ChatGPT. A fabricated Nevada case with reporter numbers borrowed from real ones.
Kim Harwell, a pro se plaintiff and military veteran with PTSD and anxiety, sued WestCare Nevada and several staff over how the treatment facility handled a hostile roommate and her reports about it. She brought claims for discrimination and retaliation under the Rehabilitation Act, negligence, and others. The defendants moved to dismiss or for a more definite statement.
In support of her negligence claim, Harwell's complaint cited Estate of Saila v. Circle K Corp., 135 Nev. 545, 353 P.3d 541 (Nev. Ct. App. 2019), with a parenthetical saying it imposed liability for "systemic failures to protect vulnerable individuals." The court could not find any case by that name.
Checking the citation, the court found the volume and page numbers pointed to entirely unrelated cases. 135 Nev. 545 led to State of Nevada v. Inzunza, and 353 P.3d 541 led to Dutton v. City of Midwest City, an Oklahoma case. The court concluded it appeared Harwell had used generative AI to prepare the complaint.
Offered for: a negligence proposition
The court granted the motion in part on the merits, dismissing several claims with leave to amend and ordering Harwell to clarify which defendants she sued for negligence. In a dedicated section it warned the parties against citing nonexistent caselaw, reminded them of their duty under Rule 11(b)(2), and noted that citing fake AI-generated cases may violate the rule. It imposed no sanction but cautioned that future violations could result in penalties or other directives under Rule 11(c)(4).
Kingsfield returned REJECT. The court could not find the case, and the citation numbers corresponded to unrelated decisions.
This matter is among the US case-law cases we ran Kingsfield against in the Charlotin benchmark. Kingsfield's existence check confirms every citation against the case-law corpus and an independent source (CourtListener) before any claim resting on it can stand. Across the run Kingsfield ruled 696 of 700 testable claims correctly, including every one of the 504 fabricated citations and 92 fabricated quotes.
Honest scope. This measures Kingsfield's ruling layer given an extracted citation, not an end-to-end run over a full brief. The case record and the AI-tool attribution come from Damien Charlotin's public database and the court order; the verdict is Kingsfield's. See the full 700-claim benchmark.