She Cited Nine Tribunal Cases To Win Her Appeal—None Of Them Were Real: Harber v HMRC

She cited nine tribunal decisions to back her appeal. Not one of them was real.

Mrs Harber was an unrepresented appellant fighting a £3,265 capital gains tax penalty—the kind of penalty that lands on people who sell a property, do not realise they had to tell HMRC, and find out only when a letter arrives. To strengthen her case, she put forward nine First-tier Tribunal decisions that supported her argument. They had names, dates, and summaries. They looked exactly like real case law.

They were generated by an AI tool. The Tribunal found, as a fact, that every one of them was invented.

This is Harber v HMRC, and it has become the leading UK warning about relying on AI-generated case law. It is also a story with two separate lessons, and this article keeps them apart on purpose. First: she would have lost her appeal regardless—the fake cases changed nothing about the outcome. Second: citing invented cases is a separately serious problem, and one that any unrepresented appellant could stumble into today. So this is both a cautionary tale and a practical guide—because the most useful thing here is learning how to find and check real case law yourself.

What The Case Was Actually About

Strip away the fake cases and Harber is an ordinary penalty appeal. Mrs Harber bought a property in 2006 for £135,000 and sold it in October 2018 for £252,000. That produced a capital gain of £82,327.69.

When you make a taxable capital gain and you are not already in self-assessment, the law requires you to tell HMRC. That duty comes from section 7 of the Taxes Management Act 1970, which says a person chargeable to capital gains tax must give notice that they are chargeable—and the deadline is six months after the end of the tax year. Mrs Harber sold in October 2018, which falls in the 2018-19 tax year, so she had to notify HMRC by 6 October 2019. She did not notify until 5 January 2021.

That delay triggered a "failure to notify" penalty under Schedule 41 to the Finance Act 2008. The CGT itself came to £16,325.56. HMRC treated her failure as "prompted but non-deliberate"—prompted because HMRC raised it first rather than her coming forward, non-deliberate because there was no finding she set out to hide the gain. That behaviour category carries a penalty of 20% to 30%, and because HMRC gave the maximum reduction for cooperation, the rate landed at 20%—a penalty of £3,265.11. On top of the tax and the penalty, tax paid late also runs up interest, which is a separate cost worth factoring into your total exposure—we cover it in interest on unpaid tax.

Here is a point worth holding onto as a reader: Mrs Harber appealed only the penalty, not the underlying CGT (judgment §39). She accepted she owed the tax. Her fight was about whether she had a reasonable excuse for telling HMRC late. You can lose the argument about the tax and still run a reasonable-excuse defence on the penalty—they are different questions. Our guides to self-assessment penalties and how HMRC penalties work explain how these behaviour categories and reductions fit together.

The Nine Cases That Did Not Exist

The fake cases nearly slipped through unnoticed. Mrs Harber had sent the Tribunal a written document—called "the Response"—setting out her argument and the nine supporting decisions. But she had not copied it to HMRC (§8). Nobody on HMRC's side had seen the cases until the very end of the first hearing, which meant the hearing had to be adjourned and reconvened months later so HMRC could check them.

Check them HMRC did. Ms Man, HMRC's litigator, searched the First-tier Tribunal website for each case, widening her search several years either side of each name. She found none of them (§16). The Tribunal independently checked the FTT website and other legal databases and also found nothing (§18).

When the Tribunal asked Mrs Harber where the cases had come from, she said a "friend in a solicitor's office" had provided them. She had no full judgments and no tribunal reference numbers—only the short summaries. When asked directly whether they might have been generated by an AI system such as ChatGPT, she said this was "possible" (§18).

The Tribunal's conclusion was unambiguous. At §21 it found "as a fact that the cases in the Response are not genuine FTT judgments but have been generated by an AI system such as ChatGPT."

Two things matter about how the Tribunal handled this. It accepted that Mrs Harber did not know the cases were fake, and that she did not know how to find or check case law on the tribunal website or other legal sites (§22). This was not treated as dishonesty. It was treated as a sympathetic but serious mistake—exactly the mistake an unrepresented person trying their best could make.

How AI Invents Case Law

To understand why this happens, it helps to know what these tools actually do. The Tribunal quoted the Solicitors Regulation Authority's description of the problem (§20(3)): "AI language models such as ChatGPT, however, can be more prone to this. That is because they work by anticipating the text that should follow the input they are given, but do not have a concept of 'reality'. The result is known as 'hallucination', where a system produces highly plausible but incorrect results."

That is the key idea. An AI tool does not look up cases in a database. It predicts text that looks right. Ask it for tribunal decisions supporting your argument and it will happily produce a list of cases with realistic names, plausible citations, and tidy summaries—none of which exist. The output looks authoritative precisely because looking authoritative is what it is built to do.

The Tribunal pointed to several tell-tale signs that the cases were fabricated, and these are worth learning to spot:

• American spelling. The word "favor" (the US spelling) appeared in six of the nine summaries—a strong hint the text was generated by a model trained largely on American material, not drawn from UK judgments.
• Repeated identical phrasing. The same phrases recurred across supposedly separate, independent decisions—real judgments written by different judges do not read like that.
• Real surnames bolted onto invented winning outcomes. Some of the fake names matched genuine taxpayers from real cases—but in the real cases those taxpayers had lost. The genuine Baker lost; the genuine Christine Perrin lost. The fabricated versions conveniently won.
• The wrong kind of penalty. Nearly all the fabricated cases were about late filing, not failure to notify—the wrong category for Mrs Harber's situation. A real adviser pulling authorities would have noticed the mismatch. (These are genuinely different regimes: if HMRC says you never told them you were chargeable, that is failure to notify under Schedule 41; if you were already in self-assessment and filed your return late, that is a separate regime under Schedule 55. Make sure any case you rely on is about your kind of penalty.)

If you ever find yourself with a list of cases that win every time, share suspiciously similar wording, and use American spellings, treat that as a red flag and check each one before you rely on it.

Did The Fake Cases Cost Her The Appeal?

No. The Tribunal was explicit: "our decision would have been the same if Mrs Harber had not provided the cases in the Response" (§5). She lost on the real law, which we come to below. The fakes did not tip the result against her.

So why did the Tribunal spend a substantial part of its judgment on the issue at all? Because it matters in its own right, independently of who won. As the Tribunal put it, "providing authorities which are not genuine and asking a court or tribunal to rely on them is a serious and important issue" (§5).

The Tribunal explained the harm in full, and the quote is worth reading whole: "But that does not mean that citing invented judgments is harmless. It causes the Tribunal and HMRC to waste time and public money, and this reduces the resources available to progress the cases of other court users who are waiting for their appeals to be determined" (§24). Fabricated authorities also corrode something less tangible—the Tribunal noted that they promote cynicism about the very idea of precedent, the system of building on what real cases have decided.

To make the point, the Tribunal drew on a now-notorious American example. In Mata v Avianca—a US case, illustrative only and not a UK authority—two US lawyers filed a court brief citing cases that, it turned out, ChatGPT had invented. When the US court ordered them to produce the full judgments, they went back to the AI tool, which obligingly generated longer fake texts. The US judge, Judge Castel, fined the lawyers $5,000.

The contrast with Mrs Harber is instructive. The US lawyers were professionals who were sanctioned; Mrs Harber was an unrepresented individual who did not know the cases were fake and was treated sympathetically. The Tribunal adopted Judge Castel's summary of the damage (§23): "Many harms flow from the submission of fake opinions. The opposing party wastes time and money in exposing the deception. The Court's time is taken from other important endeavors. The client may be deprived of arguments based on authentic judicial precedents... It promotes cynicism about the legal profession and the…judicial system." (The American spellings—"endeavors"—are in the original because the passage is from a US judge.)

The lesson is not that AI use makes you a wrongdoer. It is that fabricated citations cause real harm even when the person citing them acted in good faith—and a represented or professional party who does the same can expect a far harder reception.

It is worth being clear about what did and did not happen to Mrs Harber, because the headline can frighten people who have used AI themselves. She was not fined, sanctioned, or given any extra penalty for the citations. She was treated as having made an honest mistake, and the only thing she lost was the penalty appeal—which, as the Tribunal said, she would have lost anyway. The $5,000 fine belongs to the US lawyers in Mata, not to her.

Why She Actually Lost: The Real Reasonable-Excuse Test

The appeal was decided on a genuine piece of law: the four-step reasonable-excuse test from Perrin v HMRC [2018] UKUT 156 (TCC). This is the real Perrin—an Upper Tribunal decision that the taxpayer, Christine Perrin, actually lost. It is worth pausing on the irony: one of the invented cases was a "David Perrin" who supposedly won, while the real Perrin, on whose framework the whole appeal turned, lost. We have a full analysis of the genuine decision at Perrin v HMRC: the four-step reasonable-excuse test.

The Perrin test, which the Tribunal set out at §50, asks four questions. What facts does the taxpayer say give them a reasonable excuse? Which of those facts are proven? Were those proven facts, viewed objectively, a reasonable excuse for this particular taxpayer in these circumstances? And was the failure put right without unreasonable delay once any excuse ended? Our plain-English reasonable excuse guide walks through each step.

Mrs Harber relied on two excuses, and the Tribunal applied the test to each.

Her mental-health condition (§52-§56). This was real and undisputed. She had suffered anxiety and panic attacks for years, made worse by bereavement and a relationship breakdown around the time of the sale. But at step three the Tribunal asked whether that condition had actually prevented her from contacting HMRC—and concluded it had not. During the same period she had marketed and sold the property, instructed solicitors, worked out her likely gain, put £20,000 into government bonds (gilts), and dealt with lodgers. A condition that did not stop her doing all of that did not, on these facts, provide a reasonable excuse for failing to tell HMRC about the sale.

This does not mean a mental-health condition can never be a reasonable excuse. It can—where it genuinely prevented you from dealing with your tax in time, and you have evidence of that. Harber turned on its own facts, not on any rule that mental health does not count.

Ignorance of the law (§57-§61). Some points were in her favour: she had never completed a self-assessment return, and her solicitors gave her no CGT advice. But the facts against her were stronger. She knew she had made a gain. She knew HMRC would want money—she set aside £20,000 in gilts precisely for that. She had previously telephoned HMRC about the rent-a-room scheme, so she knew HMRC gave free advice by phone—yet she did not call about the gain. And she declined professional advice only because it was "expensive." The Tribunal found that a reasonable person who knew they had a gain would have contacted HMRC, Tax Aid, an accountant, or a lawyer to find out what to do, rather than simply setting money aside and waiting.

Importantly, ignorance of the law is not an automatic bar to a reasonable excuse. The genuine Perrin decision makes clear it can sometimes count, depending on the taxpayer and the obscurity of the rule. But here it was not objectively reasonable: Mrs Harber knew enough to know she should ask, had a free route to ask, and chose not to. The result, at §62, was that neither excuse succeeded, and the penalty was confirmed.

How To Find And Verify Real Case Law

This is the part that protects you. around 45% of tribunal appellants represent themselves, and more and more of them turn to AI tools for help. That is understandable—but a tool that invents case law is dangerous in exactly the place you most need to be right. So here is how to find genuine decisions and confirm they exist.

Use the official and trusted sources. There are three you can rely on:

• National Archives Find Case Law (caselaw.nationalarchives.gov.uk) is the official, government-run database of court and tribunal judgments. It is the best place to start for recent decisions and the most authoritative modern source. Harber itself is on there.
• BAILII (bailii.org) is a long-established free legal database. First-tier Tribunal (Tax Chamber) decisions sit under /uk/cases/UKFTT/TC/, and Upper Tribunal (Tax and Chancery Chamber) decisions under /uk/cases/UKUT/TCC/.
• The GOV.UK tax tribunal decisions finder covers First-tier decisions from around May 2022 onwards.

Watch the citation-number trap. A case's official citation is not the same as the file or reference number it is stored under. Harber is cited as [2023] UKFTT 1007 (TC)—but on BAILII its stored file number is TC09010, which bears no relation to the 1007 in the citation. If you assume the file number mirrors the citation and search on that basis, you may wrongly conclude a real case does not exist (or, worse, trust a tool that gave you a confident but wrong reference). When in doubt, search by the party name and the year.

Verify, do not trust. This is the single rule that would have saved Mrs Harber. A citation only counts if it resolves to a full judgment you can open and read. A summary is not proof—anyone, or any AI tool, can write a convincing summary of a case that never happened.

In practice, checking a case is real takes about a minute:

1. Go to caselaw.nationalarchives.gov.uk (or bailii.org).
2. Type the party names and the year into the search box—for example, "Harber" and "2023."
3. Open the result and confirm it is a full judgment, with numbered paragraphs and a named judge, not just a short summary.
4. Read the decision at the end to see who actually won, and on what point.

If you cannot find the case this way, treat it as not existing until you have proof otherwise. And when you do find it, check the citation, the court, the year, and—crucially—who won on the point you are relying on. Remember the fabricated Perrin who "won" when the real one lost. A case that decided the opposite of what you think is worse than no case at all.

For turning genuine authorities into a proper argument, see our guides to writing grounds of appeal and preparing for your tribunal hearing.

Using AI Safely

None of this means AI tools are off-limits. Used in the right place, they can genuinely help an unrepresented appellant. They are fine for explaining a concept in plain English, helping you structure an argument, understanding what a penalty notice means, or generating a list of questions you then go and research properly.

What they must never be is a source of authority. Any case an AI tool gives you—name, citation, summary, all of it—has to be found and read independently on Find Case Law or BAILII before you rely on it. The model is built to produce text that looks correct, which is precisely why its invented citations look perfect. Treat anything it tells you about the law as a lead to check, never as a fact to cite.

If You Have Already Cited A Case You Cannot Verify

If you have already put AI-generated cases in front of HMRC or the Tribunal, you are not the first, and it is not necessarily fatal. The thing that helped Mrs Harber was honesty: when the Tribunal asked where the cases had come from, she said so, and she was treated as having made a genuine mistake rather than as someone trying to deceive.

The sensible response is to correct the record promptly. Tell both the Tribunal and HMRC, in writing, that you can no longer stand behind those authorities and will rely on your genuine grounds instead. Doing that early protects your credibility and avoids wasting a hearing—the opposite of what happened in Harber, where the untested cases only came to light at the end of the first hearing and forced everyone back for a second one.

What To Do Now

If you are facing a penalty and thinking about an appeal, the practical takeaways from Harber are short.

Bring a genuine appeal on real grounds. A reasonable-excuse argument built on your actual circumstances—evidenced and honestly presented—is worth far more than a stack of impressive-looking cases. Cite only authorities you have found and read yourself.

Get free help if you need it. TaxAid and the Low Incomes Tax Reform Group (LITRG) help people on lower incomes for free, and HMRC's Self Assessment helpline answers general questions. Mrs Harber's case is a reminder that "advice is expensive" is rarely the whole story—free help exists.

And file everything on both sides, in good time. Mrs Harber's Response reached the Tribunal but not HMRC, which forced an adjournment and a second hearing. Whatever you send to the Tribunal, send to HMRC too. Tax tribunal appeals carry no fee to file, and you generally have 30 days from HMRC's decision to lodge one, so there is no advantage in cutting corners. The clock usually starts with the penalty notice or decision letter, and you normally appeal to HMRC first—or ask for a review—before going to the Tribunal; our guide to how to appeal to the tax tribunal walks through the steps. If that deadline has already passed, a late appeal may still be possible, though you will need to explain the delay.

Key Legislation And Resources

The Judgment

• Felicity Harber v HMRC [2023] UKFTT 1007 (TC) — First-tier Tribunal (Tax Chamber), Judge Anne Redston and Ms Helen Myerscough, released 4 December 2023 (note the file number TC09010 differs from the citation 1007)
• Also available on National Archives Find Case Law — the official government database

Legislation

• Section 7, Taxes Management Act 1970 — the duty to notify HMRC of liability to income tax and capital gains tax, within six months of the end of the tax year
• Schedule 41, Finance Act 2008 — the failure-to-notify penalty regime
• Paragraph 20, Schedule 41, Finance Act 2008 — the reasonable-excuse defence, available only where the failure was not deliberate

Key Cases

• Felicity Harber v HMRC [2023] UKFTT 1007 (TC) — nine cited tribunal decisions found as a fact to be AI-generated (§21); citing invented authorities is "a serious and important issue" (§5) even where it does not change the outcome
• Christine Perrin v HMRC [2018] UKUT 156 (TCC) — the genuine four-step reasonable-excuse test applied in Harber; the real Perrin lost (contrast the fabricated "David Perrin" who supposedly won)
• Mata v Avianca, No. 22-cv-1461 (S.D.N.Y. 2023) — US case, illustrative only, not a UK authority — US lawyers sanctioned $5,000 for filing a brief citing ChatGPT-fabricated decisions; adopted by the Tribunal in Harber at §23 to describe the harms

On This Site

• Perrin v HMRC: The Four-Step Reasonable Excuse Test — the genuine test on which Harber turned
• What Is A Reasonable Excuse? — plain-English primer on the defence
• Self-Assessment Penalties — the wider penalty regime, including failure to notify
• HMRC Penalties Explained — behaviour categories, prompted vs unprompted, and mitigation
• Writing Grounds Of Appeal — how to use real authorities properly
• Preparing For Your Tax Tribunal Hearing — bundles, evidence, and filing on both sides
• How To Appeal To The Tax Tribunal — the step-by-step process for lodging an appeal
• Late Appeal To The Tax Tribunal — what to do if the deadline has already passed
• Interest On Unpaid Tax — how interest adds to your overall exposure
• CGT 60-Day Reporting Appeals — a related CGT-reporting obligation for more recent property sales

This article is for informational purposes only and does not constitute legal or tax advice. For advice specific to your situation, consult a qualified tax adviser, accountant, or solicitor.