On Monday, June 8, 2026, a federal judge in Mississippi cancelled a civil trial and removed all four lawyers from the case after both sides filed briefs citing case law that did not exist.
Senior U.S. District Judge Sharion Aycock fined the four attorneys a combined $8,000, barred the two primary drafters from her court for two years, and wrote that the episode showed "the risk associated with serving as a rubber-stamp," according to coverage in the Mississippi Free Press and BNN Bloomberg.
That order, in Withers v. City of Aberdeen, is the clearest 2026 signal that AI hallucinated citations in court filings now carry personal, monetary, and career-ending consequences for the lawyer who signs the brief.
This is a practitioner's field guide to what just happened, how fake case law gets caught, and the verification workflow that keeps it off your filings.
The short answer: An AI hallucinated citation is a fabricated case, quote, statute, or pin cite produced by a generative model that a lawyer files without checking. It is sanctionable not because AI was used, but because the lawyer's non-delegable duty to verify every authority was skipped.
TL;DR
- A federal judge cancelled a trial in June 2026 and sanctioned four lawyers $8,000 total for AI-fabricated case law.
- Two lawyers were barred from the district for two years; all four were removed from the case.
- The 2026 base penalty is a $1,000 to $3,500 fine plus a mandatory AI ethics course.
- Courts now use cheap tools (Westlaw zero-result, Google Scholar, citators) to catch phantom cites.
- The fix is a documented ten-step cite-check log, not a blanket AI ban.
Key takeaways
- "Ignorance of the risks of AI usage is no longer an excuse," a D.D.C. Judge wrote in June 2026.
- Damien Charlotin's database logged 1,623 hallucination cases worldwide as of June 17, 2026.
- Supervising partners face the heaviest exposure, even when they did not draft the AI text.
- ABA Formal Opinion 512 maps existing ethics rules onto AI; it does not require disclosure.
- One Oregon attorney accrued $109,700 in combined AI-related sanctions across multiple filings.
What happened in the Mississippi case?
Withers v. City of Aberdeen (No. 1:24-cv-00218, N.D. Miss.) is a breach-of-contract dispute over unpaid legal fees tied to a solar-power project. A bench trial was set for March 2026. Judge Aycock cancelled it before it began.
The detail that makes the case a teaching tool: both sides did it. As Aycock wrote, "this case presents the court with an unusual scenario, attorneys for both litigants engaged in similar sanctionable conduct," per The New York Times, which first reported the order on June 9.
All four lawyers admitted at a January 2026 show-cause hearing that they used generative AI and never checked the output. The court found each in violation of Rule 11 of the Federal Rules of Civil Procedure, plus a local rule on resident-attorney supervision.
| Attorney | Role | Fine | Additional sanction |
|---|---|---|---|
| Kathleen M. Wilson | Plaintiff's pro hac vice counsel | $2,500 | Two-year bar from N.D. Miss. |
| Shauncey Hunter Ridgeway | Plaintiff's local counsel | $1,000 | Removed from the case |
| Kathryn Y. Williams | City's counsel (Christian & Small partner) | $3,500 | Two-year bar from N.D. Miss. |
| Mark McClinton | City's local counsel sponsor | $1,000 | Removed from the case |
Note the heaviest fine went to the senior partner. The court treated her supervisory role as aggravating, not mitigating. That is the line every law-firm AI policy should internalize: the partner who fails to check AI-assisted work carries the most risk.
Secondary coverage describes Withers as reportedly the first federal case to sanction all four lawyers on both sides simultaneously. That superlative is reported, not independently verified against PACER. The first AI-hallucination sanction of any kind remains Mata v. Avianca in 2023.
How fast are these sanctions spreading?
The pattern is no longer a few unlucky solos. It runs through every level of practice.
Mata v. Avianca (S.D.N.Y., 2023) set the template when Judge P. Kevin Castel fined two lawyers $5,000 each for six fabricated opinions, as the Guardian reported. The Second Circuit later affirmed.
Three years on, the volume is staggering. Damien Charlotin's AI Hallucination Cases Database recorded 1,623 cases worldwide as of June 17, 2026, growing by roughly 350 to 400 new decisions per quarter.
A Stanford analysis of a 114-case U.S. Sample found 90 percent of offending filings came from solo or small firms, and about half involved ChatGPT variants. The underlying Stanford CodeX/RegLab study reported that even specialized legal AI tools hallucinated 17 to 34 percent of the time on tested queries, while general-purpose ChatGPT ran 58 to 88 percent on the same benchmarks.
Read those as Stanford-found ranges from one study, not universal constants.
The 2026 escalation is visible in the penalties themselves:
- N.D. Miss. (June 2026): Withers, four lawyers, $8,000, two bars.
- Fifth Circuit (Feb 2026): a Dallas attorney fined $2,500 for 16 fabricated quotations.
- D. Or. (2026): one attorney accumulated $109,700 in combined sanctions across multiple filings.
- S.D.N.Y. (April 2026): Sullivan & Cromwell apologized to Chief Judge Martin Glenn for roughly 28 erroneous citations in a Chapter 11 emergency motion.
- D.D.C. (June 2026): a $1,000 fine plus a three-hour AI course, with the judge writing that "ignorance of the risks of AI usage is no longer an excuse."
Big Law is not immune. Plaintiff or defense, civil or bankruptcy, the tariff applies.
How do courts detect AI-generated citations?
Detection has become a craft, and most of the tools are free.
AI tell-phrases. "As an AI language model," "as of my last update," and "I cannot browse the web" still slip into filings. Judicial clerks report these appear in a noticeable and rising share of new civil filings.
Bluebook anomalies. Wrong reporter abbreviations, fabricated volume or page numbers, and parallel citations that do not parallel are cheap tells. AI usually gets the reporter right and the page wrong.
The zero-result test. The simplest check is typing the citation into Westlaw or Lexis. A growing number of judges do this in chambers and cite the empty result on the record.
Phantom and mismatched quotes. Sometimes the case exists but says nothing like what the brief claims. Opposing counsel now systematically feeds suspect briefs into the PelAIkan reference checker and the Charlotin database as a deliberate cross-check.
Citator treatment. A case offered for a proposition it was overruled or distinguished on, with no acknowledgment of the negative history, is increasingly a flag rather than a footnote.
The most reliable detection layer is the lawyer on the other side. Assume your brief will be run through a checker.
Are there court rules requiring AI disclosure?
Not uniformly. There is no national rule, and the trend is toward mandatory verification rather than mandatory disclosure.
ABA Formal Opinion 512 (July 29, 2024) does not impose a per-se disclosure rule. It maps six existing Model Rules onto generative AI: competence (1.1), confidentiality (1.6), communication (1.4), candor toward the tribunal (3.3), supervision (5.1/5.3), and fees (1.5). It does require specific informed consent before client data goes into a self-learning tool.
The Eastern District of Virginia moved early, in January 2024, requiring AI-use identification and a citation-accuracy certification in pretrial orders. Similar standing orders now exist in the Northern District of California, S.D.N.Y., D.C., Massachusetts, Delaware, and others.
A tracker-based estimate puts it at roughly one in four federal districts, which is directional rather than precise.
States track the ABA. The Florida Bar's Opinion 24-1 and NYC Bar Formal Opinion 2024-5 require verification and informed consent. On May 1, 2026, California's COPRAC proposed amendments that would, for the first time, expressly add a duty to verify every cited authority, naming AI-assisted citations in the rule text.
The honest answer for a client: it depends on the court and the bar, and the direction is toward yes.
A ten-step verification workflow for AI legal research
The cure is a documented cite-check protocol. This distills ABA, NCSC, and state-bar guidance into steps any associate or paralegal can run before filing. The National Center for State Courts frames the whole thing around one principle: human-in-the-loop, where a person owns the output.
- Pinpoint the citation. Log the court, year, reporter, and pin cite for every cite in a Cite Verification Log attached to the matter.
- Run the commercial citator. Check KeyCite or Shepard's. Anything other than clean history must be addressed or the cite removed.
- Cross-check free databases. Pull the same cite in CourtListener, RECAP, and Google Scholar Case Law. These are what judges use in chambers.
- Verify reporter and page. Confirm abbreviation, volume, and page against the opinion. The page is the cheap tell.
- Read the cited page. Open the PDF. Confirm the quote appears and the holding actually supports your proposition. Never trust an AI summary here.
- Go primary for statutes. Pull current text from Cornell LII or GovInfo. AI frequently cites repealed or renumbered sections.
- Go to the eCFR for regulations. Confirm section numbers and effective dates fresh each time.
- Follow the local rule on unpublished opinions. Confirm the circuit permits citing that authority.
- Run the smell test. Scan every cite against the AI tell-phrase list. One "as an AI language model" means the draft is not ready.
- Document and sign. Record the cite, sources consulted, verifier's initials, and date. Keep the log with the file.
If the court ever asks who verified a citation, that log is your answer.
What this means for you
Whose fault is it? Primarily the lawyer's. Vendor design choices contribute, and the Stanford finding that specialized legal tools still hallucinate 17 to 34 percent of the time shows the engineering problem is unsolved.
But the lawyer signs the brief and owes the duty of candor, and as of June 2026 no federal safe harbor for "certified" tool use exists.
Are blanket bans the answer? In litigation practice, no. The ABA, NCSC, and state bars all endorse supervised adoption over prohibition. A ban just pushes AI use into the shadows and onto less-vetted tools.
A short note on tools, dated because this field moves monthly. As of June 17, 2026, the frontier releases lawyers are most likely using include Anthropic's Claude Opus 4.6 and Google's Gemini 3.1 Deep Think, with OpenAI continuing to ship custom-tuned models for legal partners like Harvey.
Treat any vendor "we cut hallucinations by X percent" claim as marketing until it cites a peer-reviewed benchmark. Which model you pick matters less than the verification protocol you run on its output.
The lawyers at risk in 2026 are the ones who file AI output without checking it and cannot produce a Cite Verification Log when the court asks. Build the log, run the ten steps, and the next hallucination case will not be yours.
Sources
- Judge Punishes 4 Lawyers After Catching Both Sides Using A.I. (New York Times)
- Courts cracking down on error-strewn AI-assisted legal briefs (TechXplore)
- AI Hallucinations in Law Firms: What Lawyers Must Know (2026)
- Two US lawyers fined for fake ChatGPT citations (The Guardian)
- AI Hallucination Cases Database (Damien Charlotin)
- ABA issues first ethics guidance on AI tools (Formal Opinion 512)
- Guidance for implementing AI in courts (National Center for State Courts)
- NYC Bar Formal Opinion 2024-5: Generative AI in the Practice of Law