AI Hallucinated Citations: Courts Now Sanction Per Citation
Independent Verification Became a Court Rule the Same Month Discipline Stopped Requiring Harm
This article is for educational purposes and does not constitute legal advice. Lawyers and firms with questions about their own filings or professional obligations should consult qualified counsel.
The certification that you verified every citation is itself an unverified claim.
The Unit of Sanction Changed
On July 28, 2026, the Illinois Appellate Court fined an attorney $15,000 in Scott v. Illinois Human Rights Commission. The number matters less than the arithmetic behind it: $1,500 per false citation or misquotation, across four fabricated statutory quotations, one nonexistent case, and five cases cited for propositions they did not support. The court said it set the rate above prior precedent on purpose, because the problem persists across jurisdictions and fines that don't rise don't deter.
The attorney had used ChatGPT and said he'd cross-referenced the output with LexisNexis. The court's answer: verification is source-level. Checking that a case exists in a research platform isn't the same as confirming the quoted text appears in the cited source. When opposing counsel flagged the errors, the reply brief repeated them and called them technical deficiencies. That converted carelessness into willfulness, and the clerk was directed to send the opinion to the Attorney Registration and Disciplinary Commission.
Three years ago sanctions were priced per filing, and the filing was treated as an aberration. This summer they're priced per assertion, and the referral to the disciplinary body is routine.
Three Firsts in Less Than Two Months
The Illinois order (AIEL-2026-060) sits in a cluster. The court orders are indexed at AI Incident Law, the portfolio's public-record corpus of AI matters, under the record IDs noted.
Connecticut, July 31 (AIEL-2026-063). The Connecticut Supreme Court issued its first sanctions order over AI-generated citations in TOV Realty v. Suarez and Kosel Equity v. MacGregor. The attorney told the court he'd done his own research, drafted the briefs, verified his citations, and then pasted the drafts into ChatGPT to improve them. The model added new citations and altered existing ones. Proofreading didn't catch it. The court ordered six hours of continuing legal education, three of them on generative AI, and $1,000 donations from the lawyer and the firm to fund AI ethics education, and it said the order should serve as a reminder to the whole profession. The Superior Court judges had already amended their rules, effective June 23, 2026, to require that anyone using generative AI on a filed document independently verify every citation, authority, or piece of evidence it produced, and the order pointed to that rule.
Washington, July 24 (AIEL-2026-062). In LeDoux v. Outliers, Inc., the Western District of Washington sanctioned counsel $3,000 under Rule 11 for filings containing dozens of hallucinated factual and legal citations. She said she'd believed the two consumer AI tools she used worked in a closed system that referenced only the documents she supplied. The court called her blind trust unfounded, noted that a corrected filing still contained some of the original false citations, and required a certification on every subsequent brief in the case that all citations had been verified. The published takeaway on the Washington State Bar's blog was four words: ignorance is no excuse.
California, July 10 (AIEL-2026-052). The Court of Appeal in Del Biaggio v. Bansen imposed $1,500 in sanctions for fictitious quotations, certified the opinion for publication, referred counsel to the State Bar under the mandatory reporting statute, and ordered him to serve the opinion on his own client and certify that he'd done so.
USPTO, August 27. The Office of Enrollment and Discipline posted its first discipline order predicated on generative AI use, publicly reprimanding a patent attorney whose claim construction chart cited specification passages, figures, and prosecution history that didn't exist. AI Incident Law records it as AIEL-2026-069, In re Brian E. Mitchell, Proceeding No. D2026-16, arising from a joint claim construction chart in a patent case. He'd used one AI tool to draft and a second AI tool to review the draft. He corrected the chart the next day. The district court imposed no sanction. The client suffered no prejudice. The case settled. OED found violations of the competence, diligence, and misrepresentation rules anyway, with no finding of intent, and ordered a notice in the Official Gazette warning practitioners that AI citation error reaches the intrinsic record, where no research platform can catch it.
That last order is the one I'd read twice. The harm threshold is gone. Discipline attached to the behavior, not to the consequence.
August didn't slow the sequence. On August 21 Florida's Second District Court of Appeal, in Capital Standard v. U.S. Bank (AIEL-2026-067), struck two briefs carrying 46 defective citations, made counsel personally liable for the opposing party's appellate fees, and noted that Florida's rules have required since June 15, 2026 an express representation that every cited authority exists and is accurately cited. On August 28 the Northern District of Georgia, in Booker v. Kroger (AIEL-2026-066), priced four false authorities at $1,000 each and doubled the total to $8,000 because counsel had denied using AI when the court asked.
Certification Is a Self-Report
Look at what every remedy converges on. Connecticut wrote independent verification into the rules of procedure. Washington required a per-filing certification. California and Illinois made bar referral automatic. The USPTO reminded practitioners of the duty of reasonable inquiry.
Each of these controls records intent. A certification is the attorney's statement that verification happened. It's signed at the end of the process, under deadline, by the same person whose verification is in question. Nothing about it observes whether the check was performed, how thoroughly, or against what. In Scott, the attorney believed he'd cross-referenced. In Connecticut, the attorney believed he'd verified before the model rewrote his citations. In the USPTO matter, the attorney had run a second model as a reviewer and believed that discharged the duty. All three would have signed the certification in good faith. Most of us would have too.
The profession has spent three years building verification habits that run against Westlaw and Lexis. The USPTO order points at the gap: a wrong column-and-line cite to a specification, a figure number that doesn't match the drawing, a quotation attributed to a prosecution response that never said it. Not what they report, what they do. Self-assessment and behavior are different measurements, and the courts are now sanctioning the difference.
What Verification Under Pressure Looks Like
The sanctioned attorneys weren't careless people in the ordinary sense. Several described a process, and the process failed at predictable points. The points are behavioral.
Treating tool confidence as evidence. A fluent citation with a pinpoint and a parenthetical reads as verified. The Connecticut brief looked polished because the model polished it. Fluency carries no information about accuracy.
Verifying that the source exists rather than what it says. The Illinois attorney confirmed cases existed. He didn't confirm the quoted statutory language appeared in the statute.
Stopping at the first layer. Running a second model over the first model's output is asking the same witness twice. OED treated it as insufficient, and verification that never leaves the model's frame isn't verification.
Not escalating when a source can't be found. In every one of these matters, the false citation survived at least one opportunity to say "I can't locate this" and stop.
Defending instead of correcting. Scott became willful when the errors were repeated after being identified. LeDoux's corrected filing still carried some of the originals.
This is the operational drift sequence in a legal shape: adoption up, output volume up, verification rigor down, false confidence up, human review performative, drift accumulating, liability emerging. The sanctions orders are the last line of that sequence showing up on a docket.
What a Firm Can Observe Today
A firm can't audit its way out of this with certifications, because certifications are the thing that failed. It can observe the behavior directly.
PAICE (People + AI Collaboration Effectiveness) measures how a professional behaves when AI output is uncertain, incomplete, or wrong. The assessment seeds errors of the kind that reached these courts, confident and plausible and false, and scores whether the person catches them, verifies the claim that carries the decision, escalates when a source can't be confirmed, and corrects rather than defends. Someone who sounds fluent but misses the seeded error scores lower than a terse person who catches everything.
The results are structural evidence of verification behavior under pressure, produced before a filing rather than after a show cause order. They're also private by architecture: scores attach to hashed identifiers, no conversation text is retained, and individual results are unavailable to the firm rather than withheld by policy. A firm sees the reliability of its cohort. It doesn't see which associate needs a talk. The associate does.
Courts have decided that the unit of accountability is the citation and that harm is no longer required. The question left for any firm is whether it knows how its people behave when the citation is wrong, or whether it's waiting for the docket to tell it.
Want to assess your team's verification behavior before a court does? Learn about PAICE for organizations or take an individual assessment to see it firsthand.
Recommended Reading
- AI Collaboration for Contract Attorneys - Verification workflows for legal drafting
- Hoping the Model Reads the Statute Right Is Not a Strategy - Why statutory text needs a human reader
- The Cost of Getting It Wrong - What unverified output costs once it leaves the building
- When Polish Impersonates Authority - Fluency as a risk signal
Curious but short on time?
Take the 3-minute PAICE Pulse — a quick confidence check that maps how you see your own AI collaboration posture. No login required.