Are AI legal research tools safe in 2026? What the Sixth Circuit just held

Safe enough to draft with. Not safe enough to file without opening every source yourself. And in 2026 the check that matters has changed shape, which is the part nobody selling you a tool wants to explain.

On 3 April 2026 the Sixth Circuit sanctioned a court-appointed lawyer whose appellate briefs were drafted in Westlaw's CoCounsel. Not ChatGPT. A legal-specific platform, grounded in a real case law database, sold on the promise that it works from verified sources. Every authority cited in those briefs was genuine. The quotations inside them were invented, and two holdings were described backwards. If your cite-check consists of confirming the case exists, that check passes here and you get sanctioned anyway.

What actually happened in United States v. Farris

Steven N. Howe, forty years at the bar and never disciplined, was appointed under the Criminal Justice Act to argue John Farris's sentencing appeal. The court's first clue had nothing to do with the law. It was the file name of the principal brief: "CoCounsel Skill Results."

Reading further, the panel found three problem citations. Page 4 attributed language to the Sentencing Guidelines commentary at section 3B1.1, note 1. Page 10 quoted United States v. Washington, 715 F.3d 975, for the proposition that "simply facilitating the offense without exercising decision-making authority is insufficient." Page 10 also quoted United States v. Anthony, 280 F.3d 694. The reply brief repeated two of them. Correct case names, correct reporters and page numbers. The court went looking for the quoted words and could not find them in those sources, or in any other authority.

The holdings were worse than the quotes. Washington upheld the role enhancement that Howe's brief said it reversed. Anthony did vacate an enhancement, but on a narrow counting-methodology point irrelevant to Farris, and the defendant there had conceded that he directed and supervised others. Both errors ran in the direction that helped the client, which is exactly how an unverified draft tends to fail.

Howe answered the show-cause order honestly. He had told a staff member to upload the district court documents to CoCounsel to generate a first draft, then worked in that same file himself for six hours to supplement it. His office had bought the product in August 2025, after the district court proceedings ended, and he had never used it this way before.

Six hours of an experienced lawyer's attention, and three fabricated quotations walked straight through.

The sentence that should change your cite-check

The court held that citing "real legal authorities", as opposed to hallucinations featuring fictitious cases, did not absolve him. That single line retires the advice most firms adopted after the fake-citation cases of 2023 and 2024. We covered that first wave in our piece on AI hallucination sanctions, and the standard fix everyone landed on was: confirm the case is real. It is no longer sufficient on its own.

Here is the mechanism. Retrieval-augmented tools pull real documents from a real database, then generate prose about them. Correct retrieval does not make the generation correct. The model can hold the right opinion in front of it and still produce a sentence that opinion never contained. So the error has migrated from the citation to the words around it, which is the part your cite-checker glances at and your opponent reads closely.

Verify the string, not the citation. Open the opinion, search for the quoted words, and read enough to confirm which way the case came out. KeyCite will tell you an authority is good law. It will not tell you the quotation was manufactured.

A federal court just cited the hallucination research against the vendors

The most useful line in the opinion for anyone choosing a tool is this one: attorneys should not use technology without knowing how it can be misused, and that stays true "even when new tools are sponsored by trusted legal technology providers." The authority the court cited for that proposition is the Stanford and Yale study by Magesh and colleagues, published in the Journal of Empirical Legal Studies in 2025.

That study found Lexis+ AI, Westlaw's AI-Assisted Research and Ask Practical Law AI hallucinating between 17% and 33% of the time. Westlaw's tool answered accurately on 42% of queries. Lexis+ AI, the strongest performer tested, reached 65%. The paper sets those numbers directly against the marketing: LexisNexis had promised "100% hallucination-free linked legal citations", and Casetext had said CoCounsel "does not make up facts, or 'hallucinate'".

Be fair to the vendors. The testing closed in April 2024. It did not cover CoCounsel Legal, which launched in August 2025, or Lexis Protégé, from January 2025. Both companies disputed the methodology, and their objections were not frivolous.

They also did not publish independent benchmarks of their own. Two years on, the only preregistered public evaluation of these products is one the vendors dislike, and a federal appellate court has now put it in a published opinion. Until somebody funds a fresh independent test, the honest position is that you do not know your tool's error rate and neither does the person selling it to you.

Who does the checking, which is where small firms will slip

Howe did not run CoCounsel himself. He told staff to. The court cited ABA Model Rule 5.3 and Kentucky's equivalent, and found that relying on staff rather than himself or another attorney to supervise the AI-generated work product fell short of his obligations as attorney of record.

Read that as a rule about your paralegal. If a non-lawyer runs the research tool and no lawyer independently rebuilds the citations before filing, you have a supervision problem underneath your candour problem. In a two-person office that arrangement is the default rather than the exception, which is the same structural gap we wrote about in the piece on AI training and governance. The opinion also points readers to ABA Formal Opinion 512 as the collected statement of what competence now requires.

The bill for getting this wrong: no Criminal Justice Act compensation for any time on the appeal, referral to the Chief Judge of the Sixth Circuit for disciplinary proceedings, service of the opinion on the district court and on the Kentucky Bar Association's disciplinary clerk, removal from the case, briefs locked, briefing schedule reset. Farris, an indigent defendant, waits longer for his appeal because of it.

Where UK solicitors stand on this

The refreshed judicial guidance of 31 October 2025 lists legal research under "Tasks not recommended", on the reasoning that AI tools are a poor way to find new information you cannot verify independently. It is addressed to judges, not to you. It still tells you how the bench will receive a skeleton argument built on an AI answer.

Do not wait for a procedural rule to arrive and hand you a safe harbour. The Civil Justice Council working group chaired by Lord Justice Birss reported in June 2026 that respondents broadly agreed no AI-specific requirements are needed for pleadings, skeletons and advocacy documents, because professional responsibility frameworks already cover the ground. Its stated direction of travel is to maintain that position for professional drafting, with further work on witness statements, expert evidence and litigants in person. The final report is due later this year.

Which means England and Wales is about to formally decide that your existing duties are the whole rulebook. Farris is what those duties look like when a court actually applies them.

What to do on Monday

• Use the tool to find the shape of the argument, then rebuild every citation from the source document. Treat its prose as a lead, never as a draft you edit.
• Verify quotations by opening the opinion and searching the exact string. If the words are not there, the sentence goes, however good it sounds.
• Verify direction, not just existence. Howe's brief said a case reversed when it affirmed, and no existence check on earth catches that.
• The lawyer who signs the filing does the verification personally. Not staff, and not a second AI asked to grade the first one.
• Watch your artefacts. The Sixth Circuit found this through a file name, and the UK judiciary's own list of tells includes the accidental inclusion of an AI prompt in a document. If that debris reaches the court, nobody read the document.

The honest answer on safety is that these tools are safe in the way a bright trainee is safe. Quick, genuinely useful, occasionally confident and wrong, and never the person whose name goes on the filing.

We keep a running breakdown of which AI tools are worth a small firm's money, and on what terms, in our guide to the best AI tools for lawyers. If you want the confidentiality and verification checks in one place, the free guide covers the workflow we would run in your seat.

FAQ