On this page
Never file a citation an AI tool gave you without opening the actual opinion, confirming the quoted language appears at the cited page, and running the case through a citator. Language models generate text that looks like a citation, not text that is a citation, and courts have sanctioned lawyers who did not check.
That is the whole rule. What follows is why the problem exists, what it has already cost some lawyers, and a checklist that turns the rule into a habit.
What happened in Mata v. Avianca
In Mata v. Avianca, Inc., a 2023 personal injury case in the Southern District of New York, the plaintiff's lawyers filed an opposition brief that cited several judicial decisions. Defense counsel could not find them. Neither could the court. The lawyer who had prepared the research had used ChatGPT, which produced case names, reporter citations and quotations that did not exist. When the court ordered the lawyers to produce the opinions, one of them went back to ChatGPT, which generated full fake opinions, and those were filed too.
The court sanctioned the lawyers and their firm under Rule 11 of the Federal Rules of Civil Procedure, imposed a monetary penalty, and required them to send letters to the real judges whose names had been attached to the fabricated decisions. The sanctions order made a point that is worth quoting in substance: there is nothing improper about using a reliable AI tool for assistance, but the lawyer's duty to confirm that filings are accurate does not change.
Mata became the example everyone cites because it was first and because it was so complete. It has not been the last. Since 2023, courts across the country have issued orders addressing fabricated citations in filings by lawyers and self-represented parties, and many judges now have standing orders on the use of generative AI.
Why models fabricate citations
It helps to understand the mechanism, because it tells you where to look for errors.
A large language model predicts the next piece of text given what came before. It has read millions of legal citations, so it knows what one looks like: a plausible party name, a volume number, a reporter abbreviation, a page, a court and a year. When you ask it for authority on a point, it produces text in that shape. Sometimes the text corresponds to a real case because the model has seen that case cited many times. Sometimes it does not, and the model has no internal signal that anything is wrong. A fabricated citation and a real one are produced the same way.
Three consequences follow.
- Well-known cases are safer than obscure ones. A model will usually get Marbury v. Madison right. It is far less reliable on a 2019 district court decision it saw once.
- The citation can be real while the content is not. A model may attach a real case name to a holding the case does not contain, or invent a quotation and place it in a real opinion. This is harder to catch than a fake case, because a quick existence check passes.
- Pinpoints are the weakest part. Even when the case and the proposition are right, the page number is often a guess.
Tools that search a real corpus and cite from it (retrieval-based tools) are considerably more reliable than a general chatbot answering from memory, because the citation comes from a document the tool actually retrieved. They are not immune. The document may be the wrong one, or the summary may overstate it. The verification steps are the same either way; they are just faster.
The verification checklist
Do this for every citation that originated with an AI tool, and honestly, for every citation that originated with a junior associate, a treatise or your own memory. The steps are old. AI just made skipping them more expensive.
1. Pull the opinion
Find the case by its citation in a source that holds the actual text: Westlaw, Lexis, CourtListener, Google Scholar, the court's own website or the Frix Law Library. If the citation returns nothing, try the party names. If the party names return nothing, the case does not exist. Do not search for it in another chatbot.
Check that the court, year and reporter match. A model will sometimes produce a real case with the wrong court or a date ten years off.
2. Confirm the quotation
If the AI gave you quoted language, search for that exact string in the opinion. If it is not there, the quote is fabricated even if the case is real. If a similar sentence exists, quote the real one. Do not "fix" the quote by paraphrasing inside quotation marks.
3. Confirm the proposition
Read the part of the opinion the citation is supposed to support. Ask whether the court actually held that, or merely discussed it, or said it in a dissent, or quoted a party's argument before rejecting it. Models are especially prone to reporting the losing side's position as the holding.
4. Check the pinpoint
Confirm the page or paragraph number points at the language you are relying on. If you are using a free source without reporter pagination, find the star-paged version or cite to the slip opinion in a form the court accepts.
5. Run the citator
A real, accurately quoted case can still be overruled. Run it through KeyCite, Shepard's or a free citator and read any negative treatment. Our guide on how to check whether a case is still good law covers this step in detail.
6. Check the court's AI rules
Many federal and state courts, and many individual judges, now have standing orders or local rules on generative AI. Some require a certification that every citation was verified by a human. Some require disclosure of AI use. Some prohibit AI-drafted filings without review. These vary by court and by judge and change frequently, so check the judge's individual practices before every filing rather than relying on what was true last year.
7. Keep a record
Note which citations came from an AI tool and who verified them. If a question arises later, the record shows you did the work.
Signs a citation may be fabricated
Fabricated citations often share tells. None is conclusive, but any of them should raise your attention.
- The party names are generic or oddly balanced, like Smith v. United Airlines or Johnson v. City of Springfield.
- The reporter citation is plausible but the volume and year do not line up. Federal Reporter volumes and years follow a predictable sequence; a 2015 case cited to a volume from the 1990s is wrong.
- The quotation is unusually clean and on point. Real judicial prose is messier.
- The case appears nowhere except in the AI output. If neither Google Scholar nor CourtListener nor the Frix Law Library has it, it does not exist.
- The AI cannot give you a link to the opinion, or the link goes to a different case.
Where this fits in your workflow
The cheapest time to verify is when the citation is first produced, not when the brief is finished. If you use AI to draft or research, build the verification into the drafting step: every authority that goes into the outline gets pulled and checked before it goes into prose. A small firm can make this a rule for everyone who touches a brief. Our research workflow for solo and small firms shows where this sits in the wider process.
How Donna cites sources
Donna, the AI in Frix, is built so that the checklist above is short. When she answers a legal question, she searches the Law Library rather than answering from memory, and every answer shows numbered citations. Clicking a citation opens the opinion, statute or docket document at the cited page or passage. When she answers from the firm's own files, the same applies: the citation opens the PDF at that page with the passage highlighted. She runs the citator on the cases she cites and shows the flag.
That does not remove the duty to verify. It moves the verification from "does this case exist" to "does this passage support the point", which is a question you can answer in a minute by clicking the number. You can try this without an account at /ask, where three questions are free. For how this compares with a general chatbot, see Frix vs ChatGPT, and for the enterprise tools, Frix vs Harvey.
Frix does not train models on customer data, and the security page explains how firm files are handled.
Quick answers
Frequently asked questions
It is a case name, reporter citation, quotation or holding that a language model generated because it looked plausible, not because it exists. The citation may point to nothing, to a different case, or to a real case that does not say what the model claims.
Frix · Editorial team
We write about legal research and running a small firm. Every case and statute we mention links to its record in the free Frix Law Library, so you can read the source yourself. This is general information, not legal advice for any matter.