Behind AI at Work
Nine tax cases that never existed
A tax appeal in 2023 rested on nine earlier decisions. None of them were real. The story that opens AI at Work, and the habit it teaches.
AI at Work opens with a woman who owed HMRC £3,265.
She had sold a property and hadn’t told HMRC about the capital gains tax that was due, so she was fined. She appealed, arguing she had a reasonable excuse. To back that up she gave the First-tier Tribunal nine earlier decisions in which people in her position had won.
None of them existed.
What the tribunal found
The judge, Anne Redston, and HMRC’s lawyers went looking for every one. Some had names that looked like real cases, but the real cases had different years, and the people in them had lost. The tribunal decided, in December 2023, that the nine had most likely come from an AI tool such as ChatGPT.
She said she’d been given them by “a friend in a solicitor’s office”, and the judge accepted that she hadn’t known they were fake. She lost her appeal anyway. Her excuse had been built on cases nobody could find.
It would be easy to read that as a story about someone who didn’t know better. The next one isn’t.
Then it happened to lawyers
In June 2025 the High Court in London dealt with two cases together. In one, a barrister had put five cases that didn’t exist in front of the court. In the other, 18 of the 45 authorities cited in a witness statement were fictitious. Dame Victoria Sharp, who led the court, sent the judgment to the Bar Council and the Law Society, so that every barrister and solicitor in England and Wales would hear about it.
In July 2025 it was a consultancy’s turn. Deloitte’s 237-page report for the Australian government’s employment department, a contract worth AU$440,000, turned out to cite academic papers nobody had written and a quote from a court judgment that the judgment didn’t contain. The firm published a corrected version, and in October 2025 the department said Deloitte would refund the final payment.
A taxpayer, a barrister, a global firm. None of them set out to deceive anybody. Each trusted a paragraph because it read well.
Why it keeps happening
A chatbot doesn’t look things up the way a person does. It writes the most likely next words, based on an enormous amount of text it was trained on. Most of the time the likely words are also the right ones. When they aren’t, nothing in the answer tells you.
The things it invents most often are the things that look most like evidence: case names, quotes, statistics with a source attached, web addresses, page numbers. They’re exactly the details a reader is least likely to question, because they look checked.
The habit the book teaches
The first chapter of AI at Work sets out a rule that comes back again and again in the rest of the book: the more specific and checkable a claim is, the more you should check it.
You don’t have to check every sentence a tool writes. You do have to check the names, the numbers, the dates, the prices and anything presented as a quote or a rule. Chapter 9 gives you a prompt that makes the tool list them for you before you rely on the text:
List every factual claim in the text below that someone could check: names, numbers, dates, prices, quotations, laws or rules, web addresses, and anything stated as fact. For each one, say what kind of source would confirm it. Don’t judge whether they’re true; just list them so I can check them.
The tool won’t tell you which claims are wrong, and that’s deliberate. It hands you a short list to check against real sources. Nine tax cases would have gone on that list, and none of them would have survived it.
About the book
AI at Work: From Your First Prompt to Your Own AI Assistants is a plain-English handbook for UK workplaces. It starts with a free account and an email that needs fixing, and it ends with assistants, agents and automations doing real work in small businesses. Four invented people go through it with you: an office manager at a plumbing firm, a letting agent, a dental receptionist and a one-man bookkeeper.
It has 246 numbered prompts, each with the words to type and what to check afterwards.
Sources
- [Harber v Commissioners for His Majesty’s Revenue and Customs [2023] UKFTT 1007 (TC), 4 December 2023](https://caselaw.nationalarchives.gov.uk/ukftt/tc/2023/1007)
- Legal Futures, “Litigant unwittingly put fake cases generated by AI before tribunal”, December 2023
- Legal Cheek, “High Court warns lawyers over AI use after fake cases cited in submissions”, June 2025
- AP via CTV News, “Deloitte to partially refund Australian government for report with apparent AI-generated errors”, October 2025