The Sanctions Wave: What 800 Cases Actually Tell Us About AI in Legal Practice

The numbers have crossed a threshold most lawyers have not noticed

In June 2023, Judge P. Kevin Castel sanctioned two New York attorneys and their small firm $5,000 for filing a brief with six fabricated citations generated by ChatGPT. The case was Mata v. Avianca. Everyone in the profession heard about it. Most of us assumed it was an aberration.

It was not. It was the opening shot.

As of early April 2026, Damien Charlotin's HEC Paris database tracks approximately 1,317 AI hallucination cases globally. Roughly 800 of them are U.S. matters. The pace has gone from a handful of cases in 2023, to about two per week in 2024, to two or three per day by spring 2025, to a reported 17 U.S. court decisions in a single day on March 31, 2026. Bloomberg Law tracked 52 improper-AI-use rulings in February 2026 alone.

The peak single sanction has gone from $5,000 in Mata to roughly $110,000 in Couvrette v. Wisnovsky (D. Or., April 4, 2026). That is a 20-fold escalation in under three years. Q1 2026 alone produced approximately $145,000 in aggregate sanctions nationwide.

And in April 2026, something happened that the profession has not fully processed yet. Judge Anna Manasco in the Northern District of Alabama, for the first time in U.S. history, imposed a direct Rule 11 sanction on a law firm itself — not just the individual attorney. The firm. The case was Rivera v. Triad Properties. The firm was Burrill Watkins LLC. The total exposure was $47,056.90 plus disqualification plus mandatory notification of every client the firm has.

Three doctrinal phases in 34 months

The U.S. AI-sanctions case law has moved through three distinct phases. Understanding which phase we are in matters for how firms should be thinking about liability.

That is the arc. Individual-only, to firm-protected-by-policy, to firm-directly-liable. In 34 months.

Six attorneys disqualified. One federal prosecutor fired.

The monetary numbers are not the most consequential part of the doctrine. The career consequences are.

At least six attorneys have been disqualified from active cases as an AI-hallucination sanction:

• Rudwin Ayala in Wadsworth v. Walmart — pro hac vice revoked (Feb 2025).
• Maren Bam in Mavy v. Commissioner of SSA — pro hac vice revoked, removed as counsel of record, letters sent to the three Arizona federal judges she had falsely attributed citations to (Aug 2025).
• Matthew Reeves, William Cranford, and William Lunsford at Butler Snow — disqualified from Johnson v. Dunn with multijurisdictional bar referrals (July 2025).
• Joshua Watkins in Rivera v. Triad Properties — disqualified along with his firm (April 2026).

The cases where the court declined to sanction

Before any attorney assumes that sanctions are automatic when AI hallucinations appear in a filing, the no-sanction cases have to be read carefully. They are the defense playbook.

In Iovino v. Michael Stapleton Associates (W.D. Va., July 2024), attorney Thad Guyer of the Government Accountability Project filed a brief with two nonexistent cases and two fabricated quotations generated by a GPT-based cite-checker. Judge Cullen discharged the show-cause order, crediting Guyer's prompt unconditional responsibility, his remedial commitment to Westlaw verification plus junior-attorney cite-checking, and his unblemished record. "Mr. Guyer, to his credit, owned the mistake." The Virginia State Bar subsequently issued a private reprimand — so avoiding court sanction does not mean avoiding bar discipline. That distinction matters.

In United States v. Cohen (S.D.N.Y., March 2024), Judge Furman declined to sanction David Schwartz despite three nonexistent case citations generated by Google Bard, finding Schwartz's conduct "certainly negligent, perhaps even grossly negligent" but not in bad faith. In Hall v. Academy Charter School (E.D.N.Y., August 2025), Magistrate Judge Wicks declined to sanction Suryia Rahman after her law clerk's Google research produced three fake citations — citing recent bereavement, documented mental-health treatment, and full acceptance of responsibility. And in Al-Hamim v. Star Hearthstone, LLC (Colo. App., December 2024), the Colorado Court of Appeals expressly distinguished Kruse: the pro se appellant acknowledged AI use, apologized, and his brief mixed real and fabricated citations rather than being overwhelmingly fabricated.

Rivera v. Triad Properties and why firm-level doctrine is now decisive

Every firm that is paying for an AI tool, or encouraging its attorneys to use one, needs to read Judge Manasco's April 7, 2026 opinion. Not a summary. The opinion.

Joshua Watkins of Burrill Watkins LLC filed briefs with AI-hallucinated citations. He characterized the errors as "inadvertent technical error." He blamed court staff and opposing counsel. He filed "corrected" submissions that continued to mislead. He had been separately admonished contemporaneously in a state-court matter. Judge Manasco concluded his conduct was "tantamount to bad faith." That part of the ruling is not novel.

What is novel is what she did next. She turned to the firm:

Mr. Watkins's misconduct did not occur in a vacuum. The court also has serious concerns about Burrill Watkins's apparent lack of internal controls and guardrails surrounding its attorneys' use of artificial intelligence, indeed, the very AI the firm pays for and encourages its attorneys to use. Though Burrill Watkins maintains that it acted swiftly to remediate Mr. Watkins's errors, which it says it had no reason to know about, the firm has not explained how it enforced any policies about responsible AI use, how it will prevent improper AI use going forward, or any other circumstance, let alone an extraordinary one, why it shouldn't be sanctioned.

And then she closed the firm's ignorance defense permanently:

Indeed, although Burrill Watkins's purported lack of knowledge has to do with a partner's misuse of a relatively new technology, its core argument about its own ignorance is both old and ordinary.

Old and ordinary. Those three words are the ballgame.

Same judge. Johnson v. Dunn in July 2025 released Butler Snow without sanction because Butler Snow had a 2023-vintage AI policy, an AI Committee, a 50-docket internal review, and an independent Morgan Lewis audit. Rivera in April 2026 sanctioned Burrill Watkins directly because Burrill Watkins could not document any of those things. Same supervisory-duty analysis under Rules 5.1 and 5.3. Opposite outcome. The distinguishing fact is whether the firm had governance it could point to.

This is now the rule. Firms that can prove governance get protection. Firms that cannot face direct monetary exposure. The "we didn't know" defense is dead.

What a firm actually needs to have in place

Butler Snow's governance infrastructure cost the firm nothing in Johnson v. Dunn despite three of its attorneys being disqualified. Burrill Watkins's absence of governance cost the firm $47,056.90, its continued participation in the case, and the obligation to tell every client the firm has what happened.

For a small or mid-sized firm, that math is not close. Building a documented AI governance program is cheaper than a single $47,000 sanction.

Beyond fabricated citations: the categories emerging now

Fabricated citations are the category everyone reads about. They are not the only category. Four others are generating their own jurisprudence.

What this means for the small and mid-sized firm

1. The assumption that AI sanctions are a big-firm problem is wrong. Pro se appellants have been sanctioned ($10,000 in Kruse, case-dismissed in Mendones). Solo and small-firm practitioners have been sanctioned (Colorado's People v. Crabill — one-year-and-a-day suspension). The Am Law 100 firms that have been sanctioned almost without exception had better governance to fall back on than most small firms have. The risk profile for the small and mid-sized firm is higher, not lower.
2. The defense playbook is available — but only if the firm has the documentation to use it. Prompt self-disclosure requires someone at the firm to have flagged the problem before the court did. Credible remedial safeguards require those safeguards to exist. Every element of the Iovino / Cohen / Hall / Al-Hamim playbook assumes preexisting governance most firms do not have.
3. The cost asymmetry is now lopsided. Building a documented governance program — written policy, firm-wide training, enforcement mechanisms, incident response, supervisory Rule 5.1/5.3 accountability, and ongoing audit — costs a fraction of a single Rivera-sized sanction. And after Rivera, the sanctions include mandatory client notification. Try explaining to your best client why they are receiving a copy of a federal court opinion naming your firm for AI misconduct. Try explaining the same thing to your malpractice carrier at renewal.

The 800 cases are not a warning anymore. They are the rule. The next phase of the case law, likely emerging in the second half of 2026, will be about firms that saw Rivera and did not build governance in time.

The bottom line

Eight hundred U.S. cases. One hundred and ten thousand dollars in the largest single sanction. Six disqualifications. One federal prosecutor fired. And now, as of April 7, 2026, the first direct Rule 11 sanction against a law firm itself, with every client learning what happened.

If the firm is using AI without a written policy, it is not a question of whether a court sees the gap. It is a question of when.