Prosecutor Cited Nonexistent Case in High Court, Renewing AI Accuracy Concerns
A Crown Prosecution Service lawyer reportedly cited a made-up precedent in a high court child case, reviving worries about AI errors in UK courts.
According to The Guardian, a senior crown prosecutor acting in a high court case cited a legal precedent that does not exist. The lawyer, Ravinder Dhillon, was representing the Crown Prosecution Service during a dispute over whether a child should face prosecution for the rape of his younger sister.
The Guardian reports that the fabricated case came up while the court was considering that prosecution question. Court documents, which the outlet says it has seen, reveal the reference was to an entirely invented case rather than a real ruling.
The paper frames the episode as an exclusive and places it within a wider pattern. It describes the error as one of dozens of AI-related mistakes that have appeared in court proceedings, and it characterizes the incident as a fresh sign of what it calls AI hallucinations entering the justice system.
The reporting centers on the nonexistent precedent and its appearance in a case involving an alleged child rape. The article does not, based on the available reporting, describe the outcome of the proceedings or any formal response from the prosecution service beyond what is already covered in the court documents.
Background
Large language models are systems trained to produce fluent text by predicting likely word sequences. They do not look up facts in a verified database by default, so they can generate statements that sound authoritative but are false. Researchers call these errors hallucinations. When a model is asked for case law, it may produce citations with plausible names, courts, years and reporter numbers that correspond to no real decision.
Legal work depends on precedent, so citation accuracy is central. Lawyers and judges rely on reported decisions that others can check. A fabricated authority can waste opposing counsel's time, mislead a court, and, if it goes unnoticed, shape a ruling. Courts in several countries have reacted over the past few years. Some have issued practice directions requiring lawyers to verify AI-assisted material, and some have sanctioned attorneys who filed briefs containing invented citations.
The problem is not limited to lawyers working alone. Legal research tools, document drafting software and general chatbots are all used in practice, and they vary in how they retrieve and check sources. A tool connected to a verified legal database reduces some risks, but it does not remove the duty to confirm that a cited case exists and says what it is claimed to say.
Criminal cases add a further layer. Prosecutors have a duty to act fairly and to put accurate law before the court. Errors in that role can affect defendants, complainants and the wider public's confidence in the process. Human review, checking citations against primary sources, and clear internal rules for using AI tools are the standard safeguards that institutions have been urged to adopt.
Why it matters
The episode shows that AI errors are no longer confined to hobbyist experiments or low-stakes writing. When a senior prosecutor in a high court matter relies on a fabricated authority, the consequences reach the justice system directly. The case is sensitive, involving an alleged child rape, which raises the stakes for accuracy and for the public's trust that serious matters are handled with care.
It also highlights a gap between the speed at which these tools can be adopted and the slower work of building verification habits. A citation that looks real can pass a quick read. Without a step that checks each authority against an official record, errors can enter filings and hearings. The fact that The Guardian describes dozens of similar incidents suggests this is a recurring problem rather than a single lapse.
For the legal profession, the lesson is practical. Firms and public bodies may need clear policies on which tools are permitted, what must be verified, and who is accountable when a mistake is found. For technology developers, the episode is a reminder that a confident-sounding answer is not evidence of a correct one. Tools used in high-stakes settings need to make sourcing visible and make uncertainty easier to spot.
For the public, the story is a reminder that institutions are still working out how to use these systems responsibly. Trust in courts depends on the assumption that every cited authority has been checked, and incidents like this test that assumption.
What to watch
Key questions include how the court responds in the proceedings themselves and whether it sets out any formal findings about the citation. Observers will also want to know whether the Crown Prosecution Service publishes guidance or review steps on the use of AI tools by its lawyers, and whether it describes how the error was identified and corrected.
Further reporting may show whether other filings in the same matter or in related cases contained similar problems. Watch for updates from court authorities or professional regulators, as well as any new practice directions addressing AI-assisted research in criminal and civil cases.
Finally, consider whether the wider debate moves from general warnings to enforceable verification standards. The next steps will indicate whether institutions treat these errors as isolated mistakes or as a structural risk requiring firm safeguards.
Viewpoints
Commentary from our columns: three angles on this story.
The Skeptic open questions and weak points
The headline number that matters here is not the one people will remember. A single invented precedent is a striking detail, but the more useful question is how many errors like it go unnoticed. If The Guardian's account of dozens of AI-related mistakes is accurate, we are only seeing the cases that someone happened to catch. That should make us cautious about claims that these tools are getting reliable enough for legal work.
There are also open questions about what this incident tells us. Did the prosecutor rely on a general chatbot, a specialist legal tool, or something else? The answer changes the lesson. A tool marketed for legal research failing is a different problem from a general model improvising. Without clarity on that, it is easy to draw conclusions that fit whatever view we already hold.
Be wary of anyone who says the fix is simply to check citations. Checking is the right step, but it depends on people actually doing it under time pressure, and on institutions making it a requirement with consequences. A rule that exists on paper but is skipped in busy courtrooms offers little protection. The real test is whether systems catch errors before they reach a judge.
The Builder the engineer's view
From an engineering standpoint, this failure is predictable. A language model produces text that resembles legal citations: a party name, a court, a year, a reporter reference. If nothing in the pipeline checks those strings against an authoritative index, the model will happily generate something that looks right. The fix is not to make the model sound less confident. It is to build a verification step that treats every citation as a claim to be looked up.
The practical design is straightforward in principle. Extract every authority from a draft, resolve it against a trusted database of reported decisions, flag anything that fails to resolve, and require a human to sign off on exceptions. Good tools also show which passage supports which proposition, so a reviewer can confirm that the cited case says what the argument claims. None of this is exotic, but it is easy to skip when the drafting tool feels fast and fluent.
The lesson for developers is that retrieval and generation should be separated where stakes are high. A system that generates citations from memory is solving the wrong problem. Systems used in courts should only cite what they can link to a verifiable source, and should fail loudly when they cannot.
The Rulebook privacy, safety and regulation
Courts exist to make accurate, testable claims about the law. When a filing cites a case that does not exist, the harm goes beyond embarrassment. It can mislead a judge, distort an argument, and undermine the fairness owed to everyone involved. In a case concerning an alleged child rape, the seriousness of the matter makes that duty weigh even more heavily.
Regulation in this area is still catching up. Several jurisdictions have issued guidance requiring lawyers to take responsibility for what they file, regardless of what tool helped draft it. That principle is important: the duty to verify does not transfer to software. The open question is how strictly that duty is enforced, and whether professional regulators will set clear expectations for public prosecutors as well as private attorneys.
Transparency also matters. The public should be able to learn how an error like this was identified and corrected, what steps were taken, and whether similar problems have appeared elsewhere. Clear accountability builds trust. Vague reassurance does not.