The citation had a year, a court, and a number. It did not have a judgment behind it.
The Guardian reports that Ravinder Dhillon, a senior crown prosecutor in the Crown Prosecution Service rape and serious sexual offences unit, relied on a case she called CPS, R v J [2024] EWCA Crim 3529. The authority was imaginary. She was arguing that a prosecution should continue: a boy who was 11 at the time of the alleged rape of his seven-year-old sister. His lawyers had taken the CPS decision to the High Court, saying prosecutors had failed to weigh representations from social workers, youth workers, and a forensic psychologist.
If you are putting generative tools into professional work, this is the failure to price in. A general-purpose summary stood in for a source of record.
Two clocks, not one muddle
The dates only make sense if this case and the warnings around it are kept on separate tracks.
On this prosecution, The Guardian reports that the decision to proceed was Dhillon’s, taken in April and confirmed by the chief crown prosecutor. In October she admitted the citation was an error and said she had meant a real decision from 15 years earlier. The paper understands the reference came from an AI-generated summary produced during an internet search. It does not name the product, quote the summary, or identify the older case.
The challenge is the 2026 High Court judgment published as EWHC/Admin/2026/2778. The Guardian reports that two senior judges called her explanation “unsatisfactory, ” and that the fictitious reference did not affect the outcome of the boy’s challenge. The prosecution was still listed for trial in the youth courts later that week. That account does not name the judges or the chief crown prosecutor.
Around that case, the public record has three dated markers. In June 2025, in Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, the High Court treated the citing of fictitious cases as wrong even where the legal principle pinned to the fake authority happened to be correct. It asked the Bar Council and the Law Society to consider steps as a matter of urgency. Before this challenge, The Guardian reports, the CPS had already admitted a February extradition case involving two Romanians convicted of minor offences, in which a lawyer cited authorities that did not exist, likely because of AI. Opposing lawyers spotted them because they could not trace them. The service concluded that “the risk of recurrence was low.” That matter is linked to EWHC/Admin/2026/1720. In June 2026 the Solicitors Regulation Authority updated supervision guidance for solicitors. That last date belongs to a profession’s rulebook. It is not a finding in Dhillon’s case.
This was used as a legal source to argue for the prosecution of a child and that legal source doesn’t exist: that’s a serious failure.
That is Brian Flanagan, associate professor of law at Maynooth University, who studies how courts and lawyers use AI, in comments reported by The Guardian. He also put the professional point plainly. Authorities are treated as precious things. You cannot make them up. It is out of bounds.
Fluency is not a source
A generative system does not look the case up. It predicts likely text. A neutral citation is a familiar pattern, so the format of a real authority is easy to imitate, and the result can sound sure of itself while being empty. The Law Society quotes Dame Victoria Sharp, president of the King’s Bench Division, on AI-generated text that can be “astonishingly and beguilingly fluent, but false.” The Society treats this as a professional and rule-of-law problem. A quirky software bug does not explain it away. Duties not to mislead the court are not lifted because a tool helped with the sentence.
The Law Society also warns of a further risk that did not have to happen here. A hallucinated reference that reaches a judgment can shape how later readers describe the law, and then be swept back into the material the next model learns from. That is how a footnote with nothing behind it becomes a folk precedent.
Flanagan’s other warning belongs on the procurement brief. AI raises the risk “because of the time you can save.” The hours clawed back are the sales pitch. They are also the gap a detailed fiction can walk through.
A right result does not clean a fake authority
The comforting reading is that nothing turned on the ghost. The Guardian reports that the High Court said the fictitious reference did not affect the outcome, and that the youth-court trial was still listed. That finding covers the result. It does not give the citation a clean bill of health.
Flanagan separates two questions that are easy to mash together. Ayinde is about the lawyer’s conduct, and about what should follow. Citing a fictitious case is wrong even if the principle ascribed to it happens to be right. The 2026 challenge, on his account, is about what the law actually is. A court should not be stopped from applying the correct law merely because a lawyer propped it up with an authority that does not exist. He has said he is not sure the two rulings “sit together uneasily”. That phrase is his way of rejecting a clash between them. That is his argument, not a judicial finding that no harm was done.
The Law Society’s warning sits beside it. An undetected fake authority can travel. Public trust is the thing being spent. And when the lawyer works for the state, Flanagan argues, the failure “takes on a different aspect, because the power of the state is immense.”
The service had already called the risk low
The Guardian describes this as only the second known CPS case of the kind: a prosecutor citing authority that does not exist, in circumstances linked to AI. “Known” means publicly identified. It is not an audit of the file store. The February extradition matter was described as likely caused by AI, after opponents could not trace the cases. This one, as the paper understands it, came from an AI summary in an internet search. Those are related failures, with evidence of different quality.
After that February episode the CPS judged that the risk of recurrence was low. A senior specialist prosecutor then put another nonexistent citation in front of the High Court.
The wider pattern is mostly outside the CPS. The Guardian said the case added to dozens of AI-related citation errors in UK proceedings this year, mostly from people who are not legally qualified and are representing themselves, often in employment tribunals. Matthew Lee, a barrister at Doughty Street Chambers, keeps a UK tracker that The Guardian reported at 101 entries, covering confirmed, suspected, and disputed AI-related errors, plus related false citations. That mix is not a body count of proven hallucinations. Lee has said the identified cases may be the “tip of the iceberg, ” because many lower-court decisions are never published. One claim, then: published judgments understate the mess. A tracker that mixes confirmed, suspected, and disputed entries can flag the problem. It cannot count every case.
Lee has also flagged privacy, confidentiality, and the extra checking this dumps on everyone else. The reporting on Dhillon’s citation does not say that sensitive case material was pasted into a tool. An internet-search summary is a different route in. It was still enough, as The Guardian understands it, to manufacture a case name.
A policy is not a check
The CPS says it has a strict AI policy “which limits its use so that it is not used to make legal decisions and there must be human checks for accuracy.” The Guardian quoted a spokesperson saying the service “identified an error had occurred from conducting legal research using the internet, ” that the error was remedied once identified, and that an internal review looked at how it happened and whether other cases were affected. Learnings shared afterward included a line that should not have needed a review: case law must be cited from trusted legal sources.
Three things that public account does not explain:
- How a detailed reference to a 2024 Court of Appeal criminal case cleared whatever human check the policy requires.
- What the review found, beyond the statement that it happened and that trusted sources were restated. It does not identify the 15-year-old decision Dhillon said she meant.
- Whether anyone faced a sanction.
The June 2025 judgment, as the Law Society sets out the consequences discussed for lawyers who misuse AI, pointed to public admonishment, referral to a regulator, and contempt of court. Many crown prosecutors are solicitors or barristers, and so sit under the Solicitors Regulation Authority or the Bar Standards Board, and the CPS has its own conduct framework. This reporting does not say which seat applied, or whether any step was taken. The SRA’s June 2026 guidance, requiring appropriate human review, scrutiny, and professional judgment of AI-assisted work, is the direction of travel for solicitors. It is not proof of what the CPS rulebook demanded on the day, or of whether that rulebook was followed.
What transfers if you are buying the tool
A casual public search and a contracted legal tool carry different risks. One is an unbounded text predictor with a confident voice. The other can be logged, limited to material your organisation actually holds, and forced to show its working. Neither is safe if the last person in the chain is checking whether the paragraph sounds finished, rather than whether the thing exists.
That is the unglamorous core of AI automation in professional work. The tool drafts. A human verifies against a source of record. Someone is accountable if the source is missing. A workflow that files the summary into a client memo, a credit paper, or a response to a regulator repeats the High Court failure in a cleaner template. A correct final decision does not launder the fake source it leaned on.
Questions worth asking before the summary goes in the pack
-
What is the source of record?
A place a colleague can open and find the judgment. The CPS review restated that case law must be cited from trusted legal sources. An AI overview is a lead. The Guardian understands that is where this citation came from, and the public account does not name the product.
-
What check would catch a case that does not exist?
Not a skim for fluent prose. Someone has to look up the neutral citation on a source the organisation treats as authoritative and confirm the judgment is there. A line that says humans must check accuracy fails if that lookup is not the check.
-
Who signs the check?
The person who puts the authority in front of a decision-maker, before the filing, not a review team after the embarrassment. Here a senior crown prosecutor cited the case. The decision to proceed had already been confirmed by the chief crown prosecutor.
-
What happens when the check fails?
The fake citation comes out, a real source goes in, and someone asks whether other work used the same shortcut. The CPS says it remedied the error and reviewed whether other cases were affected. The June 2025 warning also treated misuse as a conduct issue. This reporting does not say whether any sanction followed.
-
Does “AI may not make the decision” cover research?
Not by itself. The CPS already says AI must not make legal decisions and that humans must check accuracy. A nonexistent case still reached the High Court. Research is part of the decision if the decision leans on it.
Cite from a place a colleague can open. Treat a polished overview as a lead, never as authority. Make the check specific enough that a fictional citation fails it. Do not congratulate the process because the final decision happened to be one you could have reached another way.
The time you saved will not defend the citation. That saved time is how the ghost got into the argument.