in rhetoric or advocacy.

Happily, it seems that at sentencing the judge recognised it as such, and so relied only on very limited factual matters which the victim had genuinely recounted during interview: namely that the offender had worn a balaclava and that she feared he intended to kill her. The Court was satisfied that these were not AI-generated fabrications and in turn held that the improperly generated VPS had not increased the sentence imposed. The appeal on this particular point therefore failed.

However, in its remarks regarding the use of AI to create witness statements, the Court of Appeal’s response was unequivocal:

“We need hardly say that no police officer should ever prepare a witness statement of any kind in this way and that no statement prepared in this way should ever be relied on in a criminal case”

The Court described the officer’s use of AI to generate the statement as “deplorable”. It is perhaps no surprise the officer concerned had since been removed from frontline duties and was under investigation in relation to his conduct in this and other cases.

Commentary

That this was a criminal case is immaterial: the views of the Court of Appeal must apply with equal force in every jurisdiction. There should be no doubt that the use of generative AI to create a witness statement with the purpose of influencing the proceedings is wholly unacceptable wherever it occurs.

Of course, there is no rule that expressly requires a witness personally to type or draft their own statement.  It is perfectly proper practice for witness statements to be written by police officers based on accounts given to them by the witness, and phrases such as “account taken over the phone and statement read back to me” commonly appear in statements. It is also unremarkably commonplace for lawyers to assist with drafting their statement, and lawyers will not transgress any rule or professional code if they review draft statements for completeness and even suggest to the witness additional matters to be covered where the statement has not addressed a relevant factual issue.

However the critical requirements when producing such statements are that:

  • the contents accurately reflect the witness’s own uninfluenced evidence;
  • the statement is limited to the witness’ factual account;
  • the statement does not contain any unauthorised statements of opinion or make advocacy-style legal/factual submissions;
  • the witness has read or had the statement read to them before adopting it;
  • the witness has indeed adopted the statement as true as indicated by signing it; and
  • the statement contains the required declaration of truth.

Importantly, it is not the function of a factual witness statement to provide a commentary on the documents in the court bundle, nor to set out opinions of the evidence of others, nor to engage in matters of argument. The overriding duty remains to provide a truthful account of facts within the witness’ own knowledge.

In JD Wetherspoon plc v Harris [2013] EWHC 1088 (Ch) (here) Briggs J firmly emphasised this point, criticising statements which were effectively drafted advocacy rather than genuine witness evidence. He decried the trend of witness statements becoming vehicles for argument, commentary and submission. Finding that in those respects the witness statement before him was an abuse and that the abusive parts should be struck out.

In this case it was not the use of AI technology itself that was the issue, rather the Court’s criticism was directed at the corruption of the evidential process by the police officer, where AI happened to be the mechanism by which it occurred.

AI is increasingly being used in legal fora, and properly employed it can perhaps be a useful and time saving tool. But many of those working ‘at the coal face’ will also have experienced the misuse of AI particularly when very poor quality AI-generated applications are made or submissions are presented to the court.

Such applications and submissions are often extremely lengthy, unfocussed and verbose, with the same proposition expressed several times in slightly different language and treating every conceivable argument as equally valuable without any focus on the determinative issue. AI-generated submissions typically include legal authorities that are summarised rather than analysed, or recite legal tests without these being applied to the facts and so do not directly address the ultimate question before the court. It’s common for them to focus on trivial points, for example in the disclosure or evidence heard, and to make these appear major issues of concern or importance. They often include lengthy discussions of fairness, proportionality, justice or public confidence but without engaging with the precise issue before the coroner. The worst examples include citations of ‘hallucinated’ or non-existent authorities. More often AI produces citations which are technically correct but irrelevant to the point in issue, or authorities said to establish a proposition of law that they never actually considered.

Some Coroners have already responded by formulating a set of directions in their court regulating the production of AI-generated submissions (one example of Coronial directions re the use of AI, that are already in use, is appended below), but guidance would be beneficial across the judiciary.

A quick change could be to amend the statement of truth to require the signatory to confirm the statement is their own and not produced by AI (or indeed a statement that they have used AI to assist – something the House of Lords have recently implemented with respect of any written evidence submitted to their Committees).

Appendix 

Coroner’s Directions on the use of Artificial Intelligence

Confidentiality and uploading documents into AI systems

  • The documents disclosed to you by the Court, and the information contained within them, are provided solely for the purposes of the inquest. Unless such information was already in your possession or is otherwise in the public domain, it must be treated as confidential.
  • You must not upload any inquest documents, or information contained within them, to any open-source artificial intelligence platform or infrastructure hosted on a public or hybrid cloud environment.
  • The Court will regard the uploading of inquest documents to open-source AI tools as a breach of confidentiality. The only exception is where a closed-source AI system is used within a secure and controlled network environment that adequately protects the confidentiality of the material.

AI-generated legal submissions

  • Interested Persons may use artificial intelligence tools to assist with legal research and when crafting their written submissions provided such use is undertaken responsibly. Care must be taken to ensure that the Court is not misled, whether deliberately or inadvertently, by submissions, legal propositions, or purported authorities that are inaccurate, fictitious, or generated as a result of AI hallucinations.
  • Any person who uses AI to produce submissions to the court or a skeleton argument of law must independently verify the accuracy of all factual submissions, any statutory material cited and any legal authorities relied upon before presenting them to the Court.
  • No legal authority should be cited unless the person relying upon it has verified that it is a genuine judicial decision and that it supports, or is at least arguably capable of supporting, the proposition for which it is cited.
  • Where AI has been used to produce arguments of law the document must include a statement by the author confirming that they have personally reviewed any legal authority or statutory provision relied upon and can confirm that it supports the proposition of law for which it is cited.
  • A self-represented Interested Person bears the same responsibility for the content and accuracy of their written and oral submissions as legal representatives acting on behalf of Interested Persons.

Length and content of submissions

  • Written submission or skeleton arguments put before the Coroner must not exceed 15 pages of A4 (when set in a font of no less than 12 pitch and 1.5 line spaced) unless the Coroner presiding directs a different page length is to apply.
  • If it is deemed necessary to present a longer written submission the permission of the Coroner presiding must first be obtained before filing and serving the document. Applications for such permission must be made in writing, on no more than two sides of A4 (when set in a font of no less than 12 pitch and 1.5 line spaced) and explain the reasons why it is said that an extended submission is necessary.
  • Where a legal authority is cited in any submission it must include a reference to the paragraph number of the judgment that contains the proposition of law the legal authority is said to establish.
  • Submissions to the court in breach of the above directions may not be read by the Coroner. Alternatively, written submissions which do not comply with these directions may be returned to the author directing amendment and compliance.