Key takeaways
- A tracked database lists more than 1,400 filings citing AI-fabricated authority, over 900 of them in US courts.
- The strong, safe uses are bounded and checkable: contract review, due diligence, document comparison, with a human verifying before anything binds.
- Every sanction shares one failure: authority generated rather than retrieved, then filed without verification, from Mata v. Avianca ($5,000) to Lacey v. State Farm (around $31,000).
- Some legal AI tools hallucinate on roughly one in three queries, with the rate rising on harder arguments, exactly when a lawyer most wants a supportive case.
- The duty to verify every citation before filing is non-delegable under Rule 11 and ABA Formal Opinion 512, whether a human or a model produced the draft.
A database maintained by a legal researcher now lists more than 1,400 court cases in which lawyers cited authority that did not exist, fabricated by AI, with over 900 of them in US courts. That number is not an argument against AI in law. Read correctly, it is a precise map: it marks exactly where the technology does real work and exactly where it ends. The boundary between the two is a single discipline, and most of the sanctioned lawyers crossed it the same way.
The work AI actually does in a law firm
Start with the win, because it is real and it is large. Roughly 79 percent of lawyers now use AI in some capacity, and the strongest uses are the bounded ones: contract review, due diligence, document comparison, first-pass extraction of clauses and obligations across thousands of pages. These tasks share a structure. The source material is in front of the model, the output is checkable against it, and a human reviews before anything binds. AI is a fast first-pass reviewer over documents that already exist, and at that job it earns its place.
That is the entire distinction. When the model works from a document you hand it and you verify the result, it is an accelerator. When you ask it to supply authority from its memory and you do not verify, it is a liability with a law license attached.
The sanctions mark the exact line
The cautionary tales are now a genre. It started with Mata v. Avianca in June 2023, when a lawyer submitted six fabricated citations from ChatGPT and drew a $5,000 sanction. It did not stop. In Lacey v. State Farm, a special master imposed around $31,000 after roughly nine of 27 citations in a brief were wrong and at least two cases did not exist, and wrote that the AI output had nearly misled him into adopting the bogus citations in an order. The Sixth Circuit sanctioned two attorneys in Whiting v. City of Athens for, among other misconduct, citing over 24 fabricated authorities. The Alabama Supreme Court sanctioned an attorney who, after being told he had cited a made-up precedent, cited nonexistent cases again in the very next sentence. Even a firm as large as Sullivan & Cromwell had to apologize for a motion with roughly 28 erroneous citations.
Every one of these is the same failure: authority generated, not retrieved, and filed without a check. A Stanford analysis found some legal AI tools hallucinate on roughly one in three queries, with the rate rising on harder arguments, which is precisely when a lawyer most wants a supportive case and is most likely to be handed a fake one.
The altitude shift
Follow one fabricated citation from the prompt to the published record. A lawyer, rushed, asks a model for support and gets a plausible case name with a plausible quote. It goes in the brief unchecked. The opposing side or the judge cannot find it, because it never existed. What was a five-second shortcut becomes a sanctions order with the lawyer’s name, the firm’s name, and the word “fabricated” in a document that is now permanent and searchable. The efficiency was measured in seconds. The cost is measured in a reputation and, increasingly, a referral to the bar.
That asymmetry is why the profession converged fast on a rule. Rule 11, ABA Formal Opinion 512, and a growing set of judicial standing orders all point to the same duty: verify every citation before filing, whether the draft came from a junior associate or a model. The duty to verify is non-delegable, and AI does not absorb it.
The rule that separates the win from the sanction
The operational fix is not to ban AI from the firm. It is to treat the model as a first-pass reviewer of material you can check, never as a source of authority you cannot. Use it to read the contract, compare the clauses, summarize the deposition, surface the obligation buried on page 340. Do not use it to tell you what the law says and then file that answer without opening the reporter. The 1,400 cases are not a story about a bad technology. They are 1,400 lawyers who let a research assistant sign the brief, and the ones doing contract review down the hall, verifying as they go, are not in the database.
The Counter Brief — one email, every Monday.
The week's AI-for-revenue moves in a 5-minute read: which tools are worth the budget and which to skip, plus what to do this week. Source-checked, no vendor decks.
Edited by Aditya Marin Gasga
Free. One click to unsubscribe.
Frequently asked questions
Is it safe for lawyers to use AI?
For bounded, verifiable tasks like contract review, due diligence, and document comparison, AI is widely used and effective, with around 79 percent of lawyers using it in some capacity. The danger is using it to generate legal authority that is then filed without verification, which is what has produced court sanctions.
Why do AI tools invent fake cases?
Language models generate probabilistic text rather than retrieving verified facts, so they can produce plausible-looking case names, quotes, and citations that do not exist. A Stanford analysis found some legal AI tools hallucinate on roughly one in three queries, with higher rates on difficult arguments.
Have lawyers actually been punished for this?
Yes. Courts have imposed fines, public reprimands, and bar referrals, from the $5,000 sanction in Mata v. Avianca in 2023 to roughly $31,000 in Lacey v. State Farm and appellate sanctions in the Sixth Circuit, with more than 1,400 such cases now tracked worldwide.
What is the rule for using AI in legal work safely?
Treat AI as a first-pass reviewer of documents you can verify, never as a source of authority you cannot. The duty to verify every citation before filing is non-delegable under Rule 11 and ABA Formal Opinion 512, regardless of whether a human or a model produced the draft.