Generative AI has reached the Employment Tribunal, mostly in the hands of the people least able to check what it produces. Published judgments now regularly record that a claimant drafted a witness statement, a skeleton argument or an application with ChatGPT or a similar tool. The picture they give is consistent: tribunals have broadly accepted that parties, especially litigants in person, will use AI. What they do not accept is unchecked output. This article looks at where tribunals draw that line, and at the ways AI-produced material goes wrong.
The scale
A review of the published Employment Tribunal judgments found at least 84 in which a party’s use of AI in bringing or conducting a claim was admitted or found by the tribunal.
| Year of publication | Judgments |
|---|---|
| 2023 | 1 |
| 2024 | 7 |
| 2025 | 15 |
| 2026 (to 1 November) | 61 |
Many judgments are published a year or more after the claim was presented, and AI use usually shows up in hearing papers rather than the claim form, so most of these judgments concern claims brought in 2023 to 2025. The table shows the direction rather than a rate.
Most users are claimants (70 of the 84), and most are unrepresented (71 of the 84). Representatives and respondents appear too: nine cases involve a claimant’s representative, and four involve a respondent or its representative.
Usually the party says so. In Wright v Extraction Solutions Ltd (ET 6031780/2025), a supplementary witness statement was in a “substantially different style” from the claimant’s other evidence, and he explained at the hearing that he had used AI to produce it (paragraph 8). Where AI use is suspected but not established, it is usually the tribunal that raises it.
Use is accepted
Several judgments say that using AI is not a problem in itself. In Ferreira v Magic Life Ltd (ET 3304596/2025), the judge said “In principle there is no objection to the use of AI by litigants” (paragraph 32). In Wandji Mboungueng v Severn Trent Water Ltd (ET 1306955/2023), “Using AI is not in and of itself a problem”; the problem is putting unchecked, untrue material before a tribunal (paragraph 22). In Okopochini v London and Quadrant Housing Trust (ET 2413147/2023), answering the respondent’s complaint that the claimant had over-relied on AI, the tribunal drew “no negative perception” and called AI a tool that can be useful (reasons, paragraph 50).
Tribunals have also recognised why unrepresented claimants turn to it. As early as Warrener-Iglesias v Coretech Security Services Ltd (ET 6000353/2023), the judge said the claimant should not be penalised for using ChatGPT to put his own evidence into more formal language, which a litigant in person could readily be understood to want, though the parts that were plainly ChatGPT’s own platitudes were given less weight (paragraphs 52 and 53). In Mallon v Energy Systems Catapult Ltd (ET 1306502/2020), the tribunal accepted that the claimant used AI in connection with his disabilities and found that use not unreasonable in itself (costs reasons, paragraph 56). In Kuzniar v General Dental Council (ET 6009997/2024), costs were refused partly because AI is a new tool the public is still getting used to (paragraph 46).
AI-assisted evidence is not automatically discounted either. In Morris v Mitchells and Butlers Retail Ltd (ET 1303431/2025), using AI to tighten up a disability impact statement did not of itself make the statement unreliable or less weighty (paragraph 14), and in Samuels v The First Element Group Ltd (ET 6004283/2025) the judge drew no adverse inference and gave an AI-assisted impact statement no less weight than if it had been produced without AI (paragraph 15).
Where acceptance stops
A helpful discussion of the limits is in Hassan v ABC International Bank plc (ET 6005897/2024), a reconsideration and costs judgment of 13 October 2025, which reviews the judicial guidance and the main authorities on false citations. The judge accepted that a litigant in person is entitled to some latitude, though it only goes so far, and declined to approach the claimant’s use of AI “as in any way inherently negative” (Annex A, paragraphs 39 and 41). Under the Guidance for Judicial Office Holders on Artificial Intelligence (14 April 2025), a judge may ask whether AI was used and what checks were made, and may remind a litigant that they are responsible for what they put forward (Annex A, paragraph 42). Drawing on Olsen v Finansiel Stabilitet A/S [2025] EWHC 42 (KB), the judge distinguished reckless disregard for whether an authority is genuine from knowing that it is false (Annex A, paragraph 48).
Hassan still ended in a costs order of £5,881.80 for unreasonable conduct: the claimant served a skeleton argument whose AI-derived citations were inaccurate, misleading or fictitious after he knew of the risk, and persisted in relying on them. The use of AI per se was not criticised; relying on citations that had not been checked, once the risk was known, was.
Where the output fails
The judgments show four recurring ways in which AI-produced material goes wrong.
Law that is invented or misquoted
This is the best-known failure. In Chanda v Royal Mail Group Ltd (ET 3311062/2023), the tribunal could not find the propositions of law in the cases cited, nor one of the cases, and said these “may have been generated by AI” (paragraph 12). The claimant’s lay representative denied using AI. The tribunal treated the false citations as unreasonable conduct “whether generated using AI or not”, but still refused to strike the claim out (paragraphs 13, 14 and 24 to 26).
The same pattern appears in judgments that never mention AI. In Sutton v Health Transportation Group Ltd (ET 1401924/2024), the judge said that two citations given for a Court of Appeal decision “are fabricated; neither exists” (paragraph 42), and said the claimants would be required, by a separate order, to explain in a signed witness statement who prepared the application (paragraph 43). In Nwogu v Royal Berkshire NHS Foundation Trust (ET 3303162/2020), authorities “with those generic names and unusual citations could not be found” in the legal reference materials available to the tribunal. In Pettitt v Elmdene International Ltd (ET 6005124/2024), the claimant’s purported authorities either did not exist or were not authority for the proposition claimed, and this was one of the reasons for a costs order (paragraph 32). None of these judgments makes a link to AI, but the absence of the word “AI” in a judgment does not mean AI was not involved.
Claims and arguments that don’t work
The less visible risk is that AI points a party in the wrong direction. Sometimes the law is simply wrong. In X v Y (ET 2413027/2023), the claimant relied on a Google extract, apparently reflecting his own AI research, which said that losses above the statutory cap could be ordered where fraudulent delay or gross misconduct caused them; the judge said this was “simply incorrect as a matter of legal principle” (paragraph 24 of the first reconsideration judgment). In Chandler v Chief Constable of Kent Police (ET 2306189/2023), the claimant followed AI advice from the internet and missed the one-month time limit after early conciliation; the judge noted the “danger when using AI that the full picture is not revealed”, time was not extended and the claim was struck out (paragraphs 25 to 27).
Sometimes AI produces claims the facts do not support. In Coll v Office of Gas and Electricity Markets (ET 8002269/2025), the tribunal found that “reliance on AI led the claimant to assert claims that were clearly not supported” by the evidence, including indirect discrimination (paragraph 99). In Heath v NHS England (ET 2304017/2023), the claimant said he had used AI to create his list of issues and expected to use only about 10% of it (paragraph 42); the judge thought many of the legal classifications in it misconceived (paragraph 41) and said “Reliance on AI seems to have done the Claimant more harm than good” (paragraph 52), and adopted the respondent’s list, adding seven of the claimant’s grievances (paragraphs 53 and 54). In Chiu v Axiom Ince Ltd (in administration) (ET 3312343/2023), an AI-assisted schedule of loss claimed losses that did not fit the claims brought, and the judge observed that “AI has its limitations” (paragraph 12).
AI can also mislead about evidence. In Massingham v The Haulage (Holdings) Organisation Ltd (ET 6005961/2024), the claimant used an AI chat to interpret a vehicle test log and relied on its reading that the truck was faulty; the document did not show that, and the tribunal gave the AI reading no weight (paragraph 100).
Words the party did not mean
AI can put words in a party’s mouth. When a claimant withdraws a claim, the tribunal will normally also dismiss it, and a dismissed claim cannot be brought again. That is the usual wording when a claim is settled, but it is not always what the claimant intends. In Sullivan v Capita plc (ET 6039320/2025), the claimant used AI to word her applications to withdraw her claims, and the wording suggested that it might be appropriate to dismiss them, the implications of which she had not appreciated (paragraphs 25 and 28). The withdrawal was effective and ended the proceedings (paragraphs 15 to 24). But the tribunal found her real intention in her request to keep the right to bring a further claim, and declined to issue a judgment dismissing the claims under rule 51 (paragraphs 28 to 32). Had it done so, she could not have pursued them again.
AI can also leave a party pursuing a case they do not understand. In Wright v SFE Chetwode Ltd and Winter (ET 2601525/2024), the judge contrasted the claimant’s ChatGPT-written statement and submissions with her oral arguments, and was concerned that she was “pursuing a claim she does not understand and cannot personally justify” (paragraphs 6 to 8). In Murphy v Secretary of State for Work and Pensions (ET 2400194/2024), the claimant was “firing off AI generated documents and applications which do not assist him”, and deposit orders followed, with a recommendation to take legal advice rather than rely on AI (paragraphs 11, 22, 49, 117 and 130).
Length and repetition
A frequent complaint is not only invented law but length. In 24 of the 105 judgments (the 84 above, plus 21 where the tribunal suspected AI use without making a finding), the tribunal or the other side complained that the material was long, repetitive, irrelevant or hard to follow. In almost all of them it was the tribunal that raised it. In Nilsen v John Lewis Partnership plc (ET 6040292/2025), the judge described an application as “repetitive and unclear, displaying characteristic use of AI”. In Mallon v Gateley Capitus Ltd (ET 1303617/2021), the tribunal said a number of the claimant’s emails were “lengthy and difficult to digest”, and that much of his AI-generated correspondence tended “to confuse rather than clarify” (costs reasons, paragraph 11). The costs order in that case rested on other findings (paragraphs 43 and 44), though the tribunal said its impression that he had put little effort into the claim was supported by his tendency “to rely overly on generic or AI generated material” (paragraph 46). And in Mallon v Energy Systems Catapult Ltd, the same judgment that accepted the claimant’s reasons for using AI found that his AI-driven emails raised irrelevant points, and told him to read and refine AI output before sending it (costs reasons, paragraphs 57 and 58).
AI transcripts of recordings raise a related problem of accuracy. In Smith v Addingtons Associates Ltd (ET 3314144/2023), the AI-generated transcript “made no sense in parts”, and with both parties’ consent an agreed transcript was ordered instead (paragraph 11). In Williams-St Louis v West Berkshire District Council (ET 6015012/2024), the claimant was refused permission to rely on AI transcripts of covert recordings, mainly because they ran to many hours and checking them would have meant postponing the final hearing; they had also not been mentioned at the disclosure stage (paragraphs 10 to 12).
What follows
In most of the judgments (69 of the 84), the AI use is recorded without anything following from it. Where tribunals do act, they act on the quality of what has been put forward rather than on the tool: disregarding the material, giving it less weight or treating it with caution, warning the party, or, as in Hassan, ordering costs. In one case the judge criticised a respondent’s solicitor for citing an authority that could not be found.
It can count against a party on time limits too. In Shah v Manchester Airport plc (ET 2400509/2024), the tribunal treated the claimant’s use of AI in the proceedings, which had helped him, as showing that he could have researched his rights at the time (paragraph 86).
Conclusion
AI use is now a regular feature of tribunal litigation, particularly for litigants in person, and it is becoming more common quickly. The published judgments do not show tribunals treating it as misconduct in itself. Several judgments say expressly that using it is acceptable, and some recognise that it helps unrepresented claimants present their case. What draws comment is the quality of the material: authorities that cannot be found, quotations that are not in the judgments cited, claims and arguments the law or the facts do not support, wording that does not say what the party meant, and submissions that are long or repetitive.
For anyone advising a claimant who is using AI, that is the point to press. Every authority needs checking against the judgment itself. Every claim needs testing against the facts. And anything AI drafts on the claimant’s behalf needs reading against what the claimant actually wants to say. The tool is not the problem; unchecked output is.