{"id":856,"date":"2026-05-06T15:23:12","date_gmt":"2026-05-06T07:23:12","guid":{"rendered":"https:\/\/connectword.dpdns.org\/?p=856"},"modified":"2026-05-06T15:23:12","modified_gmt":"2026-05-06T07:23:12","slug":"when-claude-hallucinates-in-court-the-latham-watkins-incident-and-what-it-means-for-attorney-liability","status":"publish","type":"post","link":"https:\/\/connectword.dpdns.org\/?p=856","title":{"rendered":"When Claude Hallucinates in Court: The Latham & Watkins Incident and What It Means for Attorney Liability"},"content":{"rendered":"

There is a particular kind of irony that the legal profession rarely gets to witness in such pristine form. In May 2025, Latham & Watkins a firm that routinely bills over $2,000 an hour for its partners and counts Anthropic among its clients filed a court declaration in Concord Music Group v. Anthropic<\/em> that contained fabricated citation details. The citations weren\u2019t invented by a sleep-deprived associate pulling an all-nighter. They were generated by Claude, the very AI model that Latham & Watkins was in court defending.<\/p>\n

Sit with that for a moment.<\/em><\/p>\n

The lawyer arguing that Claude is not a copyright infringement machine used Claude to format a legal citation in an active case and Claude got the authors wrong, the title wrong, and nobody caught it until opposing counsel started digging. The irony isn\u2019t just delicious. It\u2019s instructive. Because what happened inside that filing is a near-perfect X-ray of the structural problem that AI poses for legal practice: not that AI is obviously wrong, but that it is convincingly, plausibly, professionally wrong<\/em> in ways that evade even experienced human review.<\/p>\n

The Anatomy of the Incident<\/strong><\/h3>\n

To understand the legal exposure, you have to understand exactly how the error happened \u2014 because it wasn\u2019t sloppy. It was systematic.<\/p>\n

The sequence was this: a Latham colleague found what appeared to be a supporting academic source via Google Search. Dukanovic then asked Claude to format a proper legal citation for that source, providing the correct URL. Claude returned a citation with the right publication year and the correct link \u2014 but the wrong title and wrong authors. When the team ran its manual citation check, they verified the link resolved correctly. They didn\u2019t verify whether the metadata Claude generated for that link was accurate. The declaration containing those errors was filed. Opposing counsel noticed. A federal court got involved.<\/p>\n

What makes this technically significant is that Claude didn\u2019t hallucinate a phantom source \u2014 it found a real one and then misdescribed it<\/em>. This is actually harder to catch than a completely fabricated citation, because the URL resolves, the paper exists, the year is right. The error is embedded at the level of metadata, not existence. It\u2019s the legal equivalent of citing a real statute from the wrong jurisdiction.<\/p>\n

The court subsequently mandated explicit disclosure of AI usage and human verification requirements for future filings. The judge\u2019s response was proportionate but pointed. This wasn\u2019t dismissed as a technical glitch. It was treated as a professional failure.<\/p>\n

Rule 11 and the Architecture of Attorney Responsibility<\/strong><\/h3>\n

Here is where the incident stops being a story about one firm and becomes a structural question about the entire profession.<\/p>\n

Rule 11 of the Federal Rules of Civil Procedure requires attorneys to certify with their signature that every factual contention in a filing has evidentiary support and that the filing is not submitted for improper purposes. That signature is not ceremonial. It is a professional representation that the lawyer has exercised reasonable diligence to verify what they\u2019re putting before the court.<\/p>\n

The problem is that Rule 11 was written for a world where fabrication required intent or gross carelessness. An attorney who invented a case citation was either lying or catastrophically negligent. But Claude doesn\u2019t fabricate with intent. It fabricates with confidence. The output is formatted, fluent, properly punctuated, and returned in milliseconds. There is no stylistic tell, no hesitation marker, no signal that the model is operating at the edge of its competence. The professional-looking wrapper of the output is precisely what makes it dangerous.<\/p>\n

In Gauthier v. Goodyear Tire & Rubber Co.<\/em>, decided in the Eastern District of Texas in late 2024, a plaintiff\u2019s attorney submitted a brief containing citations to two nonexistent cases and multiple fabricated quotations also generated by Claude. When the court issued a show-cause order, the attorney admitted he had used Claude without verifying the output. The sanctions were relatively light: a $2,000 penalty, mandatory CLE on AI in legal practice, and an obligation to share the order with his current employer. But the court\u2019s reasoning was unambiguous \u2014 the attorney\u2019s professional obligation under Rule 11 does not diminish because an AI generated the content. The verification duty doesn\u2019t transfer to the machine. It stays with the lawyer.<\/p>\n

This is the constitutional core of the problem. Rule 11 requires certification. Certification requires diligence. Diligence requires verification. But verification, in the context of AI-generated legal content, is no longer a routine proofreading exercise. It is a technical competency task one that many practitioners are neither trained for nor culturally inclined to perform.<\/p>\n

The Duty of Competence in the Age of Plausible Output<\/strong><\/h3>\n

The American Bar Association\u2019s Model Rule 1.1 requires attorneys to provide competent representation, which includes \u201cthe legal knowledge, skill, thoroughness, and preparation reasonably necessary for the representation.\u201d The ABA\u2019s 2012 Comment 8 to this rule added that competent lawyers must \u201ckeep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology.\u201d<\/p>\n

That comment, added over a decade ago, was understood at the time to cover things like e-discovery software and encrypted communications. Nobody was thinking about large language models. But in 2025, it has become the operative text for a new genre of malpractice exposure.<\/p>\n

The key word in Comment 8 is risks<\/em>. A competent attorney using AI is not simply one who knows how to prompt Claude effectively. It is one who understands the category of errors that AI produces, the conditions under which those errors become more likely, and the verification protocols necessary to catch them. The Latham incident illustrates precisely the gap between surface-level AI literacy (\u201cI know how to ask Claude to format a citation\u201d) and functional AI competence (\u201cI know that Claude can confidently return metadata errors on real sources, and I have a protocol to catch that\u201d).<\/p>\n

Importantly, nearly 75% of lawyers cited accuracy as their biggest concern about AI tools, according to the ABA\u2019s 2024 Legal Technology Survey Report. But concern and competence are not the same thing. The same survey suggested widespread adoption of AI tools for legal research and drafting despite that stated anxiety which means lawyers are, collectively, worried about something they are proceeding to use anyway, without necessarily developing the specific verification skills that would address their worry.<\/p>\n

What \u201cVerification\u201d Actually Has to Mean Now<\/strong><\/h3>\n

The Latham incident exposes a gap in how legal professionals conceptualize verification. Traditionally, checking a citation meant confirming the case exists, pulling the relevant page, and reading the quote in context. These are tasks that reinforce themselves the act of going back to the source is itself the verification.<\/p>\n

But when an attorney asks Claude to format<\/em> a citation for a source they\u2019ve already found, the psychological dynamic shifts. The attorney has already done what they consider the hard work (locating the source). The formatting feels clerical. And because Claude\u2019s output looks correct proper journal style, accurate URL, plausible author names it doesn\u2019t trigger the cognitive alarm bells that an obviously wrong answer would.<\/p>\n

This is the hallucination failure mode that is hardest to engineer around: not the fabricated phantom, but the plausible misdescription. It requires a category of verification that legal training does not currently emphasize \u2014 what you might call metadata verification<\/em>: independently confirming not just that a source exists, but that the specific descriptive claims the AI makes about that source (authorship, title, publication date, journal name) match the actual document, line by line.<\/p>\n

For AI-generated legal citations specifically, this means: retrieve the source independently, not via the AI\u2019s link; cross-reference the author names against the byline in the original document; verify the title character by character; confirm the journal name against the masthead. It is slower than proofreading. It requires more discipline than clicking a link. And in a profession that bills by the hour and prizes efficiency, it introduces friction that firms may be reluctant to institutionalize.<\/p>\n

The Deeper Irony: Defending AI with AI<\/strong><\/h3>\n

There is a dimension to this story that legal commentary has largely let pass without examination: Anthropic was the defendant. Claude was the product at issue. The lawsuit brought by Concord Music Group and other major music publishers in October 2023 alleged that Anthropic had scraped copyrighted song lyrics to train Claude without authorization. The case was, at its core, a dispute about whether Claude\u2019s training process respected intellectual property law.<\/p>\n

And Latham & Watkins, in defending that case, used Claude to help prepare court filings and those filings contained AI-generated errors that required court intervention.<\/p>\n

This is not merely ironic. It is epistemically significant. It suggests that even the legal teams most deeply embedded in AI litigation, most knowledgeable about AI\u2019s limitations, most incentivized to use AI carefully in an AI-adjacent case, are still susceptible to the same verification failures that are sanctioning less sophisticated practitioners across the country. If Latham cannot build a reliable AI verification protocol into a high-stakes case where the AI\u2019s own maker is the client, the profession-wide challenge is considerably larger than bar associations and law school curricula have yet acknowledged.<\/p>\n

What Comes Next And What Needs To<\/strong><\/h3>\n

Courts are responding, unevenly. Several federal districts have issued standing orders requiring disclosure of AI-assisted drafting in filings. The court in Concord Music<\/em> mandated both disclosure and human verification. Judge Michael Wilner of California fined a law firm $31,000 after finding that nearly a third of the citations in a brief were AI-fabricated. These are not isolated disciplinary incidents \u2014 they are the early shape of a new jurisprudence around AI professional responsibility.<\/p>\n

What the profession needs, and does not yet have in any systematic form, is a technical taxonomy of AI failure modes translated into verification protocols. Rule 11 compliance in the AI era cannot be satisfied by a general instruction to \u201cdouble-check AI output.\u201d It requires attorneys to understand, specifically, that: reasoning models hallucinate at higher rates on document summarization than standard models; that metadata errors are less visually salient than phantom-citation errors; that confidence in AI output is not correlated with accuracy; and that formatting tasks carry as much hallucination risk as drafting tasks.<\/p>\n

Law schools are not teaching this. Bar associations are issuing guidance at a pace that lags the technology by roughly eighteen months. And law firms are deploying AI tools in client-facing work while building verification protocols that are, at best, adaptations of pre-AI proofreading habits.<\/p>\n

The Latham & Watkins incident will likely be remembered as the most clarifying data point of 2025 for AI and legal practice not because it involved a rogue actor or a spectacular failure, but because it was ordinary. It was a competent attorney, at an elite firm, using a capable AI tool, in a plausible way, and producing an error that the whole team missed. That ordinariness is the point. The question the profession must now answer is not whether AI will create liability exposure for attorneys. It already has. The question is whether the response will be serious enough to match the risk.<\/p>\n

\n

Sources<\/strong><\/p>\n