AI hallucination cases

99.4% success rate for 700 court-confirmed AI citation errors.

We ran Kingsfield against the US case-law portion of Damien Charlotin's AI Hallucination Cases Database: 1,148 real US court cases where lawyers were sanctioned for submitting AI-generated fabrications, narrowed to 700 clearly testable claims. Real lawyers, real judges, real sanctions.

The results, by failure type

504/504
Fabricated citations
case doesn't exist · 100%
92/92
Fabricated quotes
invented quotation · 100%
100/104
Misstated holdings
wrong rule · 96%
99.4%
All types combined
696 / 700 claims
Honest scope. This measures Kingsfield's ruling layer given extracted claims, not an end-to-end "feed it a whole brief" run. The 700 splits into three checks: a cite to a case that does not exist, a quote that was never written, and a real case cited for a holding it does not contain. The first two are clean by construction and run at 100%; misstated holdings is a judgment call and runs at 96%. One quote row the database had mislabeled, a real quote from In re Innovatio, Kingsfield correctly accepted; we also dropped Will v. Michigan, 491 U.S. 58, tagged fabricated but a real Supreme Court decision Kingsfield ruled real.

Every major legal AI tool

The cases below name the tool that produced the fabrication wherever the court record identifies it. Legal-grade research tools sold as accurate produced sanctioned hallucinations alongside the consumer chatbots.

ChatGPTClaudeCoCounsel (Westlaw)Lexis+AILexisNexisvLex / FastcaseCo-PilotCetientCasemine

Browse the cases · 25 ruled by Kingsfield

Isaacs v. Novartis Pharmaceuticals CorporationD. New Jersey · 2025ChatGPTREJECT Williams v. Capital One BankD. DC · 2025CoCounselREJECT Gurpreet Kaur v. Captain Joel DessoN.D. New York · 2025Claude (Sonnet 4)REJECT Coomer v. Lindell/MyPillow, Inc. (1)D. Colorado · 2025Westlaw / Co-PilotREJECT Jackie L. Miller v. Regions BankN.D. Alabama · 2026ChatGPTREJECT Garner v. KadinceUtah C.A. · 2025ChatGPTREJECT David R. Pete v. Greg Abbott, et al.E.D. Texas · 2026CetientREJECT Heriberto Perez-Castillo v. Todd W. Blanche7th Cir. · 2026ChatGPTREJECT Muhammad v. Gap Inc.S.D. Ohio · 2025ChatGPTREJECT Joy Wilson v. KIPP Texas, Inc.N.D. Texas · 2025ChatGPTREJECT Anthony C. Hill v. Workday, Inc. (1)N.D. California · 2025CoCounselREJECT Anthony C. Hill v. Workday, Inc. (2)N.D. California · 2026CoCounselREJECT Shana Jordan, et al. v. Chicago Housing Authority et al.CC Illinois · 2025ChatGPTREJECT Kim Elizabeth Harwell v. WestCare Nevada, Inc.D. Nevada · 2026ChatGPTREJECT Flowz Digital v. Caroline DalalC.D. Cal · 2025Lexis+AIREJECT Sai Malena Jimenez-Fogarty v. Thomas Fogarty et al.S.D. New York · 2026LexisNexisREJECT Brooks and Brooks Inc. v. Patel et al.Hennepin County DC, Minnesota · 2025vLex / FastcaseREJECT Jodell Dodge v. FirstService Residential Arizona LLCD. Arizona · 2025Federally LawyerREJECT Safe Choice, LLC v. City of ClevelandN.D. Ohio · 2025Casemine (Amicus)REJECT Mattox v. Product Innovation ResearchE.D. Oklahoma · 2025ChatGPTREJECT In re: the Marriage of Melinda Johnson v. Sabastian JohnsonCA Indiana · 2025ChatGPTREJECT Erica P. Kyne v. Owen Ehima; In re: Change of Name (Evelyn Ehima Kyne)SC Connecticut · 2026ChatGPTREJECT Harvey v. Torrent Leasing & U.S. BankD. Nevada · 2025ChatGPTREJECT Gustafson v. Amazon.comD. Arizona · 2025tool not identifiedREJECT Gjovik v. Apple Inc.N.D. California · 2025tool not identifiedREJECT

The landmark cases that started it

These predate the slice we tested (the database is roughly 95% 2025 to 2026 filings), so they are field context rather than cases Kingsfield ruled on. They are the ones that put AI hallucinations on the record.

Mata v. Avianca · S.D.N.Y., June 2023 · ChatGPT. The first widely reported sanction. Two lawyers fined $5,000 for six invented cases.
People v. Crabill · Colorado, 2023 · An attorney suspended after filing ChatGPT-fabricated cases and initially blaming a legal intern.
United States v. Cohen (Michael Cohen) · S.D.N.Y., 2024 · Google Bard produced fake cases passed to counsel and filed in a federal motion.

Method

Each citation is checked by its exact reporter location, not its name (a fabrication often reuses a real case name with an invented reporter cite). The oracle is CourtListener, which indexes comprehensive US case law. Fabricated citations were confirmed not found, or found with a name mismatch; fabricated quotes were confirmed absent from the cited opinion. Source records come from Charlotin's public database; the verdicts are Kingsfield's.

Questions

Are these real court cases?

Yes. Every case is a documented filing in Damien Charlotin's AI Hallucination Cases Database, a public catalog of court orders where lawyers were sanctioned or corrected for submitting AI-generated fabrications. The court records are real and public; the verdicts shown are Kingsfield's.

What is an AI hallucination in law?

An AI hallucination in a legal filing is a citation, quotation, or holding that an AI tool produced and that does not exist or does not say what the brief claims. The three common types are a fabricated citation (the case does not exist), a fabricated quote (a real case, an invented quotation), and a misstated holding (a real case cited for a rule it does not support).

How do you check whether a case citation is real?

Kingsfield resolves each citation against a case-law corpus and an independent source (CourtListener), confirms the case name and reporter match, and checks that any quoted text appears in the opinion. If the citation resolves to nothing, or to a different case, it returns REJECT.

Which AI tools produce fake legal citations?

Charlotin's database identifies a range of tools across these cases, including ChatGPT, Claude, Westlaw's CoCounsel, Lexis+AI, LexisNexis, vLex/Fastcase, and others. Both consumer chatbots and legal-grade research tools have produced citations that courts found fabricated.