The first reported AI hallucination sanction in a US court arrived in 2023. As of May 29, 2026, this site tracks 363 sanctioned matters: an attorney filing contained AI-generated material and a court issued a final order. (Pending matters where a show-cause is open but no final ruling has issued are tracked separately.) This post reads the docket sideways. Patterns emerge by tool, by error type, by jurisdiction, by court action, by dollar amount. Most are unsurprising once seen in aggregate.
Cases are the unit of analysis, not the lawyers. Names of individual attorneys appear only as links to the underlying case page, where the public record speaks for itself. Our framing is what firms can do to stay out of this database. It is not who already is in it.
Which AI tools appear most often
ChatGPT dominates the docket. The aggregator counts 50 sanctioned-matter appearances in which an order names ChatGPT specifically as the tool that produced the offending text. Appearances by every other named tool sit in low single digits: 8 Claude, 5 Gemini, 7 Microsoft Copilot, 10 Westlaw AI Assistant or CoCounsel, 2 Perplexity. A small number of orders name two or more tools. Those filings are counted once per named tool, not allocated to a single bucket. A handful of pending matters add Vincent AI and others that do not yet appear in the sanctioned-matter count.
Most filings name no tool at all. In 271 matters, the order identifies the output as AI-generated by inference. The signature pattern is fabricated case names, invented reporters, and citation errors that no human lawyer would produce. The vendor is not on the record.
Two readings of the tool distribution are worth holding at once.
Read it as product-share first. ChatGPT is the consumer tool a generalist lawyer reaches for in a hurry, so its over-representation reflects raw market exposure rather than any defect specific to that model. Legal-grade vendors (Westlaw AI Assistant, Lexis+ AI, CoCounsel, Vincent, Harvey) appear less because they have less raw market share among solo and small-firm filers. Less raw share, less raw exposure.
Read it the other way too: the legal-grade vendors are not immune. Vincent AI appears in Creech v. City of Raleigh, a North Carolina Court of Appeals matter pending as of May 29, 2026. Counsel cited two fictitious cases generated by Vincent AI, a tool marketed for legal research; counsel filed a motion to amend after the citations were noticed. The case is the first NC appellate test of how to handle AI-hallucinated filings. Whatever marketing language a vendor uses, the duty under ABA Formal Opinion 512 sits with the filing attorney.
A firm AI policy that bans only ChatGPT misses the pattern. The risk is not the brand. It is the practice of citing a quotation, a holding, or a case caption without verifying the underlying source.
What kinds of errors trigger sanctions
The dominant error is the fabricated citation. A non-existent case is paired with a real legal proposition, and the filing attorney either did not check or did not know to check.
Adjacent errors visible across the docket:
- Invented quotations from real cases. The case is real. The quote is not. Westlaw and Lexis return the case as valid, so the quote check requires opening the opinion itself.
- Misattributed holdings. Real case, real holding, wrong proposition. Frequently produced when a tool summarizes an opinion the filing attorney has not read.
- Mismatched parallel citations. An accurate F.3d cite paired with a fabricated state-reporter cite. Bluebook formatting is technically correct, which is the giveaway. Human lawyers rarely pair accurate federal authority with a fabricated state pin cite.
- Fake statute or rule numbers. Less common but persistent. Filings argue rules of court that do not exist or cite statutes the legislature never enacted.
The error type that is conspicuously rare across the docket is the substantive misreading of an opinion the lawyer actually read. Models are weaker at retrieval than at analysis. Once given the right authority, the analysis is usually serviceable. The lesson, then: AI is not unfit for legal work. It is unfit for legal research in the narrow sense of locating authority that actually exists.
Two cases illustrate the retrieval failure. Mata v. Avianca is the canonical example. The filing cited six fabricated cases, opposing counsel raised the issue, and the SDNY issued a Rule 11 sanction. Park v. Kim is the Second Circuit follow-up that put attorneys on notice that the duty to verify is not new law. It is the existing duty under Federal Rule 11 applied to a new tool.
Where in the country it happens
Filings landed in courts spread across federal and state systems. The aggregator distinguishes the federal districts, state courts, and bankruptcy and specialty venues separately. The headline split: 271 matters in federal court, 92 in state court.
Top concentrations by state of the court of record. Only state courts and single-state federal districts are counted. Circuit cases that bind multiple states are excluded so a single appellate sanction does not inflate every state in the circuit:
- California with 55 matters
- New York with 40 matters
- Florida with 26 matters
- Texas with 19 matters
- Illinois with 17 matters
Read this list with two adjustments. First, raw filing volume drives raw sanction volume. The Southern District of New York and the Central District of California sit at the top of every federal docket statistic, not just AI ones. Second, a state’s bar rules shape the visibility of the issue. States with formal AI ethics guidance (see the California opinion, New York Part 161, Florida Bar Opinion 24-1, Texas Opinion 705) make the duty explicit. Sanctions in those states are less ambiguous because the rules are public.
The state that is conspicuously missing from the top of the list is also informative. Pennsylvania has been propagated across aggregator sites as having a “statewide August 2024 disclosure mandate.” We have not located a primary source for that order. Until one surfaces, treat the claim as unverified. Aggregator consensus is not validation.
What courts have done about it
Courts have taken every action available to them, with one notable exception.
What courts have done:
- Rule 11 monetary sanctions are the most common remedy. Cash penalties issue under Federal Rule 11 and parallel state-court inherent authority. 326 matters in the dataset took this path.
- 28 U.S.C. § 1927 fee-shifting is rarer but heavier. Courts award opposing counsel’s reasonable fees under Section 1927 on findings of unreasonable and vexatious conduct. Fee awards can dwarf any flat sanction.
- Striking the offending filing. Briefs disappear from the record. The case proceeds on what remains.
- Mandatory CLE. Several courts have ordered offending attorneys to complete remedial training on AI before further filings.
- Bar referrals. 15 matters resulted in referral to state bar disciplinary counsel. Discipline can include suspension, public reprimand, and required restitution.
- Adverse inferences and credibility findings. Several judges have found that a fabricated cite undermines the credibility of every other representation in the filing. Downstream effects on the merits are harder to quantify.
- Pre-filing injunctions. Some orders bar offending attorneys from filing further documents in that court without pre-clearance. This is the most procedurally serious remedy short of formal discipline.
One remedy is conspicuously absent: disqualification from the case. Courts have stopped short of removing the attorney as counsel, even where the conduct supports it. Reasoning usually runs that disqualification disserves the client, who did not make the AI error.
For ongoing court-level guidance, the court orders tracker lists the standing orders judges have issued requiring AI disclosure or verification on every filing.
What it costs
Among sanctions matters that include a parseable dollar amount, here are the headline numbers:
- Cases with a stated monetary penalty: 116
- Total assessed across those cases: $2,894,812
- Median: $2,500
- Largest single sanction: $1,671,560
Three reasons that total is a floor, not a ceiling.
First, many sanctions orders shift opposing counsel’s fees without quantifying the dollar figure inside the order. Fee awards land later on a supplemental affidavit and frequently do not appear in the order text the aggregator parses. Second, bar discipline (suspensions, restitution, referral fees) carries economic costs that do not map cleanly to a single dollar number. Third, the parser sums every dollar amount written into the sanction string. Where the order specifies ”+ $X each” per attorney without naming the count, the per-attorney multiplier is not applied, so totals understate multi-attorney joint-and-several sanctions.
A rough ladder, in ascending severity:
- Public admonishment, no monetary sanction. Court goes on the record but does not move money.
- Flat sanction in the $1,000 to $5,000 range. Most common shape; $2,500 is the median.
- Fee-shifting under Section 1927. Variable amount; five figures or more on contested motions is normal.
- Bar referral with disciplinary outcome. Suspension, public reprimand, and ancillary fees vary by state. Insurance and licensing consequences sit on top.
- Pre-filing injunction or vexatious-litigant designation. Procedural restriction; cost shows up as lost clients and inability to file in that court.
Insurance carriers reading the docket have begun pricing this exposure into renewal underwriting. That pattern is the subject of Why Your Malpractice Carrier Is Starting to Ask About AI.
Practical reading
A managing partner who has read this far is asking the right question. Where does my firm sit in the distribution, and what would it cost to land in it?
Three lessons emerge from the docket.
One: the duty is old, even if the tool is new. Rule 11 has always required reasonable inquiry into citations. These cases prosecute the same duty against a new factual pattern. Firms with clean Rule 11 records need to extend existing verification practice to AI-assisted work, not invent a new compliance regime from scratch.
Two: disclosure rules vary by court. Getting disclosure wrong is a separate sanction risk from the underlying citation error. Our court orders tracker lists disclosure regimes by judge and by court. Filings that comply with the underlying duty but violate a disclosure rule still draw an order.
Three: the policy artifact matters at renewal even more than the policy itself. Firms with a written AI policy, an attorney acknowledgment file, a verification log, and an incident-response procedure have the documentation a malpractice carrier asks for. Firms that have only the practice but no documentation carry the same actual risk and a worse renewal posture. Our Carrier-Renewal Packet ships those nine artifacts pre-customized. For firms preferring a guided rollout, consulting runs vendor interviews, training, and carrier-submission prep.
A note on what is in the dataset
All case data is sourced from public court records: the federal PACER system, individual state court electronic filing portals, and the CourtListener RECAP archive. Case captions, judges, dates, and outcomes are checked against the docket entry before publication. Where a primary source is unavailable, the matter is omitted from this post and flagged on the case page.
Reversals propagate. If a sanction is later vacated, modified, or withdrawn on appeal, the case page updates and the aggregate counts in this post update on the next deploy. Manual editing of these counts is a regression. The dataset reflects what the public record shows on May 29, 2026.
No part of this post lists, ranks, or characterizes the conduct of any individual attorney. Across the database, framing is neutral reporting from court records under the fair-report privilege. Readers who believe a case page contains a factual error should send a correction notice; on receipt we verify against the docket and update the page promptly.
This article is for informational purposes only and is not legal advice. Verify all citations against primary sources, including the underlying court orders, before relying on them.