A May 2025 declaration filed by Latham & Watkins in Concord Music Group v. Anthropic included a supporting citation whose listed title and authors did not match the linked source. The error entered the court record after opposing counsel flagged the discrepancy. Latham, a firm that routinely bills over $2,000 an hour for partners and counts Anthropic as a client, had used Anthropic’s Claude to format the citation.
The mistake unfolded in a specific sequence. A Latham colleague located an academic source via Google Search. Attorney Dukanovic provided the correct URL to Claude and asked the model to format a proper legal citation. Claude returned a citation that showed the correct publication year and a working link but supplied the wrong title and wrong authors. Team members verified that the link resolved, did not detect the metadata mismatch, and the declaration was filed as submitted.
Crucially, Claude did not fabricate a nonexistent document. Instead it identified a real web resource and misdescribed the embedded metadata. That failure mode is harder to spot than an entirely fabricated citation because surface checks — such as confirming a URL resolves or matching a year-can appear satisfactory while the citation’s descriptive elements remain inaccurate.
The court treated the filing as a professional lapse rather than a harmless glitch. The judge ordered explicit disclosure when AI tools are used and imposed human — verification steps for future filings in the case. The incident raises questions under Rule 11 of the Federal Rules of Civil Procedure, which requires attorneys to certify that factual contentions have evidentiary support and that filings are not submitted for improper purposes.
The episode echoes an Eastern District of Texas decision from late 2024, Gauthier v. Goodyear Tire & Rubber Co., in which a brief contained citations to two nonexistent cases and fabricated quotations generated by Claude. After a show-cause order, the plaintiff’s lawyer admitted using Claude without verifying its outputs; the court imposed a $2,000 penalty and required continuing legal education on AI in practice.
For law firms and tool builders the case underscores concrete needs to manage AI-assisted drafting. Recommended safeguards include provenance and metadata validation for citation tools; automated cross — checks that compare fetched document metadata to generated citations; audit logs that record when and how AI suggestions were used; and clear internal policies requiring human verification before filing documents or signing declarations. Because large language models can produce fluent but confidently wrong metadata in milliseconds, technical and workflow protections matter not as theoretical best practices but as practical measures to reduce exposure under Rule 11 and preserve the integrity of court filings.
Sources
Replies (0)
No replies in this topic yet.
