Two fake legal authorities. Two apologies to the High Court. One very simple lesson that every trainee, paralegal and future solicitor needs to take seriously before AI catches them out the same way.

If you needed proof that AI-in-law anxiety is not overblown, here it is. The Crown Prosecution Service, the body responsible for prosecuting criminal cases across England and Wales, has admitted putting two completely fictitious legal authorities before the High Court. Not a small firm. Not a single overworked junior. The CPS.

Here is what happened, why it matters, and what it tells you about the exact skill firms will be testing for in every interview from now on.

The case, Andreea-Maria Tobosaru v Court of Law Craiova, Romania, was an extradition appeal. The CPS opposed the appeal, and in doing so its submissions relied on two legal authorities to support the continued application of established House of Lords reasoning.

There was just one problem. Neither authority existed.

The fabricated citations first appeared in the CPS's grounds of opposition, prepared the previous year. From there, they were carried forward into a further document headed "Respondent submissions", prepared by junior counsel. They sat before the court at the permission stage, unnoticed. By the time senior and junior counsel prepared the skeleton argument in February 2026, the false citations had been identified internally and were quietly dropped, but the skeleton never explained to the court that they had ever been wrong in the first place.

It was the appellants' own legal team who eventually raised the issue, after they were unable to locate either case in any legal database.

Two fake cases slipping through is bad enough. What happened next made it worse.

When the CPS first wrote to the court to address the error, it stated the mistake had been spotted by the appellants' counsel. That turned out not to be the full story. It later transpired that the CPS's own junior counsel had independently identified the fake cases and raised concerns internally, just before the appellants did. The CPS had to write to the court a second time to correct its own explanation.

So this was not simply an AI error. It was an AI error that survived multiple stages of review, moved between documents, and was then followed by an account to the court that itself needed correcting. Mr Justice Sweeting accepted there had been no deliberate attempt to mislead. But the sequence of events speaks for itself.

The CPS's own explanation is worth reading closely, because it is the exact explanation you will hear again and again as AI use grows across the profession. The Chief Crown Prosecutor told the court the citations were likely to have originated from the use of generative AI. But the CPS was careful to draw a distinction: while the immediate source of the error was AI, the operative cause was human error, specifically, the reviewing lawyer's failure to verify the authorities before the document was filed.

That distinction matters, and it is the whole point of this article. The tool did not fail here. The check did.

Mr Justice Sweeting dismissed both extradition appeals on their substantive merits, so the fictional authorities ultimately had no bearing on the outcome. But he did not let the episode pass without comment.

He acknowledged it would be naive to assume AI use in legal work will not keep increasing, and accepted that this may be both necessary and beneficial. At the same time, he was clear that the case highlighted the risks of using AI without appropriate oversight, particularly in legal research. The message was not "stop using AI." It was "start checking what it gives you."

If this felt like a one-off, it would still be a good story. It is not a one-off. Earlier this year, a City law firm was criticised by a High Court judge after a junior lawyer used AI to draft letters in an insolvency matter that turned out to contain misleading information, prompting the firm to refer itself to the SRA. Legal commentators tracking AI hallucinations in the courts have noted a steady rise in similar incidents over the past two years, across firms of every size and, now, prosecuting authorities too.

The pattern each time is strikingly similar. A fabricated citation or quotation is generated, it is not checked properly before filing, and it then survives because later readers assume someone earlier in the chain already verified it. That assumption is exactly how a single hallucination becomes embedded in the official record.

You are almost certainly going to use AI tools in your legal career, whether that is for research, first drafts, or summarising documents. Firms know this too, and increasingly expect it. What separates a safe use of AI from a career-ending one is not whether you used it. It is whether you checked what it gave you before anyone else relied on it.

If you get an interview question about AI and the legal profession, and you will, resist the urge to give a generic answer about efficiency and innovation. Everyone in the room will say that. Instead, show that you understand the actual risk: AI-generated legal research can produce citations, quotations and even entire case summaries that sound completely plausible and are entirely fictional, and once one slips through unchecked, it tends to spread through a document, then through a case, before anyone catches it.

The solicitors who will be trusted with AI tools going forward will not be the ones who use them the most. They will be the ones who never file anything they have not personally verified.

A single missed verification step turned a routine extradition appeal into a story that made headlines across the legal press and a case study every trainee should know. The technology was not the failure here. The failure was assuming someone else had already checked.

Whatever tools you end up using in practice, treat every AI-generated citation the same way you would treat a fact given to you by someone you have just met: helpful, plausible, and entirely your responsibility to verify before you put your name to it.

Do you know someone who would benefit from this? Share it with your friends, colleagues and family. Knowledge is better shared.