If you used an AI tool anywhere in drafting a brief — even just to outline arguments or find candidate cases — you have exactly one job before you file: verify every citation against a primary source, not against your own sense that it “looks right.” AI-generated legal text is specifically optimized to look right. That is the entire problem, and it is why researcher Damien Charlotin's public AI Hallucination Cases database now tracks well over 1,700 documented instances of fabricated or misused AI citations reaching courts, across firms of every size.
Why this is not a fringe problem
The Stanford RegLab/HAI study "Hallucinating Law" (Dahl, Magesh, Suzgun & Ho, 2024) found general-purpose large language models hallucinate on legal queries at rates between 58% and 82%, depending on the model and question type. A follow-up Stanford benchmark, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools," tested purpose-built legal RAG products — Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI — and still found hallucination rates ranging from roughly 17% to 34% depending on the product. The takeaway is not "avoid AI tools." It is that no category of tool, including expensive purpose-built legal research platforms, eliminates the need for independent verification.
Step 1: Extract every citation, not just the ones you remember adding
Go through the document line by line and list every case citation, statute, and quoted passage. It is common for an AI draft to insert a citation you did not consciously request, buried inside a sentence that otherwise reads as your own writing. Missing one because it "felt like something I wrote" is how fabricated citations survive review. Treat this as a mechanical inventory step, separate from evaluating the substance of any argument.
Step 2: Check existence first, separately from everything else
For each case citation, confirm it exists at the exact reporter, volume, and page cited. A fast way to do this by hand is CourtListener's free citation lookup, maintained by the nonprofit Free Law Project. Do this before you evaluate whether the case supports your argument — there is no point analyzing the legal reasoning of a case that does not exist. In Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023), this exact step, run by opposing counsel, is what unraveled the entire filing: several cited cases simply could not be located in any reporter or database.
Step 3: Check every quotation against the actual opinion text
A real case can still carry a fabricated quotation. If your brief puts language in quotation marks and attributes it to a case, find that exact language in the opinion. This is the step most lawyers skip because it is tedious — it is also, per the Charlotin database, one of the most common failure modes courts have caught, sometimes in cases that otherwise exist and are correctly cited to the right reporter and page.
Step 4: Confirm the case actually supports the proposition
A citation can pass both of the above checks and still be wrong: a real case, accurately quoted, cited for a holding it does not actually stand for, or from a portion of the opinion (a dissent, a vacated section, an overruled holding) that is not good law on the point. Read the actual passage you are relying on, not just the headnote or the AI tool's summary of it.
Step 5: Keep a record that you did this
If your verification process is ever questioned by a court or bar authority, "I checked and here is my documented process" is a materially stronger position than "I trusted the tool." A simple checklist with a timestamp, kept in the file, costs nothing and has real evidentiary value later.
Case examples from the public record
Beyond Mata, the pattern recurs across practice areas and firm sizes. In Park v. Kim, 91 F.4th 610 (2d Cir. 2024), the Second Circuit flagged a fabricated citation in an attorney's reply brief and referred the matter for further review, noting the citation "does not exist." In Wadsworth v. Walmart Inc. (D. Wyo. 2025), attorneys from a national plaintiffs' firm were sanctioned after AI-generated fictitious case citations appeared in a motion, and the firm subsequently adopted new internal AI-verification policies. These are not obscure outliers; they are three of the more widely reported entries in a growing, searchable public record.
Expert perspective
Damien Charlotin, who built and maintains the hallucination-cases database used throughout this piece, has noted in his own published analysis that the recurrence rate is the real story: this is not a problem that was solved after Mata made headlines in 2023. New entries continue to be added on a roughly weekly basis, spanning solo practitioners, mid-size firms, and pro se litigants alike. That pattern is the strongest evidence available that verification needs to be a mandatory workflow step, not a matter of individual diligence or tool selection.
A checklist for before you file
- Every citation extracted into a single list, including ones buried mid-sentence.
- Every case citation checked for existence against CourtListener or an equivalent primary source.
- Every quotation checked word-for-word against the actual opinion text.
- Every relied-upon holding checked against the actual passage, not a headnote or AI summary.
- Good-law status confirmed (not overruled, vacated, or reversed on the relevant point).
- A dated record of this process kept in the file.
A common question
Do I need special software to do this, or can I do it by hand?
You can do every step by hand for free — CourtListener's citation lookup and full-text search cover existence and quote checks at no cost. A dedicated tool mainly buys you speed and a documented, repeatable process across an entire document at once, which matters more as document length and deadline pressure both increase. See our comparison of verification tools for what each category actually checks.