When AI Invents Court Cases: 1,981 Real Examples

At a glance

What this is: A plain-English look at a public database of court cases where someone filed AI-invented material, and what it teaches anyone who uses AI for anything official.

Who it is for: Anyone who uses ChatGPT, Claude or Gemini to draft something that another person will rely on. You do not need any legal background.

The rule to take away: AI is best at faking the things that come in a fixed format, like citations, figures and dosages. Those are exactly the things people skip checking.

A researcher named Damien Charlotin keeps a public database of court decisions in which someone filed material an AI made up. When we checked it on 29 August 2026 it held 1,981 cases, 1,363 of them in the United States. It has grown steadily since it began logging cases in 2023.

The story usually told about this is that lawyers keep embarrassing themselves. That is part of it. It is not the biggest part.

Table of Contents

Most of the people caught are not lawyers

Filter the database by who put the material in front of the court, and the order is not what the headlines suggest:

Self-represented litigants: 1,136 cases. People handling their own case without a lawyer.
Lawyers: 793 cases.
Judges: 29, in their own rulings. Experts: 15. Prosecutors: 5.

The largest group here is ordinary people, usually without money for representation, who asked a chatbot for help with the most consequential paperwork of their lives and had no way to check what came back. A lawyer who files a fake case faces a sanction and a referral to the body that licenses them. Someone representing themselves can lose the case itself.

Damien Charlotin’s AI Hallucination Cases database, checked 29 August 2026. Note the party filter: self-represented litigants outnumber lawyers.

What is actually going wrong here?

The database sorts the failures by type, and the pattern is worth knowing because it is the same pattern outside a courtroom. Of the logged cases, 1,646 involve fabricated material, 824 involve real sources misrepresented, and 539 involve false quotes pulled from real documents. One case can fall into more than one category, so those do not add up to the total.

Only the first category is invention from nothing. The other two are more dangerous for a beginner, because the source is real. The case exists. The law exists. The AI has just described it wrongly, or put words in it that were never there. You can look up the citation, find it, and still be wrong. This is what an AI hallucination usually looks like, and it is why “I checked the link worked” is not checking.

Why does the AI sound so certain?

Because a language model is built to produce text that reads like the thing you asked for. A legal citation has a very regular shape: names, a volume number, a court, a year. That shape is easy to reproduce perfectly while the contents are wrong. The more formulaic and authoritative a format looks, the better AI is at faking it and the worse you are at spotting it.

That is why it shows up in citations, statistics, medical dosages and specifications rather than in ordinary prose. If you have wondered whether you can trust AI search results, this is the specific failure to watch for.

The courts stopped treating it as a simple slip

Early rulings mostly let it go with a warning. The database now logs filings thrown out, people ordered to explain themselves in person, costs awarded against them, and reports to the body that licenses them. Browse the recent entries and you find sanctions running into five figures alongside plenty of cases resolved with nothing worse than a warning. Warnings still happen. Serious sanctions did not, two years ago.

The check that would have caught nearly all of these. Open every specific the AI gave you, in the original source, and read the part it claims to be quoting. Not the search result or the summary, the document itself. If you cannot find the passage, do not use it. That holds whether the AI invented it or you simply cannot reach the source.

How do you use AI for serious paperwork without this happening?

Use it to understand, not to cite. Asking AI to explain how something works is low risk. Asking it to supply the authority you will hand to someone else is where people get hurt.
Ask for the reasoning without the references. You lose nothing, and you remove the exact output that tends to be invented.
Check every specific in the original. Names, numbers, dates, quotes. This is dull and it is the whole job.
If you are asked whether you used AI, say yes. Several of the logged cases got worse when the filer denied it.
Get a person involved when a deadline carries a penalty. When losing costs you your home or your children, get a human involved.

What AI cannot do here

It cannot know which facts a decision-maker will actually care about in your situation.
It cannot tell you that your strongest argument is one you have not mentioned.
It cannot take responsibility. When a filing is wrong, the name on it is yours.

The same lesson turned up in a classroom this summer. A professor hid an instruction in white text inside an essay prompt, and 32 of his 35 students handed in answers containing the nonsense it produced, because none of them read their own work before submitting it. We covered that in our guide to AI for teachers. Same failure, different room: nobody read their own work before handing it in.

One AI story a day, explained properly

Human curated and edited, for people who want to use AI well without reading the AI press.

Free forever. Unsubscribe anytime.

Common questions about AI in court filings

Does this mean AI is useless for legal questions?

No. It is good at explaining what a term means, what a process usually involves, and what questions to ask a professional. It is unreliable at supplying the specific authority for a claim. Ask it what a term means. Ask a professional what to file.

Do the paid versions invent less?

Somewhat, and not enough to skip checking. Live search helps most with outright invention. It helps far less when the AI has a real document in front of it and still describes it wrongly, which is the harder failure to spot.

Is this only a United States problem?

No. The database logs cases across more than 35 jurisdictions, including 214 in Canada, 98 in Australia and 62 in the UK.

How would I check a citation if I am not a lawyer?

Search the case name in a free court records service and open the decision itself. CourtListener covers United States federal and state courts, BAILII covers the UK and Ireland, and most other countries have an official equivalent. If the name returns nothing, or returns an unrelated dispute, the citation is wrong. Then read the passage the AI attributed to it. Most failures are visible at that step.

Where else does this failure show up?

Anywhere the format is rigid and the reader is unlikely to check: academic references, medical dosages, tax rules, technical specifications, and statistics with a decimal point in them. Courts publish their decisions, so these mistakes are countable. In most other fields nobody is keeping score.

Sources

Damien Charlotin — AI Hallucination Cases database (all counts read 29 August 2026; database last updated 28 August 2026)
Norton Rose Fulbright — AI in litigation: update on generative AI sanctions in 2026
Grokipedia — Hallucination (artificial intelligence)

What Claude Remembers About You