It started with one lawyer and a ChatGPT subscription in 2023. By early 2026, the Fifth Circuit is calling the problem one that "shows no sign of abating." Federal appeals courts are issuing sanctions. Am Law 100 firms are getting caught. Twice.
This is the definitive list. Every known case where AI-generated hallucinations made it into court filings, what happened, what it cost, and what it means. We update this tracker as new cases emerge. Bookmark it.
The Case Tracker
Below is every documented instance where a court found (or strongly implied) that AI-generated fabricated content appeared in legal filings. Cases are ordered chronologically.
Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023)
$5,000 SanctionsWhat happened: Attorney Steven Schwartz of Levidow, Levidow & Oberman used ChatGPT to research case law for a personal injury suit against the airline Avianca. The brief he filed contained six entirely fabricated case citations. The cases did not exist. The courts named in the citations never issued the opinions described. When the judge questioned the citations, Schwartz doubled down, submitting additional filings that continued to rely on the fake cases. He even asked ChatGPT to confirm they were real. It said yes.
Outcome: Judge P. Kevin Castel found bad faith, citing "acts of conscious avoidance and false and misleading statements to the Court." Schwartz and co-counsel Peter LoDuca were sanctioned $5,000. Both were referred for potential disciplinary proceedings.
Key takeaway: This was the shot heard round the world. It established that "I didn't know the AI made things up" is not a defense. Attorneys have an independent duty to verify every citation, regardless of the source.
United States v. Cohen, No. 18-CR-602 (S.D.N.Y. 2024)
No Sanctions ImposedWhat happened: Yes, that Michael Cohen. Donald Trump's former personal attorney used Google Bard to find case citations for a motion seeking early termination of his supervised release. Cohen found three cases that looked promising. They were completely fabricated by the AI. He passed them to his attorney, David Schwartz (no relation to the Mata lawyer), who inserted them into a court filing without checking.
Outcome: Judge Jesse M. Furman declined to impose sanctions, finding no bad faith. Cohen was not a practicing attorney and had believed the cases were real. Schwartz, his lawyer, had failed to verify them but was not found to have acted with intent to deceive. The court called the episode "embarrassing" but stopped short of punishment.
Key takeaway: Bad faith matters. The absence of sanctions here was not because the hallucinations were acceptable. It was because the court found the errors were genuinely inadvertent. That distinction has become thinner with every subsequent case, as courts increasingly expect lawyers to know better.
Park v. Kim, 91 F.4th 610 (2d Cir. 2024)
Referred to Grievance PanelWhat happened: A plaintiff's attorney in a Second Circuit appeal used ChatGPT to find case citations for her reply brief. One of the cited cases simply did not exist. The attorney had already received two deadline extensions to prepare the brief, making the lack of verification even harder to excuse.
Outcome: The Second Circuit panel referred the attorney to the court's Grievance Panel under Second Circuit Rule 46.2. The court also ordered her to serve her client with a copy of the opinion. No monetary sanctions were imposed, but a grievance referral is a serious professional consequence that can lead to suspension or disbarment.
Key takeaway: Appellate courts are not going to be gentler about this than trial courts. The Second Circuit's referral to the Grievance Panel signals that AI hallucinations in filings could be treated as a disciplinary matter, not just a sanctions issue.
In re Jackson Hospital & Clinic (Gordon Rees, First Incident)
$55,000 Fee Repayment · Public ReprimandWhat happened: This is where things escalated. Attorney Cassie D. Preston of Gordon Rees Scully Mansukhani, one of the largest law firms in the country (1,800+ lawyers), submitted filings in an Alabama bankruptcy case on behalf of creditor Progressive Perfusion Inc. The filings were riddled with fabricated citations generated by AI. When the court issued an order to show cause, Preston initially denied using generative AI. Gordon Rees later admitted that was false: AI had been used, and Preston knew it.
Outcome: Gordon Rees said it was "profoundly embarrassed." The firm repaid $55,000 in fees and updated its AI policies to include a cite-checking requirement. The bankruptcy judge publicly reprimanded Preston and ordered her to send copies of his opinion to all her clients, opposing counsel, and presiding judges in her pending cases. The firm itself avoided formal court sanctions.
Key takeaway: Two firsts here. First, this was the first Am Law 100 firm to get caught. The scale of the firm made the story impossible to dismiss as a solo practitioner problem. Second, the initial denial of AI use made everything worse. Lying to the court about how the fabrications got there transformed a carelessness problem into a credibility crisis.
Villalovos-Gutierrez v. Pol (Gordon Rees, Second Incident)
ReprimandWhat happened: Just weeks after the Jackson Hospital debacle, Gordon Rees was reprimanded again. This time in a Colorado federal court matter. The filing contained citations that appeared to be AI-generated hallucinations. According to court records, the partner on this case was the same partner involved in the Jackson Hospital matter.
Outcome: The court issued a December 3, 2025 order reprimanding the firm. The full scope of sanctions is still developing as of this writing.
Key takeaway: Once is a mistake. Twice is a pattern. When the same firm, and reportedly the same partner, gets caught with suspected AI hallucinations in a second case within weeks, the "we updated our policies" defense loses all credibility. This is the case that transformed Gordon Rees from a cautionary tale into a case study in institutional failure.
Huynh v. Redis Labs (Gordon Rees, Third Allegation)
PendingWhat happened: In February 2026, opposing counsel in a Northern District of California motion to compel battle alleged that Gordon Rees had once again submitted a brief containing AI hallucinations. The reply brief cited the firm's two prior incidents and argued a pattern of unreliable filings. As of this writing, the court has not yet ruled on the allegations.
Outcome: Pending. But the reputational damage is already done. Above the Law's headline read: "Am Law 100 Firm Accused of Filing Brief Riddled with AI Hallucinations... AGAIN!"
Key takeaway: Reputation compounds. Whether or not the latest allegations hold up, Gordon Rees now faces a situation where every brief it files will be scrutinized for AI artifacts. That is a self-inflicted institutional wound that no policy update can quickly heal.
Jaffer v. Experian et al., No. 25-20086 (5th Cir. 2026)
$2,500 SanctionsWhat happened: Attorney Heather Hersh of FCRA Attorneys (formerly Jaffer & Associates) filed a reply brief in the Fifth Circuit that the court concluded was substantially, if not entirely, drafted by AI. The panel identified 21 fabrications or misrepresentations in the brief. When the court pressed Hersh for an explanation, she was not forthcoming about the AI's role.
Outcome: The Fifth Circuit ordered Hersh to pay $2,500 in sanctions. More significantly, the panel issued a written opinion expressing frustration that the problem "shows no sign of abating," explicitly referencing the growing list of AI hallucination cases across circuits.
Key takeaway: This is the first federal appellate court to impose monetary sanctions specifically for AI-generated hallucinations in a brief. The Fifth Circuit's language signals that appellate courts are running out of patience. The "21 fabrications" finding also shows courts are getting better at spotting AI-generated content, not just isolated fake citations but pervasive patterns of hallucination throughout a filing.
The Pattern: What's Getting Worse and Why
Look at the timeline. In 2023, it was one lawyer who genuinely did not understand what ChatGPT was doing. By 2026, it is Am Law 100 firms, federal appeals courts, and lawyers who absolutely should know better.
Three trends are driving this escalation.
First, AI tools are getting easier to use. The barrier to generating a legal brief with AI has dropped to zero. Any lawyer with a browser can paste a prompt into ChatGPT and get something that looks like competent legal writing in seconds. The output reads well. It follows proper formatting. It cites real-sounding case names with plausible reporter citations. The problem is that "real-sounding" is not "real."
Second, verification takes effort. Checking a citation takes minutes. Checking every citation in a 30-page brief takes hours. When a firm is under deadline pressure (and when is a firm not under deadline pressure?), the temptation to trust the output is enormous. This is a classic automation bias problem: the more polished the AI's output looks, the less likely humans are to scrutinize it.
Third, the tools are marketed as legal research aids. When an AI company positions its product as useful for legal research, lawyers reasonably (if incorrectly) assume it does actual research. It does not. It generates text that resembles research. That gap between marketing and reality is where much of this liability originates.
Court Rule Changes: The Disclosure Wave
Courts have not waited for the profession to self-correct. Since Mata v. Avianca, a wave of standing orders, local rules, and judicial directives has swept across federal and state courts.
The approaches vary, but they cluster around two requirements.
Disclosure mandates. Many courts now require attorneys to certify whether generative AI was used in researching or drafting any filing. The U.S. District Court for the Eastern District of Texas adopted such a requirement in June 2024. Multiple judges in the Southern District of New York have individual standing orders requiring AI disclosure. Similar rules exist in courts across the country.
Verification certifications. Some courts go further, requiring not just disclosure of AI use but affirmative certification that all AI-generated content has been independently verified for accuracy. This is the more robust approach. Disclosure alone does not prevent hallucinations. Verification requirements create accountability.
New York state courts have been particularly active. Multiple judges in Kings County Supreme Court, New York County, and other jurisdictions have adopted specific AI disclosure rules, each with slightly different requirements. As Greenberg Traurig noted in a November 2025 analysis, the diversity of approaches creates a compliance challenge for firms practicing across multiple jurisdictions.
The trend is clear: within the next two years, some form of AI disclosure will likely be standard in most federal courts and many state courts. Firms that do not already have policies in place are behind.
Prevention Frameworks: What Firms Should Implement Now
Based on the patterns across every case in this tracker, here is what a responsible AI use framework looks like for a law firm in 2026.
1. Mandatory Verification Protocols
Every citation in every filing must be independently verified against a primary legal research database (Westlaw, LexisNexis, Bloomberg Law) before submission. This sounds obvious. The Gordon Rees cases prove it is not happening consistently, even at major firms. The protocol needs to be written, trained, and enforced, not just announced in an email.
2. AI Use Logging
Attorneys should document when and how AI tools are used in the preparation of any filing. This creates an audit trail that protects the firm in two ways: it enables compliance with court disclosure requirements, and it provides evidence of good faith if a hallucination slips through despite reasonable precautions.
3. Tiered Review Requirements
AI-assisted drafts should require at least one additional level of review beyond the drafting attorney. For high-stakes filings (appellate briefs, dispositive motions, anything going to a federal court), consider requiring senior attorney sign-off specifically on citation accuracy.
4. Training That Goes Beyond "Be Careful"
Most firm AI policies amount to "use AI responsibly." That is not a policy. It is a platitude. Effective training explains why LLMs hallucinate (they are not doing research; they are generating statistically probable text), shows attorneys what hallucinated content looks like, and walks through real examples from the cases in this tracker.
5. Tool Selection and Configuration
Not all AI tools carry the same hallucination risk. Legal-specific AI tools that use retrieval-augmented generation (RAG), grounding their outputs in actual case law databases, are significantly safer than general-purpose chatbots. Firms should evaluate tools based on their hallucination rates in legal contexts, not just their marketing materials.
6. An Honesty-First Response Plan
The Gordon Rees saga teaches one lesson above all others: if AI hallucinations are discovered in your filing, do not lie about it. Cassie Preston's initial denial that AI was used turned a bad situation into a catastrophic one. Have a response plan that prioritizes immediate candor with the court, voluntary withdrawal of affected filings, and proactive disclosure to clients.
Malpractice Implications
Every case in this tracker also carries potential malpractice exposure. When a lawyer files a brief containing fabricated citations, the client suffers harm. At minimum, the client's credibility with the court is damaged. In some cases, the client's legal position is compromised because the arguments built on fabricated authorities collapse entirely.
The malpractice analysis is straightforward. An attorney has a duty of competence. Filing a brief without verifying its citations, regardless of whether AI or a summer associate generated them, breaches that duty. The standard of care requires verification. The cases in this tracker establish that courts expect it. A malpractice plaintiff does not need to prove that the attorney intended to file fake citations. Negligence is enough.
For law firm insurers, this is a growing risk category. The frequency of incidents is increasing (look at the timeline above), and the damages can be substantial. A client who loses a case, or receives a worse outcome, because their lawyer's brief contained fabricated authorities has a real claim.
The malpractice risk is especially acute for solo practitioners and small firms. Large firms like Gordon Rees can absorb the reputational hit and the $55,000 fee repayment. A solo practitioner sanctioned $5,000 and referred for disciplinary proceedings, like Schwartz in Mata, faces potentially career-ending consequences.
The Expert Witness Angle
AI hallucination cases are creating a new category of expert testimony. Courts and litigants need witnesses who can explain, in technically precise but accessible terms, why the fabrication happened and whether the deployer's practices met the standard of care.
In my work as an expert in these matters, the technical questions I am most frequently asked include:
Was the content AI-generated? This is the threshold question. Courts are getting better at recognizing AI-generated text (the Fifth Circuit identified 21 separate fabrications in the Hersh brief), but the determination often requires technical analysis. AI-generated legal text has characteristic patterns: plausible but nonexistent case names, reporter citations that follow correct formatting but reference fake volumes, and legal reasoning that sounds authoritative but does not trace to any actual opinion.
Was the hallucination foreseeable? This goes to negligence. If the attorney used a general-purpose LLM for legal research, hallucination was not just foreseeable but virtually certain. The hallucination rates of general-purpose models for specific citations and factual claims are well documented. An expert can testify about baseline hallucination rates and explain why a particular deployment was, or was not, reasonable.
Could it have been prevented? This is the standard of care question. If RAG-based legal research tools were available and the attorney chose a raw chatbot instead, that choice is relevant. If the firm had no verification protocol, that absence is relevant. The expert's role is to establish what a reasonably competent attorney or firm should have done, given the known risks of the technology.
What Comes Next
The Fifth Circuit's February 2026 opinion contains a line that should concern every attorney who uses AI: the problem "shows no sign of abating."
That is not just judicial frustration. It is a prediction. And the data supports it.
The cases are accelerating. One in 2023. A handful in 2024. A major firm caught in 2025. Multiple federal court incidents in early 2026. As AI tools become more embedded in legal practice, the opportunities for hallucinations to reach court filings will only multiply.
Several developments are likely in the near term.
Sanctions will increase. The progression from $5,000 in Mata to the Fifth Circuit's $2,500 (on top of the underlying sanctions in that case) suggests courts are still finding their footing on appropriate penalties. As the "I didn't know" excuse becomes less plausible, expect sanctions to climb.
Disciplinary proceedings will become common. The Second Circuit referred the Park v. Kim attorney to its Grievance Panel in 2024. Expect more referrals as bar associations develop specific guidance on AI use. Several state bars have already issued ethics opinions on the topic.
Malpractice claims will follow. We have not yet seen a wave of malpractice suits arising from AI hallucinations, but the foundation has been laid. Every sanctioned attorney in this tracker is a potential malpractice defendant. It is only a matter of time before clients pursue those claims.
Insurance will adapt. Legal malpractice insurers will begin asking about AI use policies during underwriting. Firms without robust AI governance frameworks may face higher premiums or coverage exclusions.
The technology itself is also evolving. Newer models hallucinate less frequently than their predecessors, and RAG-based legal research tools are improving rapidly. But "less frequently" is not "never." And the core problem remains: LLMs generate plausible text, not verified facts. Until that fundamental architecture changes, the risk persists.
The question is no longer whether your firm uses AI. It is whether your firm has the safeguards to catch AI's mistakes before they reach a judge's desk.
This tracker will be updated as new cases emerge. If you are aware of a case we have not covered, contact us.
The Criterion AI provides expert witness services and litigation support for matters involving artificial intelligence, machine learning, and algorithmic decision-making. For a confidential consultation on an active or anticipated matter involving AI hallucinations or related liability, contact us at info@thecriterionai.com or call (617) 798-9715.