Most AI hallucination cases in the legal system involve professionals, lawyers who should know better and get sanctioned when they do not. Omar Rafique v HMRC is a useful variation, because Rafique was representing himself, and his story shows how the same tool that gets attorneys fined can quietly cost an ordinary person their entire case.

A five-figure tax fight

Rafique, a former director of a venue in Worthing, was contesting a tax demand from HM Revenue and Customs of roughly £51,200. He brought his VAT appeal to the First-tier Tribunal (Tax Chamber), the UK forum that hears tax disputes. So far this is a normal, if stressful, situation: a person who believes the tax authority has it wrong, asking an independent tribunal to take a look.

Then came the submissions. Rafique repeatedly filed correspondence that cited legal authorities to support his position. The problem was that the authorities were, in the tribunal's assessment, illusory. Some references were simply wrong. Others pointed to cases that did not exist. And the ones that did correspond to real cases did not actually support the propositions he was citing them for. This is the full hallucination spread: invented cases, misattributed cases, and real cases bent to say things they never said.

The tell

The detail that makes this one land is the residue. Some of Rafique's filings still contained the unfilled placeholders that a language model leaves when it spits out a template the user never customized. Phrases like "[Name of Agent/Third Party]" and instructions to "insert date if known" survived into documents submitted to a tribunal. That is the legal-filing equivalent of sending an email that still says "Dear [FIRST NAME]." It is hard to argue you carefully reviewed your authorities when the document itself is still wearing the assembly instructions.

Those brackets are a small thing and a damning thing at once. They are proof not of using AI, which is allowed, but of not reading the output, which is the entire failure. A person who opened the document and read it to the end would have seen "insert date if known" sitting in their own submission and felt the floor drop. Rafique apparently did not, or did and filed anyway.

What the tribunal did

The appeal had already been struck out for non-compliance, and the question before the tribunal was whether to reinstate it. UK tribunals do not decide reinstatement on vibes; they apply a structured test from a case called Martland, which weighs the seriousness of the default, the reasons for it, and all the circumstances, including the need for litigation to be conducted efficiently and at proportionate cost.

Run through that framework, Rafique's position did not survive. The tribunal confirmed the strike-out as the consequence of his non-compliance and declined to reinstate the appeal. It also did the thing these decisions increasingly do: it issued a warning, on the record, that fabricated authorities would not be tolerated going forward. The roughly £51,200 demand he was fighting was left standing. He did not lose on the merits of his tax argument; he lost the chance to make the argument at all, because the vehicle he was using to make it kept arriving full of fake law.

The tribunal's patience was clearly thinning, and it is not alone. UK courts and tribunals have now worked through enough of these episodes that fabricated AI authorities are treated as a known and unacceptable hazard rather than the honest stumble of someone new to litigation. A growing public database tracks AI hallucination cases across jurisdictions, and the number climbs every month, which means each new litigant who leans on a chatbot is walking into a system that has already seen the trick and lost its appetite for it.

A trap for the self-represented

There is a sympathetic reading here, and it is worth stating plainly. A self-represented litigant fighting a five-figure tax bill is exactly the kind of person who turns to a free, confident, always-available chatbot for help, because the alternative is paying for a lawyer they may not be able to afford. The tool feels like a great equalizer. It produces fluent, lawyerly prose and citations that look real. To someone without legal training, there is no obvious way to tell a genuine authority from a fabricated one; they both look like a case name, some numbers, and a year.

That is the trap. The tool is most appealing to the people least equipped to catch its mistakes, and the tribunal's patience for "I did not know it could do that" is shrinking as the warnings accumulate. UK courts and tribunals have now seen enough of these to treat fabricated AI citations as a known hazard rather than a novel surprise, which means the grace period is closing for litigants too, not just for lawyers.

The lesson, even for non-lawyers

The uncomfortable takeaway is that a language model will help you lose a case you might have had. It does not know which cases are real. It generates citations because citations are predictable text, and a fabricated one is often more confident-looking than the real, messy, partially-on-point authority you actually needed. If you cannot verify what it produces, by finding the actual case and reading it, you are not strengthening your argument; you are smuggling a defect into your own filing and signing your name under it. Rafique signed his, brackets and all, and the tribunal closed the door.