A family court judge has publicly named a lay advocate who submitted four entirely fabricated case citations generated by AI. The ruling is a stark warning for anyone using AI tools for legal research without verification.
On 5 March 2026, Recorder Howard delivered judgment in Re A, B, C, D [2026] EWFC 71 (B) at Bournemouth Family Court. The case concerned children welfare proceedings — serious enough on their own terms. But the judgment became notable for something else entirely: the conduct of Layla Parsons, an unregistered barrister acting as a lay advocate for one of the parties, who had filed a skeleton argument containing four case citations that did not exist.
The citations were not real cases that had been mischaracterised or incorrectly referenced. They were complete fabrications — cases that had never been heard, in courts that had never sat on the matters described, producing judgments that were never written. They were, in the judge's words, “another example where AI hallucinations have led to the court being misled.”
What Happened
Parsons had prepared a skeleton argument for the hearing that included citations to four authorities. When the court attempted to verify these references, none of them could be located. The case names, neutral citation numbers, and legal propositions attributed to them were entirely fictitious — bearing all the hallmarks of content generated by a large language model that had been asked for supporting case law and had obliged by inventing it.
Parsons was not a registered practising barrister. She was acting as a lay advocate — a person who assists a litigant in person but does not have rights of audience or the regulatory obligations that come with being a practising member of the Bar. Nevertheless, she reported herself to the Bar Standards Board after the issue came to light.
Recorder Howard took the unusual step of naming Parsons in the judgment. In family proceedings, where anonymity protections are strong and the court rarely identifies participants beyond the parties themselves, this was a deliberate and pointed decision. The judge considered it necessary to mark the seriousness of what had occurred: fabricated authorities had been placed before a court dealing with the welfare of children.
Why This Matters Beyond the Individual Case
The Parsons case is not the first instance of AI-generated citations appearing in court proceedings. In the United States, a New York lawyer was sanctioned in 2023 after submitting a brief containing six fictitious cases generated by ChatGPT in Mata v Avianca. In Canada, a litigant submitted AI-fabricated authorities in a British Columbia case the same year. But the Parsons case is among the first reported instances in the English and Welsh courts, and it arrives at a moment when the legal profession is actively debating how AI should be integrated into legal practice.
The timing matters. The Judicial Office, the Solicitors Regulation Authority, and the Bar Standards Board have all published or are developing guidance on the use of AI in legal proceedings. The common thread across all of this guidance is that practitioners bear personal responsibility for every citation and every legal proposition they put before the court, regardless of which tool generated it.
What makes AI hallucination particularly dangerous in the legal context is its plausibility. A fabricated citation from an AI tool does not look like a mistake. It looks exactly like a real citation: correct formatting, a plausible court name, a realistic neutral citation number, and a legal proposition that sounds like something a court might say. Without checking the original source, there is no way to distinguish a hallucinated citation from a genuine one. The error is invisible until someone looks for the case and finds it does not exist.
The Verification Problem
The core issue is not that AI tools sometimes produce errors — every research tool has limitations. The issue is the nature of the error. When a traditional legal database returns no results for a search, you know you haven't found what you're looking for. When an AI tool generates a response, it always gives you something. It is architecturally incapable of returning “I don't know” in the way a database search returns zero results. Instead, it generates the most plausible response it can construct from its training data, and when that training data doesn't contain a real answer, the response is fabricated with the same confidence and formatting as a genuine one.
This creates an asymmetric verification burden. Every AI-generated citation must be independently verified against an original source — not just checked for existence (does the case number return results?) but checked for accuracy (does the case actually say what the AI claims it says?). The second type of error — citing a real case for a proposition it doesn't support — is arguably more dangerous than outright fabrication, because the case exists and a surface-level check will not catch the mistake.
A Stanford study published in 2025 found that even purpose-built legal AI tools from Westlaw and LexisNexis hallucinated in 17–33% of queries. These weren't general-purpose chatbots — they were AI features integrated into established legal research platforms with access to comprehensive case databases. If professional legal AI tools backed by billions of pounds of investment hallucinate at those rates, the risk from using general-purpose AI tools like ChatGPT for citation generation is substantially higher.
What Practitioners Should Take From This
The lesson from the Parsons case is straightforward, even if its implications are uncomfortable for a profession that is increasingly enthusiastic about AI adoption.
Look up the real authorities on professional conduct and AI — with verified citations from official databases.
Never submit AI-generated citations without verifying them against original sources. This means finding the actual judgment — on The National Archives' Find Case Law service, on BAILII, or through a verified legal database — and confirming that the case exists, that the citation is correct, and that the judgment supports the proposition for which it is being cited.
The duty to the court has not changed. Whether you are a solicitor, barrister, or lay advocate, you have an obligation not to mislead the court. The SRA Code of Conduct, the BSB Handbook, and the overriding objective in the CPR all require that representations to the court are accurate. “The AI told me” is not a defence, and the Parsons case demonstrates that judges are prepared to take public action when this obligation is breached.
The choice of research tool is now a risk management decision. There is a material difference between AI tools that generate citations (and may fabricate them) and research platforms that retrieve citations from official databases (and can verify them against original sources). Practitioners should understand which category their tools fall into and adjust their verification workflows accordingly.
A Different Approach to AI Legal Research
Search the Law was built specifically to address the hallucination problem. Rather than asking AI to generate citations, the platform searches 21 official UK legal databases simultaneously and retrieves case law directly from The National Archives, legislation.gov.uk, Hansard, and other government sources. Every citation links to the original published judgment. The AI layer is used for query understanding, research structuring, and citation network analysis — not for generating citations.
The platform tracks over 191,500 citation pairs, classifying how each authority has been treated by subsequent courts: applied, considered, distinguished, doubted, or overruled. This citation network is built from the actual text of judgments, not generated by a language model. When the platform tells you a case has been distinguished, it's because a subsequent judgment used that language — and you can follow the link to read the passage yourself.
This approach cannot fabricate a case that doesn't exist. It cannot attribute a holding to a judgment that doesn't contain it. The error modes are structurally different from generative AI, and the verification path is transparent: every citation can be checked against its official source in one click.
Re A, B, C, D [2026] EWFC 71 (B) was handed down on 5 March 2026. The full judgment is available through Find Case Law. Search the Law provides free access to 129,000+ UK judgments with verified citation networks at searchthe.law.