Ndaryiyumvire v Birmingham City University: a County Court wasted-costs order after AI-fabricated case citations from legal software

The problem

Generative AI — including the “research” features now built into mainstream legal case-management software — can produce fluent citations to cases that do not exist. When a lawyer files that output without checking it, false authorities enter the court record. This case is one of the first in England to test what happens when the fabrication comes not from a public chatbot but from a firm’s own trusted software. As His Honour Judge Charman recorded, the firm made “an error in relying on the LEAP software and it appears, failing to appreciate that the so called AI feature or legal research feature was in fact in substance a large language model”.

What was built

There is no deployment to celebrate here — this is an AI-governance honest-negative. Gloriose Ndaryiyumvire sued Birmingham City University; on 10 July 2025 her solicitor Raphael Newton’s firm, Gordon & Thompson, filed an application to amend her particulars of claim that referred to two authorities the University’s solicitors could not locate. When the matter came before District Judge Bradshaw and it was pointed out that the cited cases did not exist, the application to amend was dismissed and the claim struck out. As the judgment records, “The fictitious authorities cited were a case referred to as Qureshi v Qureshi [1998] 1 WLR 174 and a case said to be Z Ltd v A Ltd [2011] EWCA Civ 110” — and “these cases are fake cases; the cases cited simply do not exist.” On Mr Newton’s evidence, the document had been “generated by LEAP legal software which includes a built-in research function that can automatically suggest case law”, and was a draft “work in progress” that administrative staff filed in error using his initials, without his review.

The outcome

HHJ Charman found the conduct “improper and unreasonable and negligent” and held that “the failures in administration mean that in this case in my judgment it is appropriate to make a wasted costs order against Gordon & Thompson Solicitors.” He was careful to note the order’s nature and to reserve the sum: “a wasted costs order is compensatory, it is not penal … and I will hear submissions as to the amount of those costs.” No monetary figure was fixed in this judgment — the amount was left to further submissions.

Applying the leading case Ayinde v London Borough of Haringey, the judge held that placing false citations before a court would “usually be appropriate for the court to refer the matter to the regulator”. But he found that “it was not Mr Newton that placed false material before the court, it was administrative staff at his firm”, and that “the failure in this case was in substance a failure of management at the firm more than the failure of Mr Newton as an individual solicitor.” Because a wasted costs order is itself automatically referred to the Solicitors Regulation Authority, he declined to make a further separate referral of Mr Newton, and held it would be “wholly disproportionate” to consider contempt proceedings or a police referral, this case being “not at the more serious end”. He directed that the transcript be prepared at public expense and published.

Two independent legal newsrooms reported the outcome firsthand. Legal Futures wrote that “given the ‘improper and unreasonable and negligent conduct and the failures in administration’, it was appropriate to make a wasted costs order against Gordon & Thompson.” The Law Gazette named both fake cases — “The fictitious cases were Qureshi v Qureshi and Z Ltd v A Ltd, neither of which existed” — and set out the regulatory position: “A referral to the SRA was made by the court as a result of the wasted costs order (pursuant to s51(7A) Senior Courts Act 1981).”

On the AI attribution — the honest limit: the court did not name ChatGPT or any external chatbot. It accepted the solicitor’s account that the citations came from the built-in research function of LEAP case-management software, and characterised that feature as “in substance a large language model”. The judge added that “it was not the use of LEAP per se that was the problem, but what was done with the product of LEAP.”

Verification status: PENDING — checker-verified at confidence 1.000, awaiting HUMAN client-confirmation. Not verified (green is human-only).

Path to green

The court-stated facts need no client confirmation: the origin is a published, approved transcript of a County Court judgment, byte-tied to a contemporaneous organic Wayback capture (local PDF sha1-b32 XJNTCFGABC4JL2AO5L774KTMNOJK44VG = CDX digest of captures 20251119113146 / 20260118134746). Two deliberate limits remain for a human/checker: (1) no monetary figure is claimed because the amount was reserved — any later quantum order, if published, would be a separate document; do not infer a sum; and (2) confirming the outcome of the automatic SRA referral and that no appeal disturbed the wasted-costs order — a post-judgment currency check, not a gate on the court-ordered facts. No green badge is sought: this is a court-origin honest-negative and verified is human-only.