The fastest way to tell whether an AI gave you a real legal citation is to open the actual case yourself, in a source the AI didn't hand you -- not its summary, not its restated holding, the opinion itself. That single habit, run every time, would have stopped every incident below before it reached a judge, from a $5,000 sanction in 2023 to a bar-discipline referral still pending in 2026.
Nothing about this problem is contained to one chatbot or one country. Paris-based researcher Damien Charlotin has tracked, since the practice first surfaced, every court decision where a judge found or clearly implied a party relied on hallucinated AI material; as of October 2, 2026, the database held 2,125 cases. The same tracker had counted 1,598 as of June 9, 2026 -- the figure cited when OpenAI launched a dedicated legal-research product into exactly this backdrop -- meaning the total grew by more than 500 cases in under four months.
Where this is actually happening
The United States accounts for most of the recorded cases, but not all of them -- a reminder that this isn't a quirk of one country's court system or one AI product's rollout.
Where the tracked cases are happening
Five incidents, two years apart, same underlying mistake
The specifics vary -- different courts, different AI tools, different stakes -- but the failure is identical every time: nobody opened the primary source before it went into a filing.
From a $5,000 fine to a certification requirement
- Jun 22, 2023 — A federal judge fines two ChatGPT-citing attorneys and their firm $5,000 in Mata v. Avianca -- the case that put AI-fabricated citations on the map.
- May 23, 2025 — A magistrate judge strikes part of Anthropic's own expert declaration in Concord Music Group v. Anthropic after Claude fabricates an article's title and authors while formatting citations.
- Mar 20, 2026 — The Nebraska Supreme Court refers attorney Greg Lake for discipline over a divorce-appeal brief it calls 'filled with fictitious cases ... and fictitious quotations,' after he denied using AI.
- Mar 24, 2026 — The Sixth Circuit sanctions two attorneys $15,000 each, plus fees and double costs, over 24-plus fake citations in Whiting v. City of Athens -- 'the stiffest penalty available.'
- Jun 3, 2026 — The Ninth Circuit suspends a firm's attorneys from its practice for six months in Lnu v. Blanche and orders every future filing to certify citations were personally verified.
The Anthropic case is worth sitting with, because it shows the failure isn't limited to a lawyer typing a loose prompt into a general chatbot. Anthropic's own data scientist used Claude.ai to help format citations in a sworn declaration -- not to do legal research, just formatting -- and the tool invented a plausible article title paired with two authors who had never worked together. Magistrate Judge Susan van Keulen struck the relevant paragraph, writing that the error 'undermines the overall credibility' of the entire declaration. If a formatting task can introduce a fabricated citation into a filing from the company that makes the model, a substantive research task from any tool deserves the same scrutiny -- which is exactly what a 2024 Stanford RegLab study found when it tested commercial legal-AI products built specifically to avoid this: Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI each still hallucinated between 17% and 33% of the time, directly contradicting vendor marketing describing the tools as eliminating or avoiding hallucination entirely.
The five-minute check before you file
None of the five steps below require a specialized tool or a paid verification product. They require treating an AI's citation exactly the way you'd treat an unsourced tip from a stranger: a useful starting point, worthless as a final answer.
Verify an AI-generated citation before it reaches a filing
- Search the case name and number in a free, independent database -- Google Scholar, CourtListener, or your court's own public docket -- never the AI's own restated summary.
- A citation can point to a real case and still be false -- Concord Music Group v. Anthropic involved a real-sounding article title with entirely fabricated authors, not an invented case name.
- Use Shepard's, KeyCite, or a free citator (CourtListener and Google Scholar both show how later cases have treated an opinion) before relying on its holding.
- The Ninth Circuit's order in Lnu v. Blanche now requires one firm to certify every citation was personally verified -- a growing number of districts have their own standing orders, and they are not uniform.
- Attorneys who disclosed AI use early have been sanctioned less severely than those who didn't; the Nebraska Supreme Court's complaint against Greg Lake centered on his initial denial as much as the fabricated citations themselves.
Running that check once closes the actual gap. The ways it gets skipped anyway cluster around four specific, recurring mistakes.
Four ways this check gets skipped when it shouldn't be
None of this is an argument against using AI for legal research -- the time it saves is real, and refusing to use it is not itself a safeguard. It's an argument for treating a citation exactly the way a broader check for AI fabrication already recommends treating any specific, checkable claim: useful as a lead, never as the final word, until you've personally opened the source. Anthropic and OpenAI both power tools now marketed directly at legal work, and neither has claimed its product is immune to the failure mode documented above -- which is precisely why the check belongs to the person signing the filing, not the vendor that built the tool.
- A public tracker counted 2,125 worldwide court cases of AI-fabricated legal citations as of October 2026.
- Specialized legal-AI research tools hallucinate too -- Stanford found 17%-33% error rates, not zero.
- Courts have fined, suspended, and referred attorneys for discipline over fake citations since 2023.
- One 2026 appeals ruling now requires attorneys to certify every citation was personally verified.
- Opening the real source yourself is the only check that counts -- a tool's label never is.