Can an AI Detector Catch a Fake Legal Citation Before It Gets You Sanctioned?

Quick answer

On September 3, 2026, the D.C. Court of Appeals struck a brief filed on behalf of Deutsche Bank because it cited four cases that don't exist. The attorney, Loishirl Hall, admitted she'd used Google's AI search tool to find case law and never checked whether the cases were real before filing. She's not an outlier. A public database tracking these incidents has logged over 2,000 court decisions worldwide involving fabricated or misrepresented AI-generated legal citations, and it grows by several new cases a week. An AI-writing detector is genuinely useful here, but only for one specific failure mode. It can flag whether a document's prose was likely drafted by AI, which tells a reviewer where to look harder. It cannot tell you whether a citation inside that prose points to a real case. Those are two different problems, and conflating them is exactly how firms like the one representing Deutsche Bank end up in front of a disciplinary panel.

What actually happened to Deutsche Bank's lawyers

The case is Douglas v. Deutsche Bank National Trust Co. In June 2026, the D.C. Court of Appeals discovered it couldn't locate several of the case citations in the bank's appellate brief and ordered the firm to show cause why the brief shouldn't be struck for "citing nonexistent cases that are possibly the product of artificial intelligence (AI) hallucinations." Attorney Loishirl Hall responded that she'd used Google's generative AI tool to help find case authority, never verified what it gave her, and filed it anyway. She has since left the firm. The firm claimed it didn't know until the court raised the issue.

The court wasn't persuaded that any of that mattered. Its order, reported by the legal blog The Volokh Conspiracy on September 4, put it bluntly: "citing to even a single fake case can be sanctionable," regardless of whether other, real citations in the same brief were accurate. The court struck the brief and referred the matter to the D.C. Office of Disciplinary Counsel. One concurring judge went further, citing the Ninth Circuit's recent opinion in Malkeet Lnu v. Blanche, which separated AI legal-research errors into two categories: "fabrications," where the AI tool invents a case that doesn't exist at all, and "inaccuracies," where it cites a real case but misstates what the case actually holds. The Ninth Circuit's own numbers on that second category are the more alarming ones: legal-specific AI research tools from Westlaw and Lexis produced hallucinated answers 17% and 33% of the time, respectively, on a representative set of legal queries tested in 2024.

This isn't a one-off. It's a documented pattern, growing weekly

Legal researcher Damien Charlotin maintains a public database of court decisions where a judge or tribunal found that a party relied on AI-hallucinated material, almost always fake or misrepresented case citations. As of September 10, 2026, it logs over 2,000 cases across more than 35 countries, with the United States accounting for the largest share by far. Lawyers account for more than 800 of the logged incidents; self-represented litigants, who don't have a law license on the line, account for even more. The consequences documented in the database range from a warning, to five-figure monetary sanctions, to bar referrals and, in at least one UK case, an attorney being struck from the register entirely. Days before the Deutsche Bank ruling, a California appeals court ordered pro se litigants Rajesh and Mahima Varma to pay the Bank of New York Mellon's entire appellate attorney fee bill over a brief riddled with fabricated case citations and misquoted authority. The Varmas hadn't written the brief themselves. At a sanctions hearing, Rajesh Varma admitted they'd paid a third party, Trina Patterson, to draft it. The court held the Varmas responsible anyway, since they'd signed it, and separately referred Patterson herself to the State Bar to investigate potential unauthorized practice of law.

What a detector can actually help with here, and what it can't

We've made this same distinction before about corporate research reports: a PwC report that invented an entire government partnership got caught because a reporter clicked through to the primary sources and read them, not because a detector flagged a fake citation. The same logic applies to legal filings, with a state bar complaint sitting where a retraction notice sat before.

An AI-writing detector, ours included, measures something different from citation accuracy. It scores the statistical likelihood that a passage of prose was generated or heavily edited by a language model, the same kind of analysis we've written about in detail elsewhere on how detection actually works. That's useful triage: if an entire brief was drafted by an AI tool end to end, the prose itself will usually carry a detectable signature, and a high score is a legitimate signal to slow down and manually verify every citation before the document goes out under a lawyer's bar number. That's the same workflow we recommended for corporate reports, and it holds here too.

But the Deutsche Bank case is a cleaner, and in some ways worse, version of the problem, because the brief's prose probably wasn't AI-written at all. Hall's own account was that she used an AI search tool to find case law, not to draft the brief. If that's accurate, a detector run against that brief's text would likely have scored it as human-written, correctly, because a human wrote it. The fabrication didn't come from the writing process a detector is built to catch. It came from an unverified research step earlier in the pipeline, when a human writer trusted an AI tool's output without checking it, then typed the fake case name into a document herself. No AI-writing detector, including ours, is built to verify that a case citation resolves to a real, on-point decision. That's a fact-checking and legal-research-verification task, not a stylometric one, and treating detection as if it covers both is precisely the gap that let four fake cases reach a federal appellate court's docket in the first place.

The two failure modes look identical from the outside, and need different fixes

A fake citation in a filed document is where both failure modes end up. Before a judge, the outcome looks the same too: a struck brief, a referral, sometimes a fine. But preventing each one takes an entirely different fix.

• AI-drafted documents. When a brief, memo, or report is substantially written by an AI tool, an AI-writing detector is a legitimate first tripwire. A section that scores as heavily AI-generated is exactly the section that needs a slower, citation-by-citation manual check before anyone signs their name to it. This is the workflow behind our own explainer on why a detection score is a signal, not a verdict applied to a document that's about to go out the door, not one already published.
• AI-assisted research, human-drafted prose. When the writing itself is a lawyer's own words but the underlying case law was pulled from an AI search tool and never independently verified, an AI-writing detector has nothing to flag, because there's no AI-generated prose to catch. The only fix here is procedural: treat every citation an AI tool surfaces as unverified until someone opens the actual case and confirms it says what's being claimed, the exact discipline the D.C. Court of Appeals said Hall's firm owed the court regardless of who typed the words.

Knowing which failure mode you're facing matters a lot. A firm that only runs its filings through a writing-style detector, then calls that due diligence, would have waved the Deutsche Bank brief through with a clean bill of health. It stayed clean, too, right up until a judge started reading the citations herself.

Check what's actually in a draft before it goes out under your name

Run any AI-assisted draft that will carry your name or your client's through our free demo first. It flags the sections that read as machine-written, and that's your starting point for a manual citation check, not the finish line. A clean detection score has never meant every fact and citation in a document is real.

FAQ

Can an AI detector catch a fake legal citation? Not directly. An AI-writing detector measures whether prose was likely generated by a language model. It doesn't verify whether a cited case exists or says what a brief claims it says. That's a separate fact-checking step no writing-style detector, ours included, is designed to replace.

Which lawyers have actually been sanctioned for AI-hallucinated citations? Deutsche Bank's outside counsel is a recent, high-profile example: brief struck, matter referred to the D.C. Office of Disciplinary Counsel, in September 2026, over four cases that don't exist. It's far from an isolated incident. Researcher Damien Charlotin's public database has logged over 2,000 comparable court decisions worldwide, and the sanctions have ranged from a plain warning up to five-figure fines and, in one UK matter, a solicitor struck from the register entirely.

What's the difference between a "fabrication" and an "inaccuracy" in an AI-assisted legal filing? Per the Ninth Circuit's framing in Malkeet Lnu v. Blanche, a fabrication is a citation to a case that doesn't exist at all. An inaccuracy is a citation to a real case that's been misstated or doesn't actually support the claim it's attached to. The court noted legal AI research tools produced hallucinated answers in a meaningful share of tested queries, meaning inaccuracies, which are harder to catch on a quick read, may be the more common failure in practice.