1. The Crime Scene (The Symptom)
A filing submitted to the U.S. District Court for the Southern District of New York cited three Second Circuit decisions. Two cases exist but were mischaracterized; one citation paired a real case name with a non-existent reporter pin cite and invented procedural holding. The errors originated from a consumer AI chatbot used to "format and strengthen" the motion.
Unlike total fabrication, this failure mode is more dangerous: real case names lend false credibility while holdings and pin cites drift from the canonical text.
2. The Artifact (Raw LLM Output)
"United States v. Cohen, 939 F.3d 167 (2d Cir. 2019) — establishes that
pro se litigants may not rely on AI-generated citations without independent
verification by counsel of record."
"See Park v. Kim, 91 F.4th 212 (2d Cir. 2024) — the Second Circuit held that
district courts must sua sponte strike AI-fabricated authorities even where
no party objects."
"United States v. Schwimmer, 892 F.2d 237 (2d Cir. 1989) — sanctions for
frivolous filings apply with equal force to generative-AI-assisted pleadings."
3. The Evidence Breakdown (The Core Contradiction)
Claimed Authority: United States v. Cohen, 939 F.3d 167 (2d Cir. 2019).
The Reality Check: United States v. Cohen exists — but at 797 F.3d 115 (2d Cir. 2015), and concerns tax evasion sentencing, not AI citation duties. The reporter cite 939 F.3d 167 is wrong; the holding is invented.
Claimed Authority: Park v. Kim, 91 F.4th 212 (2d Cir. 2024).
The Reality Check: Federal Reporter 4th series volume 91 does not contain this page for any Park v. Kim disposition in 2024. The case name may echo unrelated docket captions; the authority is unsupported.
Claimed Authority: United States v. Schwimmer, 892 F.2d 237 (2d Cir. 1989).
The Reality Check: Schwimmer exists and concerns frivolous appeals — but the opinion does not mention generative AI. The LLM imported a contemporary policy concern into a 1989 sanctions framework.
4. The Dali Verification Check
Dali semantic pass on Cohen: authority_exists=true, proposition_supported=false. The case is real; the attributed rule is not in the opinion text. This is the failure class most citation-link validators miss.