What a hallucinated citation is
A hallucinated citation is a case that a generative AI tool invented. The caption sounds plausible, the reporter volume and page numbers are formatted correctly, and the parenthetical describes a holding that would be perfect for your argument. The case simply does not exist. In other variations, the case exists but the AI attributes a quote, a holding, or a pincite to it that appears nowhere in the opinion.
Doe v. Smith, 478 F.3d 419, 425 (1st Cir. 2009) (holding that conclusory allegations cannot survive a motion to dismiss)
Large language models generate text by predicting what plausible legal writing looks like. They are not querying a case database, so nothing in the generation process checks whether the authority is real. When the model needs a citation to complete a sentence, it produces one that fits the pattern. That is why hallucinated cites are so dangerous: they are formatted exactly like the real thing.
The sanctions record so far
Courts started sanctioning lawyers for AI-fabricated authority in 2023, and the pace has only increased. A few of the widely reported early decisions:
- 2023
Mata v. Avianca (S.D.N.Y.)
Attorneys filed an opposition citing six nonexistent cases generated by ChatGPT, then doubled down when questioned. A $5,000 sanction against the lawyers and their firm, and national headlines. - 2023
People v. Crabill (Colorado)
A lawyer filed a motion with fabricated ChatGPT citations and did not correct it after discovering the problem. Suspended from practice. - 2024
Park v. Kim (2d Cir.)
A reply brief cited a nonexistent case produced by ChatGPT. Referred to the circuit grievance panel. - 2024
Gauthier v. Goodyear (E.D. Tex.)
A response brief included nonexistent cases and fabricated quotes from real cases. Monetary sanction plus mandatory AI education. - Today
Dozens of orders and counting
Similar orders keep issuing in federal and state courts, and judges now search filings for fabricated authority on their own initiative.
The pattern in every one of these matters is the same. The sanction was never for using AI. It was for filing authority no one verified, and in the worst cases for standing behind it after opposing counsel or the court raised questions.
Why AI invents cases
Language models are trained to continue text convincingly. Real judicial opinions cite authority constantly, so when a model writes in the register of a brief, it has learned that a citation belongs in the sentence. If its training data contains the perfect case, it may reproduce it. If not, it interpolates: a plausible party name, a real reporter, a volume and page in the right numeric range.
- Blended authority. The model merges two real cases into one that never existed, borrowing a caption from one and a holding from another.
- Real case, fake content. The citation resolves to a real decision, but the quoted language or the proposition attributed to it appears nowhere in the opinion. These are harder to catch than fully invented cases.
- Wrong pincite. Everything is real except the page. The proposition lives at page 570 and the brief says 544. Small, but it is the first thing a careful clerk checks.
Newer models hallucinate less, not never. And tools that wrap AI around real legal databases reduce fabrication but still misstate holdings and misattribute quotes. Every generation of the technology has produced its own sanctions orders.
How to catch fake citations before you file
Every citation in a filing should clear four checks. This is true whether the draft came from an associate, co-counsel, or an AI tool, but AI drafts make the first check non-negotiable.
| Check | Question it answers | How to do it |
|---|---|---|
| Existence | Is this a real case? | Pull the case by citation in Westlaw, Lexis, or a free source like CourtListener. If the citation resolves to nothing, or to a different case, stop. |
| Quote accuracy | Does the opinion say what the brief says it says? | Compare every quotation word for word against the original. Paraphrases inside quotation marks are misquotes. |
| Pincite support | Does the cited page support the cited point? | Read the actual page. The proposition must appear there, not three pages later. |
| Status | Is the authority still good law? | Run the citator. A real case that was overruled is as damaging as a fake one. |
Done by hand, this takes several minutes per citation, which is why it gets skipped under deadline pressure, and why the sanctions orders keep coming. The failure mode is never that a lawyer could not check. It is that no one did.
Where CiteScreen fits
CiteScreen automates the mechanical part of that checklist. Upload a brief and it extracts every citation, matches each case to its source, confirms the pincite against the cited page, and compares quotations word for word with the original opinion. A citation that resolves to nothing gets flagged as likely fabricated, with the reason stated in plain English.
The output is a report a reviewing attorney can act on: what cleared, what needs judgment, and why. It does the checking a sanctions order says someone should have done, in minutes instead of an afternoon.
Frequently asked questions
No reported sanction has been for AI use itself. Sanctions follow from filing unverified or fabricated authority and from failing to correct it. Using AI to draft and then verifying every citation before filing satisfies the duty of reasonable inquiry.
Tools grounded in real case databases fabricate far less than general chatbots, but studies and court filings have documented misattributed quotes and misstated holdings from every category of tool. Verification against the published decision is still required.
Pull it by citation, not by party name. A fabricated cite either resolves to nothing or to an unrelated case. Also check internal consistency: reporter volumes map to date ranges, so a 2009 decision cannot appear in a volume published in 2007.
Correct it immediately and candidly, before the court or opposing counsel finds it. In the reported decisions, the harshest outcomes followed attempts to defend or paper over the fabricated authority, not the initial mistake.