We searched 9,000 UK trade mark decisions for AI. Here is what we found

Earlier this month the Intellectual Property Office used its newsletter to pass on a warning from two IPEC judges: do not rely on AI-generated information in legal claims, because AI tools have produced case citations that do not exist, and claims built on them have failed.

We wanted to know how often this has actually reached the UK trade mark tribunal. Not how often it is discussed, how often it has happened, in real decisions, with real consequences.

So we searched our own database.

What we searched, and what we did not

TMGuard holds a structured copy of the UK Intellectual Property Office's published tribunal decisions, the BL "O/" series: oppositions, invalidations, revocations, rectifications, examination appeals, and Appointed Person appeals. As of this search the database held 9,525 such decisions, 9,194 of them with machine-readable text, running from 1998 to late April 2026.

That is not every UK trade mark decision in existence. It does not include routine registry actions that never produce a published decision, and it does not include High Court, Court of Appeal or CJEU judgments, which appear in our data only when a tribunal quotes them. What it is, is a near-complete record of the IPO's own contested trade mark decisions over more than two decades, in a form we can search precisely.

We searched the full text of every decision for references to AI tools, generative models, hallucinated authorities, and fabricated citations, using word-boundary matching and proximity rules to keep ordinary language out (the word "invented" appears constantly in trade mark decisions, almost always about invented words as a measure of distinctiveness, not invented cases). We ran the same search across 2024 as a control. It returned no fabricated-authority cases, which told us the method was tight enough to trust.

Then we read every genuine hit in full, against the actual decision text, and sorted the real cases from the noise.

The headline

Across 2025 and 2026, eleven decisions in our database genuinely involve AI: five where fabricated or AI-generated material reached the tribunal in a serious way, and six more where AI was used or suspected in the conduct of a case.

Four of these have been written up by legal commentators. The other seven have not, as far as we can find. That is the value of searching a structured corpus rather than waiting for the cases that happen to get blogged: the full pattern is larger than the visible one.

Here is what the decisions actually say.

The first one: a trade mark appeal where both sides used ChatGPT

The case usually cited as the first is Pro Health Solutions Ltd v ProHealth Inc (BL O/0559/25), decided by the Appointed Person Phillip Johnson on 20 June 2025. It was an appeal in a passing-off opposition, and both sides had used ChatGPT.

The appellant, a litigant in person, filed grounds that listed real cases but attached quotes that, in three instances, were not in the judgments cited. He confirmed at the hearing that he had drafted the documents with the assistance of ChatGPT, and apologised. The Appointed Person noted that even setting the fabricated citations aside, the AI output had not actually helped him.

The respondent's representative, a trade mark attorney, did not escape scrutiny either. His three cases were real and correctly cited, but the propositions he attached to them were not supported by the judgments, and at the hearing he could not point to where the cases said what he claimed.

Johnson used the decision to set out, for the first time in a trade mark context, how the tribunal should treat AI-fabricated material, drawing on the Divisional Court's landmark ruling in R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin). And he suggested the IPO should consider giving litigants a clear warning about the risks of AI, for represented parties and litigants in person alike.

One note on this case, because it matters for a piece about accuracy. A line has circulated in some commentary suggesting Johnson said a professional cannot "abdicate responsibility to an algorithm." Those words do not appear in the decision. We checked the full text of every Phillip Johnson decision in our database; the phrase appears in none of them. It is a paraphrase that hardened into a quotation somewhere in the retelling, which is its own small lesson.

The pattern after ProHealth

ProHealth was not a one-off. Over the following six months the issue returned to the IPO repeatedly.

In Warwick Econometrics Ltd v University of Warwick (BL O/0938/25, 3 October 2025), a skeleton argument cited a case the hearing officer could not locate. Asked directly if AI or ChatGPT had produced the skeleton, the representative denied it with a simple "no." A later email attributed the bad citation to an automatically generated Google summary, which the hearing officer noted did not sit easily with the earlier denial. No admission of AI use was ever made, and no penalty was attached to the point. We mention the detail because it is a good example of how these situations actually unfold: not a confession, but an unlocatable citation and an explanation that raises more questions than it settles.

In Orthofix S.R.L. v OscarTech UK Ltd (BL O/1013/25, 30 October 2025), the applicant had cited two case references that turned out to be wrong: one pointed to a patent decision, the other to an unrelated case. At the hearing he explained that he had used an AI tool to help him prepare, and withdrew his reliance on the two references. The hearing officer used the moment to restate the risks of AI in legal research, citing Johnson's decision in ProHealth directly.

In Onyinye Udokporo v Enrich International Ltd (BL O/1141/25, 5 December 2025), Phillip Johnson again found fabricated references in grounds of appeal, this time two cases that simply do not exist, alongside quotes from real cases that were not in the judgments as quoted. The appellant admitted she had used AI to write submissions for an earlier part of the proceedings. The decision is a clear statement of the standard: material cited to the tribunal must be real, must be relevant, and any quotation relied on must be accurate.

Then it reached the evidence itself

The cases above are about fabricated or AI-drafted legal argument. By 2026, AI was also turning up in the evidence.

In a 2026 opposition concerning the mark Kersaall (BL O/0290/26, 31 March 2026), the opponent filed ChatGPT responses as evidence that the competing marks were confusingly similar. The hearing officer quoted the Ayinde warning at length and declined to give the AI material any weight, observing that it showed no actual marketplace confusion and simply represented hypothetical scenarios generated by artificial intelligence. The costs award was reduced partly because that evidence was not relevant. This is, as far as we can find, an unreported case, and it is a clean example of AI-generated evidence being rejected rather than AI-drafted argument.

We also found genuine AI use surfacing in other forms: an Appointed Person noting that grounds of appeal may have been drafted by generative AI and holding the quality unacceptable regardless (Black Swan Vodka, BL O/0820/25); a hearing officer observing that a counterstatement bore "the hallmarks of having been generated (at least in part) by AI" (Grayn, BL O/0746/25); a party submitting logos produced by ChatGPT as evidence on distinctiveness (The Hoxton, BL O/0908/25); and a hearing officer suspecting submissions had been AI-produced because of multiple incorrect citations, while declining to impose any sanction given the litigant's inexperience (Misfits Boxing, BL O/0010/26).

The case that runs the other way

The most instructive case of all is the one that does not fit the pattern.

In a 2026 opposition concerning the mark noosh (BL O/0302/26, 7 April 2026), one party argued that the other side's authorities were of no relevance and "may have been generated by artificial intelligence." The representative responded that they had been included on legal advice. The Appointed Person rejected the AI accusation: the authorities were genuine, at least two of them plainly relevant to the point at issue, and, in the decision's own words, "They are not hallucinations."

This is the inverse of every other case here. It is not AI fabricating law. It is a party wrongly accusing the other side of using AI to fabricate law. As AI hallucination becomes a known phenomenon, "this was generated by AI" becomes an accusation people reach for, and tribunals now have to judge those accusations too. Sometimes, as here, the accusation is the thing that is wrong.

We include it deliberately. A survey that only reported the cases supporting a single narrative would be doing, in miniature, exactly what this whole subject warns against.

What the genuine cases have in common

Read together, the eleven cases share a single mechanism. In each one, a person asked an AI tool for legal material, the tool produced something that looked right, and nobody checked it against a real source before putting it in front of the tribunal.

That is the failure mode worth understanding. A large language model trained on a vast amount of text has learned what a citation looks like: the names, the court, the year, the neutral citation, the confident summary. Asked to support an argument, it produces citation-shaped text. Sometimes that text points to a real case. Sometimes the case is real but the quote is invented, as in ProHealth. Sometimes the whole authority does not exist, as in Udokporo. The model has no way to tell the difference, because it is not connected to a verified source of law. It is generating something plausible, not retrieving something true.

The conclusion is not that AI has no place in trade mark work. It is narrower and more useful than that: AI without a grounded, checkable source has no place in a legal claim.

Why our own search is the counter-example

There is a category difference between AI applied to the open internet and AI applied to a known, structured body of source documents. The search behind this article is itself an illustration of the second kind.

Every case named here was found by searching the actual text of real decisions, then verified by reading those decisions, with the paragraph numbers checked against the source. When we could not confirm something, the "abdicate responsibility to an algorithm" line, we left it out and said why. Nothing here was generated; it was extracted and checked.

That is the same discipline behind the platform itself. When TMGuard structures an opposition decision, it works from the decision as published by the IPO, taken directly from the register. The text is processed into structured fields, and those fields are measured against a manually-reviewed gold standard before anyone relies on them. If a decision does not exist, it cannot appear, because the system is not inventing decisions. It is reading real ones.

The IPO was right to put the IPEC warning in front of nearly 20,000 readers. Our own data shows why it matters: in the space of a year, the tribunal has had to deal with fabricated citations, AI-drafted argument, AI-generated evidence, and even a wrongful accusation of AI use. The common thread is not the technology. It is the missing step between what an AI says and what is actually true. Close that gap, and AI is a useful tool. Leave it open, and it is the thing that gets your case struck out.

The eleven cases

For reference, the AI-relevant UK trade mark decisions found in our database, 2025 to 2026:

Pro Health Solutions Ltd v ProHealth Inc (BL O/0559/25, 20 June 2025, Appointed Person Phillip Johnson): fabricated quotes attached to real cases; ChatGPT-assisted drafting.

Grayn (BL O/0746/25, 11 August 2025, Mrs J Pike): counter-statement appearing to be AI-generated.

Black Swan Vodka (BL O/0820/25, 8 September 2025, Appointed Person Phillip Johnson): grounds of appeal possibly AI-drafted, held unacceptable regardless.

happygum (BL O/0836/25, 11 September 2025, Ms L Bailey): a conversation with a website AI bot submitted as evidence.

The Hoxton (BL O/0908/25, 29 September 2025, Ms C Boucher): ChatGPT-produced example logos submitted as evidence on distinctiveness.

Warwick Econometrics Ltd v University of Warwick (BL O/0938/25, 3 October 2025, N Barratt): an unlocatable citation; AI use asked about and denied.

Orthofix S.R.L. v OscarTech UK Ltd (BL O/1013/25, 30 October 2025, Ms Rhea Morris): AI tool used in preparation; two incorrect case references withdrawn.

Onyinye Udokporo v Enrich International Ltd (BL O/1141/25, 5 December 2025, Appointed Person Phillip Johnson): two non-existent cases cited; admitted AI use in earlier submissions.

Misfits Boxing (BL O/0010/26, 8 January 2026, Teresa Pinto): incorrect citations suspected AI-produced; no sanction applied.

Kersaall (BL O/0290/26, 31 March 2026, Ms S Wallace): ChatGPT responses filed as evidence and given no weight.

noosh (BL O/0302/26, 7 April 2026, Appointed Person Antony Craggs): an allegation that authorities were AI-generated, rejected by the tribunal.

A note on method

This article is based on a search of TMGuard's database of UK IPO trade mark decisions, conducted on 29 May 2026. Higher-court authorities mentioned (Ayinde [2025] EWHC 1383 (Admin); and, in the wider debate, Harber v HMRC [2023] UKFTT 1007 (TC)) are referenced from the public record. Each UK trade mark decision cited has been read in full and its paragraph references checked against the source text. Anyone is welcome to verify the BL-series decisions on the IPO's decisions database.