The judiciary is also exploring how to deal with AI. One observation that has accumulated over the past two years: even the most expensive legal-specific AI cannot achieve zero hallucinations. Lawyers must verify every citation generated by AI, or risk losing their license. But this is not an entirely new problem brought by AI. The legal profession has always had a verification tradition for subordinates and co-counsel: drafts written by junior associates must be read by senior partners; documents submitted by out-of-state lawyers must be reviewed by local counsel. AI has merely changed the delegate from a person to a tool — the verification obligation has not changed one bit. In June 2026, a ruling by the U.S. District Court for the Northern District of Mississippi laid this all out at once: four lawyers on both sides of the litigation were simultaneously ejected from the case, because every link in the chain assumed someone else had already verified.
The case originally began as an ordinary contract dispute over attorney’s fees. Tom Withers III sued the City of Aberdeen, Mississippi, claiming the city had not fully paid his legal fees from another case. Dozens of such disputes are filed in U.S. federal courts every day, but this one ultimately became a precedent no lawyer wants to be cited against them.
The plaintiff’s out-of-state counsel, Kathleen Wilson, who practices in Louisiana, used AI to draft the legal brief submitted to the court, which was stuffed with fabricated cases. At a hearing on January 20, 2026, she told Judge Aycock that she “did not know AI could hallucinate.” Two months later, she was sanctioned again in a separate case in the Western District of Louisiana Bankruptcy Court for the same AI hallucination problem. Her apology did not work, because she had not added any verification step into her workflow.
The defendant’s out-of-state counsel, Kathryn Williams, was in worse shape. Her firm, Daniel, Williams & Associates, had a written AI policy explicitly requiring lawyers to independently verify AI-generated citations. As a partner, she directly violated her own firm’s policy. The AI software she used was designed only for Texas law, and she knew this, yet she used it on a Mississippi case. Judge Aycock wrote in the order that Williams tried to deflect attention and avoid acknowledging her knowledge of the software’s design flaw. Worse still, to avoid being subpoenaed to appear, she provided the judge with a false scheduling conflict.
The local lawyers also failed to do their duty. Plaintiff’s local counsel Shauncey Ridgeway and defendant’s local counsel Mark McClinton each admitted to the court that when they signed, they genuinely believed their out-of-state colleagues had already verified the content. They received drafts from the out-of-state lawyers, did not open or read them, simply signed their names and handed them to the court. Judge Aycock revoked Wilson’s and Williams’s pro hac vice admissions, barred them from appearing in the Northern District of Mississippi for two years, imposed individual fines on all four, and canceled the civil trial outright. A contract dispute over attorney’s fees became a hollow proceeding no one could continue, because not one of the four had read the documents they signed.
The Mississippi case is not an isolated one. French lawyer Damien Charlotin maintains an AI hallucination case database. As of February 2026, it had catalogued 239 U.S. cases of lawyers sanctioned for AI hallucinations. In 2025 alone, courts worldwide issued 712 related rulings, with over 90% concentrated in that single year. From the 2023 Mata v. Avianca case (where a lawyer used ChatGPT to fabricate six nonexistent cases and was fined $5,000) to 2026, both the frequency of cases and the severity of penalties have risen exponentially.
Fines of a few thousand dollars no longer deter lawyers’ wishful thinking. In 2026, the Sixth Circuit imposed $15,000 in punitive sanctions on each of two violating lawyers in Whiting v. City of Athens, and required them to bear double the litigation costs. This is one of the highest punitive amounts at the circuit level to date. The court wrote in its opinion that it intended to send “the loudest signal” that this kind of submission of fabricated citations is not permitted in this court or any other.
The Northern District of Alabama went even further in Johnson v. Dunn. The judge did not impose a fine, but directly disqualified the lawyer from representing the client in subsequent proceedings and reported the matter to the relevant state bar associations. The violator was not a junior lawyer, but the practice group co-leader of a large firm. The judge wrote in the opinion that if fines and public exposure were effective deterrents, there would not be so many cases to cite. A fine cannot reflect the degree of dereliction involved in fabricating citations, nor measure the multiple harms it causes.
Many lawyers believe that buying the most expensive legal-specific AI will let them rest easy. A test report from Stanford RegLab and HAI dismantles this belief. They tested Lexis+ AI, Thomson Reuters’ Westlaw AI-Assisted Research, and Ask Practical Law AI using over 200 pre-registered legal queries. The results: Lexis+ AI and Ask Practical Law AI had error rates exceeding 17%, and Westlaw AI-Assisted Research had a hallucination rate exceeding 34%. At least one in every six queries produced an error. No independent evidence supports manufacturers’ claims of being “hallucination-free.”
In April 2026, the Sixth Circuit delivered a heavy blow in U.S. v. Farris. The lawyer delegated drafting to an assistant who uploaded it to CoCounsel, and spent only six hours supplementing the work. The legal brief generated by CoCounsel cited real cases in its citations, but the direct quotes were fabricated. The court found that three citations could not be traced to corresponding source text. The court denied the lawyer appellate compensation, referred him for disciplinary action, and removed him from the case. Thomson Reuters had been promoting CoCounsel as “hallucination-free” — this case directly disproved the vendor’s headline claim.
The Central District of California dealt with a more complex version in Lacey v. State Farm. Lawyers from two firms simultaneously used three tools — CoCounsel, Westlaw Precision, and Google Gemini — to generate a court outline. About 9 out of 27 citations were problematic, including two entirely nonexistent cases. The special master presiding over the hearing, Michael Wilner, wrote in his ruling:
Drawn in by the authorities they cited, I began to look up these decisions, only to find they did not exist. This terrified me. It almost led to an even more horrifying result: incorporating this fabricated material into a judicial ruling and order.
The two firms were fined a combined $31,100. Wilner’s remark is the most chilling part of the entire case: had the judge not checked, these AI-fabricated citations would have become part of judicial precedent.
The essence of this issue is not that AI has gone wrong, but that the management problems already present in the legal profession have erupted en masse under AI’s acceleration. The legal profession has long maintained the tradition of delegation and verification. A senior partner would not directly sign and submit a draft written by a junior associate without reading it first. A local lawyer would not forward an out-of-state lawyer’s documents to the court unchanged without reviewing their content. This has been a practice convention for a century, codified in Federal Rule of Civil Procedure 11: by signing a document, a lawyer certifies having conducted a reasonable inquiry. This system constrains not AI, but every link in the delegation chain.
After AI emerged, the delegate changed from a person to a tool, but the verification obligation has not been reduced one bit. Wilson delegated verification to AI, Ridgeway delegated verification to her out-of-state colleague, and McClinton delegated the signing obligation to “I assumed she already checked.” Williams went even further — she delegated her judgment obligation to the vague consensus of “I heard AI works,” while violating her own firm’s written policy. Every link assumed someone else had verified, and not a single link in the entire chain was truly accountable.
Wilson’s case is the most illustrative. Two months after apologizing before Judge Aycock, she was sanctioned again in the Louisiana bankruptcy court for the same AI hallucinations. A verbal apology did not work, because she had not internalized verification as her own responsibility. What she lacked was not knowledge of AI risks, but an institutional habit of embedding verification into her workflow. As long as internal management does not undergo qualitative change, she will make the same mistake again next time.
Faced with this situation, the U.S. judiciary did not create entirely new AI-specific laws. Judges found that existing procedural rules were already sufficient. In spring 2024, the Fifth Circuit established an AI subcommittee to consider whether to mandate AI usage disclosures from lawyers. After 13 public comments, the committee recommended against adopting specialized rules, finding that existing Rule 11, Appellate Rule 46(c), and the court’s inherent sanction authority were adequate. The February 2026 Fletcher v. Experian case validated this approach. Chief Judge Jennifer Walker Elrod wrote in the opinion:
If ignorance of AI risks was ever an excuse, it is certainly no longer one. To take responsibility for legal arguments and citations, lawyers must ensure that AI-generated content is reliable.
This passage has been directly quoted by multiple district courts, becoming authoritative grounds for the proposition that “the ignorance defense is dead.” Judge Aycock also cited the same passage in the Mississippi case.
Some judges have gone further. Judge Brantley Starr of the Northern District of Texas issued the nation’s first AI standing order in May 2023, requiring lawyers to submit a two-option certificate when appearing: either declare that no AI was used to draft any document, or declare that AI-generated content has been manually verified citation by citation using traditional legal databases. As of November 2025, 32 states plus Puerto Rico had issued 286 AI-related court rules or orders.
Industry self-regulatory bodies are also catching up. The American Bar Association’s Formal Opinion 512, issued in July 2024, embedded the verification obligation into competence requirements: indiscriminately trusting AI output constitutes a violation of the duty of diligence. Lawyers must understand the capabilities and limitations of the tools they use, and the signer is responsible for the authenticity of the document. This is not a new rule — it is an old rule re-emphasized.
This judiciary-led tightening is reshaping the industry’s competitive landscape. Clio’s 2025 report shows that 79% of lawyers nationwide are already using AI at work, up from 19% just two years earlier. Bloomberg Law’s figure is even higher, at 83%. But firms’ institutional development lags far behind. 8am’s 2026 industry report presents a glaring statistic: 69% of lawyers individually use AI, but only 34% of firms have adopted legal-specific AI tools. 43% of firms have no AI usage policy whatsoever, and 54% provide no AI-related training.
On the client side, the picture is different. LexisNexis’s 2024 survey shows that 80% of corporate legal executives expect outside counsel to reduce fees due to AI use, but only 9% of law firm leaders have heard this demand. 59% of in-house lawyers do not know whether outside counsel are using AI, and 41% of firms have never disclosed AI usage on client bills. Mark Smolik, general counsel of DHL Supply Chain Americas, said something that captures the client attitude: “We’re not waiting anymore.”
This gap between adoption rates and governance coverage is splitting law firms into two groups. A small number of firms have embedded AI into pricing, training, credentials, and partner KPIs. Morgan Lewis requires lawyers to complete Coursera courses, hands-on exercises, and a final assessment before using tools, earning an internal digital badge upon passing. Sidley Austin has new associates participate in AI hackathons during orientation. Most firms remain without policy, without training, and silent toward clients.
The Mississippi case is a signal. When the execution cost of drafting a legal document drops to near zero, the value of judgment and verification ability rises instead. The four Mississippi lawyers failed not because AI was hard to use, but because they delegated something they should have done themselves and did not follow up to check the result. The court is using sanctions to weld this responsibility back onto the signer’s hands: you signed it, you read it. This has nothing to do with whether AI is a new technology, and everything to do with whether a lawyer is a responsible signer.