Insights•3 min read

Citations verified in the draft, then altered by an AI rewrite, reached the court

October 12, 2026 by Asif Waliuddin

Evidence WatchWatchlistVerification DebtAI Governance

Where fabricated legal citations get caught

On July 31, 2026, the Connecticut Supreme Court issued an order in TOV Realty, LLC v. Suarez / Kosel Equity, LLC v. MacGregor, 355 Conn. 902. The attorney verified the citations in his drafts. Then he pasted the text into ChatGPT to improve the writing. The AI pass added or altered citations, and about seven erroneous citations reached the court. The verification happened. It happened on the wrong version.

The Ninth Circuit drew the line in a published order on June 3, 2026, in Lnu v. Blanche, No. 24-4790. The briefs there had "multiple nonexistent cases, misattributed quotations, and gross misrepresentations of real cases". The court said where the failure sits: "the rules are not violated at the point of research and drafting, but at the point of signing and filing."

The scale signals point the same way. Damien Charlotin's AI Hallucination Cases database shows "2180 cases identified", last updated 9 October 2026. A preprint by Zhao et al. (arXiv:2605.07723) audited 111 million references across 2.5 million papers and found "a sharp rise in non-existent references following widespread LLM adoption", with "a conservative estimate of 146,932 hallucinated citations in 2025 alone".

I ran a small test of my own. I wrote two synthetic paragraphs and scanned them with Faultline (CLI 0.11.1, Gemini provider). Both fabricated citations came back "contradicted", including "Delmont v. Arcadia Freight Systems", a case I invented that has never existed anywhere. The second run also produced a false pass: a sentence fragment containing the fake case name was marked "supported" on an unrelated 1992 case. So my rule is to read every contradicted line. One contradicted citation means the sentence fails, whatever else is green. The product owner traced the false pass to the engine splitting text at every period, which cut the case name into fragments and checked one fragment alone. That is fixed in the hosted version as of October 12, 2026: re-running my second input there returns the fake citation contradicted and no fragment rows, while the npm CLI 0.11.1 I tested keeps the split until its next release. One gap is still open: my holding sentence about the fake Ninth Circuit case still reads supported, because extraction checks the holding apart from the fake case as a generalization, and the product owner is building a guard so a claim from the same sentence as a contradicted citation cannot read supported.

That directly reinforces the Trust Laundering argument. A citation check that ran on a draft says nothing about the text that was signed. The check has to run on the final version, because that is the version someone puts their name to and files.

There is counterevidence worth stating. None of this says AI drafting is unusable. The Ninth Circuit's own line is that AI use is not the violation; it put responsibility at signing and filing.

What remains unproven: the Charlotin database is a curated corpus, not an incidence rate, and it includes some decisions where AI use was alleged but not confirmed. The Zhao et al. figures come from a preprint, not a peer-reviewed paper. My own test is n = 2 with one false pass. It is a demonstration, not an accuracy rate.

Source: Lnu v. Blanche, Ninth Circuit order; TOV Realty v. Suarez / Kosel Equity v. MacGregor, Connecticut Supreme Court order; AI Hallucination Cases database; Zhao et al., arXiv:2605.07723

What I would test next: run the same citation check on a verified draft and on its AI-polished final version, and record which citations change between the two.