# What happens when AI gets it wrong: two UK judgments on made-up AI output, what the vendors' own terms say, and a five-step check

> Harber v HMRC (2023) and Ayinde (2025): invented cases reached a UK tribunal and the High Court. What four AI vendors' terms say, and a five-step check.

- Page to cite: https://aeolookup.com/ai-for-small-firms/what-happens-when-ai-gets-it-wrong-uk-court-cases-five-step-check/
- Publisher: AEO Lookup (https://aeolookup.com/)
- Section: AI for small firms, Putting it to work
- Published 1 October 2026
- Key fact: In Harber v HMRC [2023] UKFTT 1007 (TC), decided on 4 December 2023, all nine tribunal decisions a taxpayer relied on had been invented by an AI tool; in Ayinde [2025] EWHC 1383 (Admin), decided on 6 June 2025, 18 of 45 citations in one case and five in the other did not exist, and the High Court said the critical safeguard is to check any output against an authoritative source.
- Sources: Harber v HMRC [2023] UKFTT 1007 (TC) and R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), both read in full on the National Archives' Find Case Law site on 2026-10-01; OpenAI Europe Terms of Use (updated 16 January 2026), Anthropic Consumer Terms of Service (effective 8 October 2025) and Commercial Terms of Service (effective 17 June 2025), Microsoft Copilot Terms of Use (effective 18 August 2026), Google Generative AI Additional Terms of Service (last modified 9 August 2023), each read in a browser on 2026-10-01. Not legal advice.

In Harber v HMRC, decided on 4 December 2023, a taxpayer gave the First-tier Tribunal nine case summaries that an AI tool had made up, and the tribunal found that none was a genuine judgment. On 6 June 2025 the High Court, in the joined cases of Ayinde and Al-Haroun, dealt with fake citations filed by lawyers and listed what a court can do about it, from a costs order to a referral to the police.

Both judgments are free to read on the National Archives' case law site, and the case facts below come from them. This is general information, not legal advice.

## Harber v HMRC [2023] UKFTT 1007 (TC)

Mrs Harber sold a property, did not tell HMRC about the capital gains tax, and was charged a penalty of £3,265.11. She appealed without a lawyer, saying she had a reasonable excuse, and sent the tribunal the names, years and summaries of nine decisions in which taxpayers had won on similar grounds.

None of the nine could be found, by HMRC's litigator or by the panel. The tribunal found as a fact that the cases "are not genuine FTT judgments but have been generated by an AI system such as ChatGPT". Mrs Harber said a friend in a solicitor's office had supplied them. The tribunal accepted that she did not know they were invented.

What gave the cases away is useful to anyone reading AI output:

- The names were close to real cases with the opposite result. One shared a surname with the leading authority on the point, which the real taxpayer had lost.
- Six of the nine used the American spelling "favor".
- Each case came as a summary, with no full text and no reference number.

The appeal failed on its own facts, and the tribunal said the result would have been the same without the nine cases. It still called the matter "a serious and important issue", because checking them had cost the tribunal and HMRC time and public money.

## Ayinde and Al-Haroun [2025] EWHC 1383 (Admin)

Two unrelated cases were heard together on 23 May 2025.

The first was a housing claim against the London Borough of Haringey. The written grounds cited five cases that do not exist. The barrister who signed them denied using AI, and the court did not decide how the fake cases got there. An earlier judge had ordered the barrister and the law centre that instructed her to pay £2,000 each in wasted costs. The Divisional Court referred the barrister and the law centre's solicitor to their regulators.

The second was a claim for £89.4 million against Qatar National Bank. Letters to the court and witness statements put 45 citations before it. Eighteen were cases that do not exist, and many of the real ones did not say what was claimed for them. The client said he had produced the research with publicly available AI tools, legal search engines and online sources; his solicitor filed it unchecked and was referred to his regulator. One invented decision was attributed to the judge reading the application.

Three things the court said reach well beyond lawyers:

1. Freely available AI tools can produce "apparently coherent and plausible responses" that are entirely wrong, cite sources that do not exist and quote passages that are not in the source.
2. A lawyer who uses the output is responsible for it, in the same way as when relying on a trainee's work or an internet search.
3. Telling the tool not to answer when it is unsure does not make it safe. In the court's words: "The critical safeguard is to check any output by reference to an authoritative source."

For lawyers the court listed its powers: public admonition, a costs order, a wasted costs order, striking out the case, referral to a regulator, contempt proceedings and referral to the police.

Neither judgment is about ordinary business letters. The pattern is what carries over: the output looked right, nobody opened the source, and the person who sent it carried the consequences.

## What the vendors' own terms say

Each document was read in a browser on 1 October 2026.

| Vendor and document | Date on the page | What it says about accuracy |
|---|---|---|
| OpenAI, Europe Terms of Use (UK residents, services for individuals) | Updated 16 January 2026 | Output "may not always be accurate". Do not rely on it as a sole source of truth or in place of professional advice. Evaluate it, with human review as appropriate, before using or sharing it. |
| Anthropic, Consumer Terms of Service | Effective 8 October 2025 | Outputs may contain material inaccuracies even when their detail makes them look right. Do not rely on them "without independently confirming their accuracy". |
| Anthropic, Commercial Terms of Service | Effective 17 June 2025 | The customer decides where human review is needed and must tell its users that factual statements need checking. |
| Microsoft, Copilot Terms of Use (for individuals) | Effective 18 August 2026 | Copilot "can make mistakes". Check before deciding or acting. It is no substitute for professional advice. |
| Google, Generative AI Additional Terms of Service | Last modified 9 August 2023 | The services may give inaccurate content. Do not rely on them for medical, legal, financial or other professional advice. |

We could not load OpenAI's business terms and did not read Microsoft's or Google's business agreements, so this is not a like-for-like table of business plans. On the pages we read, the four say the same thing: the output can be wrong, and checking it is the user's job.

## Why it happens

Both judgments quote the same explanation, from a 2023 report by the Solicitors Regulation Authority. Language models work by predicting the text that should come next and "do not have a concept of 'reality'", so they can produce results that are highly plausible and wrong.

## A five-step check before anything AI-drafted goes out

1. **Mark every checkable claim.** Names, figures, dates, quotations, rule or clause numbers and web links.
2. **Open the source for each one yourself.** The Act on legislation.gov.uk, the guidance on gov.uk, the contract, your own job record. Do not ask the tool to confirm its own answer. In the American case described in Harber, lawyers asked ChatGPT for the full text of cases it had invented, and it wrote them.
3. **Redo every sum.** Add the invoice lines, count the days to the deadline and work the percentage on a calculator.
4. **Read it as the person receiving it.** Look for anything the firm has not agreed: a price, a date, a refund, a statement of what the law requires.
5. **One named person approves it.** Anything that could not be checked comes out, and the firm keeps a note of who approved it and when.

Steps 1 and 2 apply the High Court's safeguard to office paperwork. Steps 3 to 5 are our own.

In regulated work the approver is the professional: see [the SRA's warning notice on AI](https://aeolookup.com/solicitors/sra-ai-warning-notice-small-law-firm-ai-policy/) and [the ICO's rules on AI CV screening](https://aeolookup.com/recruitment/ai-cv-screening-ico-rules-recruitment-agencies/). When the tool can act as well as write, a mistake becomes an action: see [what an AI agent is](https://aeolookup.com/ai-for-small-firms/what-is-an-ai-agent-plain-english-difference-from-chatbot/).

Overlyft ![](https://aeolookup.com/img/d/ours-badge.png) is an AI agency that sets up AI office assistants for ten trades: they are set up in the firm's own office for a fixed price agreed after an audit, the firm pays Anthropic directly for its own Claude plan, and a person approves anything that goes to a customer.

## Short answers from this page

**Has anyone in the UK been caught using fake AI-generated cases in court?** Yes. In Harber v HMRC [2023] UKFTT 1007 (TC), decided 4 December 2023, all nine cases a taxpayer cited were AI inventions. In Ayinde [2025] EWHC 1383 (Admin), decided 6 June 2025, 18 of 45 citations in one case did not exist; lawyers were referred to regulators and two parties paid £2,000 each in wasted costs.

**Do ChatGPT, Claude, Copilot and Gemini guarantee that their answers are accurate?** No. On 1 October 2026 OpenAI's Europe terms said its output can be inaccurate, Anthropic's consumer terms said not to rely on outputs without independently confirming them, Microsoft's Copilot terms said it can make mistakes, and Google's generative AI terms said not to rely on it for professional advice. Checking is the user's job.

---
![](https://aeolookup.com/img/d/owner-line.png)

General information for UK small firms, not legal, tax or financial advice.