When the Crown Prosecution Service admits that hallucinated cases went before the High Court under its name, every law firm using AI should read the judgment. In Tobosaru [2026] EWHC 1720 (Admin), a Romanian extradition appeal decided on 8 July, Mr Justice Sweeting set out how two authorities that do not exist entered the CPS's written case. The judgment also records what the service did once it knew. The account is short, uncomfortable and useful.
Two cases that never existed
The invented authorities were Vlad v Romania [2017] EWHC 2060 (Admin) and Jeziorski v Poland [2019] EWHC 3885 (Admin). Both look entirely plausible. The names fit the subject matter, the years are believable and the court is the right one for an extradition point. Neither case has ever been decided. A search of the law reports returns nothing, because there is nothing to return. They first appeared in the CPS's grounds of opposition in June 2025, then travelled into counsel's submissions days later, which is how fabrication spreads through litigation, each document trusting the one before it. The errors came to light before the hearing and the final skeleton argument dropped them, so the false material played no part in the outcome. The court dismissed both appeals on their merits.
Human error, not machine error
The CPS told the court that the citations "were likely to have originated from the use of artificial intelligence", and then said the more important thing. The operative cause was not the tool but "human error in the failure to verify the authorities relied upon in formal submissions placed before the court". That framing deserves attention, because it is how English law now allocates responsibility for AI in professional work. The UK Jurisdiction Taskforce's recent legal statement, which we examined in when not using AI becomes negligence, settles the duty on whoever deploys the tool rather than on its maker. A filed document belongs to the lawyer who files it, whatever produced the first draft. Mr Justice Sweeting accepted that nobody intended to mislead the court, and observed that the episode "highlights the risks of its use without appropriate oversight particularly for legal research".
The 78-case review is the real lesson
What the CPS did next is the part your firm can borrow. It ran a full internal review. Senior managers re-examined 78 other cases conducted by the same lawyer and found nothing similar. The service reinforced verification obligations across its Extradition Unit and sent the learning points around the organisation. None of that needed new technology or new rules. It needed the organisation to treat one bad filing as a question about its processes rather than about one person's afternoon, which is the reading a court respects. Now hold your own practice against that standard. If a fee earner filed an invented authority tomorrow, would you know which other matters to check, and would anything on the file show who verified what before it left the building? A firm that can name the affected work and evidence its checks has an isolated lapse. A firm that cannot has a pattern, and a regulator, an insurer or an opponent will read it that way.
The practical direction is the one we set out in verify before you file. A named person opens and reads every authority in anything that leaves the firm, from the live law report or database rather than from the tool's output, and the file records that it happened. The CPS employs specialist prosecutors and still let two phantom cases travel through two documents. A busy practice with no written checking discipline is more exposed, not less. Read the judgment on the National Archives and notice how much of it concerns process rather than technology. Then give someone in your firm the job of answering what your equivalent of the 78-case review would look like, before an incident asks the question for you.
If you would rather find the gaps in your checking discipline before a judgment finds them for you, our free AI Readiness Check takes about two minutes: see where your firm stands.