Lawyer Fined for Using AI-Generated Legal Documents with Fake Citations

A lawyer is sanctioned for using AI when the lawyer signs a filing containing authority the AI invented and nobody verified it. Courts worldwide have logged more than 2,100 decisions involving authority an AI invented or misstated. Where an outcome was recorded, 379 drew a monetary sanction and 162 drew a disciplinary referral.

This article covers what gets a lawyer sanctioned and what the penalties in the record are. It then sets out the three defect types behind them, which attorneys have been penalized, what courts now require before a filing goes out, and whether blaming the tool has worked. The pattern reaches law firms running high volumes of third-party paper as much as it reaches litigators, because the failure underneath is ungrounded generation rather than anything specific to a brief.

[cta-1]

Key takeaways

1. Lawyers are sanctioned for failing to verify AI output, not for using AI.
2. Monetary sanctions were imposed in 379 recorded decisions and disciplinary referrals in 162.
3. The most common penalty band in the record is 1,000 to 5,000 dollars, which holds 86 of the 191 US-dollar penalties logged against all filers rather than lawyers alone.
4. Fabricated cases are only one of three failure modes, alongside misrepresented holdings and false quotations from real authority.

What gets a lawyer sanctioned for using AI

What gets a lawyer sanctioned for using AI is the certification, not the tool. The duty sits in Rule 11 of the Federal Rules of Civil Procedure, which provides that by signing a filing, an attorney certifies that the legal contentions in it are warranted by existing law after an inquiry reasonable under the circumstances. A citation the lawyer never opened fails that inquiry whether a junior associate, a contract attorney or a language model produced it. Courts imposing sanctions have located the breach at the signature rather than at the software.

Three layers of authority reach the same place:

1. Rule 11 was never amended for generative AI and did not need to be, because the certification it requires does not depend on what produced the text.
2. Many state courts apply a close analogue, so the duty follows the signature in state practice as well as in federal practice.
3. The American Bar Association's Formal Opinion 512 reaches the same conclusion from the ethics side, treating competence, candor to the tribunal and supervision as the operative duties rather than creating a separate AI rule.

The reasonable inquiry, in practice

What a reasonable inquiry means is narrower than it sounds, and it is what the sanctioned filings skipped:

1. Pull the full text of every cited authority from a reporter or a citator, not from the generated summary.
2. Read the pinpoint page the proposition is attributed to, because a real case can still fail to say what the brief claims.
3. Check that the authority is still good law and has not been reversed, vacated or superseded.
4. Record who performed the check, because the signing attorney carries the duty regardless of who did the work.

The supervision point

The duty is non-delegable in the sense that matters: a partner who signs a brief drafted by a junior associate, a contract attorney or a vendor has certified it personally.

How much are lawyers fined for AI fake citations

How much a lawyer is fined for AI fake citations depends on the court, and the record clusters in a narrow band. The 1,000 to 5,000 dollar band holds 86 of the 191 US-dollar penalties logged in the AI Hallucination Cases database, against 37 in the 5,000 to 10,000 band, 33 below 1,000, 29 between 10,000 and 50,000, and 6 above 50,000. The database records penalties against all filers, including self-represented litigants, rather than against lawyers alone.

The penalties that reach the headlines sit in the band holding 6 of the 191 recorded US-dollar penalties, so the widely reported figures come from the rarest recorded outcome in the record of AI hallucination cases. The most common recorded penalty is an order of magnitude lower.

The penalty bands in the record

What else the courts imposed

The money is not the whole exposure, and the financial penalty is often not the most consequential part. The record carries non-financial outcomes alongside the fines, such as:

1. A disciplinary referral to the relevant bar, logged in 162 decisions, which is the route by which a sanction becomes a licensing problem.
2. Disqualification from the matter, which hands the client a new counsel search mid-litigation.
3. Mandatory continuing legal education on AI use, which a court can attach as a condition of the sanction.
4. An order to pay the opposing party's attorney fees, which runs separately from the sanction itself.
5. A written opinion naming the lawyer and the firm, which is permanent, searchable and quoted by opposing counsel afterwards.

Three types of AI errors in legal citations

Fabricated cases are the most familiar type of AI error in legal citations, and they are not the only one. The AI Hallucination Cases database records 3,527 fabricated items, 1,744 misrepresented items and 1,065 false quotes drawn from real authority. Misrepresentation and false quotation together account for a little under half of the recorded defects, and both survive a check that confirms only that the case exists.

The three types fail differently, and only the first one is caught by a citator check. Stanford's RegLab found that purpose-built legal AI models still hallucinate on one in six or more benchmarking queries, so the defect rate is a property of the category rather than of one consumer tool.

Fabricated authority

The cited case, statute or docket number does not exist. This is the defect that produced the early sanctions and the one a citator catches in seconds, because the lookup simply returns nothing.

Misrepresented holdings

The case is real and the proposition attributed to it is not. A citator confirms the case exists and is good law, and reports nothing about whether it says what the brief claims. Catching this requires reading the opinion at the pinpoint cite.

False quotations from real authority

The case is real, the holding may even be close, and the quoted sentence appears nowhere in the opinion. These are the hardest of the three to catch, because the citation, the court and the holding can all check out while the sentence in quotation marks was never written. Only a text comparison against the opinion catches it.

What each check catches

The practical consequence for a verification workflow is specific:

1. A citator check catches fabricated authority and nothing else.
2. A pinpoint read catches a misrepresented holding.
3. A string search of the opinion text catches a false quotation.
4. None of the three is caught by re-reading the brief, because the defect sits in the authority rather than in the prose.

Which attorneys have been sanctioned for using AI

Attorneys have been sanctioned for using AI in courts across the United States, and a handful of those decisions set the markers other courts cite. Amir Mostafavi was fined 10,000 dollars by a California appellate panel after 21 of the 23 quotations in his opening brief turned out to be fabricated. Two California law firms were ordered to pay 31,100 dollars in Lacey v. State Farm. Richard Bednar was sanctioned by the Utah Court of Appeals over a brief citing a case that did not exist.

Amir Mostafavi, California Court of Appeal

The panel fined Mostafavi 10,000 dollars after finding that 21 of the 23 quotations in his opening brief were fabricated, in what CalMatters reported appears to be the largest penalty a California court has issued over AI fabrications. He told the court that he did not read the generated text before filing it. He told CalMatters that he had written the appeal himself and then used ChatGPT to improve it, without knowing that it would add citations.

Lacey v. State Farm, Central District of California

Two law firms were ordered to pay 31,100 dollars over what the court described as bogus AI-generated research in a brief submitted to a special master. The judge noted that he had come close to citing the fabricated material in a judicial order of his own, which is the systemic risk the sanctions regime is aimed at.

Richard Bednar, Utah Court of Appeals

Bednar's brief cited Royer v. Nelson, a case that does not exist, and the court ordered him to pay the opposing party's attorney fees, refund the fees his client had paid for the filing, and donate 1,000 dollars to a legal non-profit. The brief had been drafted by an unlicensed law clerk. That did not reduce the sanction, because the signature was Bednar's.

The Wyoming Walmart filings

Three lawyers were fined 5,000 dollars in total over fabricated citations in a product liability case against Walmart, with one of them fined 3,000 dollars and removed from the case, as Reuters reported at the time. The case is a useful counterpoint to the headline sanctions because the firm self-reported and withdrew the filing, and the penalty landed in the modal band rather than above it.

[cta-2]

Mata v. Avianca: the case that started the count

Mata v. Avianca was the first widely reported sanction over AI-fabricated citations, and it remains the reference point courts use. Two New York lawyers and their firm were fined 5,000 dollars in June 2023 after a filing cited cases ChatGPT had invented. It was one of six decisions logged in its quarter. Courts now log several hundred a quarter.

The full account of the hearing, the lawyers' explanations and the court's reasoning is set out in our piece on the New York lawyers who submitted a ChatGPT-drafted filing. The New York Times covered the sanctions hearing in detail at the time.

What courts now require before a filing goes out

The court rules for lawyers using AI in filings fall into three layers: the signature rules that already applied, the standing orders individual judges have issued, and the certification language some courts now require in the filing itself. Rule 11 and its state analogues were never amended for AI and did not need to be. The standing orders add a disclosure step on top, and they vary by judge rather than by jurisdiction, which is why the check belongs in the filing workflow.

The standing orders

No reliable public count of federal judicial standing orders on AI exists, because the orders are issued judge by judge and the trackers that aggregate them sit behind paywalls. What can be said without a number is that checking the assigned judge is now a standard step rather than a cautious one, and that the orders are not uniform. The orders take several recurring shapes:

1. Disclosure orders, which require a filing to state whether generative AI was used in preparing it.
2. Certification orders, which require a statement that a human verified every citation and quotation.
3. Prohibition orders, which bar generative AI from drafting the filing at all.

The verification step

The verification steps that catch these defects are specific:

1. Pull and read every cited authority in a citator before the filing goes out, including the ones carried over from an earlier brief.
2. Compare every quotation against the opinion text, not against the case name, because a real case can carry an invented sentence.
3. Check the assigned judge's standing order at the point the matter is assigned, not at the point of filing, so a prohibition does not surface after the draft is written.
4. Record who verified what, with a name against each authority, so the certification has something behind it.
5. Treat a brief drafted by a junior associate, a contract attorney or an outside vendor as unverified until someone reads the authorities.
6. Re-run the check when a filing is recycled, because an authority that was good law when the original brief went out may not be now.

The supervision, disclosure and billing layer

Three questions sit outside the citation check and reach the engagement rather than the filing:

1. Supervision: the firm's AI policy has to reach everyone who touches a draft, including contract attorneys and outside vendors who never sign anything.
2. Client disclosure: whether a client is told that generative AI was used in the matter is a function of the engagement letter and the client's own policies, and in-house clients increasingly set terms on it.
3. Billing: time saved by a tool can raise a reasonableness question about what is billed, which ABA Formal Opinion 512 addresses directly.
4. Records retention: an order requiring certification implies keeping the verification record for the life of the matter, not for the life of the draft.

The broader shift in how this work gets staffed is covered in our overview of the changes AI is driving across legal practice.

Can a lawyer blame AI for fake citations

A lawyer cannot blame AI for fake citations, because the duty sits on the signature rather than on the software. Of the 763 decisions recording how the filer responded, 94 blamed a tool, 59 blamed a junior and 41 doubled down on the fabricated authority. Withdrawing or correcting the filing was recorded 397 times and an apology 383 times.

The filers' own explanations are worth reading, and 404 Media collected eighteen of them in a piece on why the lawyers caught using AI say they did it. The recurring account is not recklessness. It is a filing deadline, an unfamiliar tool treated as a search engine, and an assumption that something which produced a correctly formatted citation had retrieved it from somewhere.

Bednar's case shows what happens to the delegation version of the argument. The brief was drafted by an unlicensed law clerk, which is a real supervisory failure, and the Utah Court of Appeals sanctioned Bednar anyway. The signature does not move, which is the certification duty seen from the other direction.

What can and cannot be delegated

The distinction that holds is between the work and the responsibility:

1. Drafting can be delegated to a junior, a vendor or a tool.
2. Verification can be delegated, provided someone is named against each authority.
3. The certification cannot be delegated at all, because it is an act of the signing attorney.
4. The record of who checked what belongs with the matter file, because it is the only evidence the inquiry happened.

What the record does not support is the stronger claim that AI is displacing the judgment involved, a question taken up in our piece on whether generative tools stand in for a lawyer.

Where the same failure shows up in contract work

AI hallucinations in contract drafting and review come from the same failure as the fabricated citation: a general-purpose model generating fluent language with nothing underneath it. Contract disputes are the single largest legal field in the hallucination record, accounting for 454 of the 1,863 decisions with a field recorded.

Spellbook's Review feature works from a team's own playbooks and from a set of 20 million real contracts rather than from a model's general training. It flags risks and proposes redlines as native Track Changes in Word, which a lawyer accepts or rejects. The retrieval runs against the team's positions and the benchmark set, so the output points at clause language that exists rather than at language the model composed.

Where it is not the right tool

Spellbook specializes in commercial contracts:

1. It does not check litigation citations and it is not a citator.
2. It would not have caught any of the briefs described above.
3. The failure mode is shared between contract work and filings; the remedy is not transferable between them.

A firm that wants fabricated authority caught needs a citator and a verification step in the filing workflow, which is a different control from a grounded drafting tool.

Frequently asked questions

Can you get in legal trouble for using AI?

Yes, a lawyer can get in legal trouble for using AI, though the trouble follows from filing unverified output rather than from using the tool. The sanctions in the record are imposed under the signature rules that already governed filings, which means the exposure attaches at the point of certification. Using a tool to draft and then verifying the authorities is not sanctionable conduct in any of the logged decisions.

Is it legal for attorneys to use AI?

Attorney use of AI is permitted in the United States, and the ethics authorities that have addressed it treat the question as one of competence and supervision rather than permission. ABA Formal Opinion 512 and the state bar opinions that followed it set conditions on confidentiality, candor and billing rather than prohibiting the tools. The conditions, and where the bar opinions diverge, are set out in our guide to the permissibility question for generative tools in practice.

Can a lawyer be disbarred for using AI?

Disbarment is possible in principle and rare in the record. The route runs through the disciplinary system rather than the court imposing the sanction: 162 of the logged decisions recorded a referral to the relevant bar, and what the bar does with a referral is a separate proceeding with its own range of outcomes. The far more common result is a monetary sanction in the low thousands, sometimes with a disqualification or a continuing legal education condition attached.

Do I have to tell the court I used AI?

It depends on the assigned judge, which is why the check belongs in the filing workflow. Undisclosed AI use is one of the three failure patterns courts have penalized, alongside fabricated citations and failure to verify, and a number of judges have issued standing orders requiring either disclosure of AI use or a certification that a human verified every citation. The practical answer is to read the standing order when the matter is assigned.

[cta-3]

Final thoughts

The filings that drew penalties failed at verification rather than at the choice of tool. They were not unusual in their use of a tool; they were unusual in reaching a court without anyone opening the authorities. The duty that produced them predates the tools and was not amended for them, so the control that prevents the outcome is a workflow step rather than a technology choice.

In contract work the equivalent control is grounding: an output traceable to a playbook position or a real precedent is one a lawyer can check. Try Spellbook free to see how the Review feature works against your own playbooks.

> About this content: This article is provided for informational purposes only. It does not constitute legal advice, and reading it does not create an attorney-client relationship. For legal advice on a specific situation, consult a licensed attorney in your jurisdiction. Spellbook is a software product; it supports legal professionals' work but does not replace their judgment. References to AI model capabilities reflect publicly available information as of the publication date and may change.