USPTO Reprimands Patent Attorney in First Discipline Case Over AI-Hallucinated Citations to Intrinsic Record

The legal profession's rush to embrace generative artificial intelligence has collided with the unforgiving reality of professional liability. For months, the debate over AI in the courtroom has centered on a specific, almost comical failure mode: chatbots inventing fake judicial opinions out of whole cloth and attorneys blindly submitting them to judges. But the tension between technological efficiency and ethical duty has just shifted from the theoretical to the highly technical. The U.S. Patent and Trademark Office (USPTO) has drawn a hard new line in the sand, proving that the consequences of AI hallucinations are no longer confined to extrinsic case law, but can corrupt the very core of a technical dispute.[1]

On August 27, the USPTO's Office of Enrollment and Discipline posted a final order publicly reprimanding Brian E. Mitchell, a California-based patent attorney. The disciplinary action stems from a federal patent infringement lawsuit, Magpul Industries Corp. v. Mission First Tactical Group, Inc., litigated in the Eastern District of Pennsylvania. In that case, Mitchell utilized two separate generative AI tools to draft and subsequently review a Joint Claim Construction Chart—a highly technical document used to define the exact scope and meaning of the patent terms at issue. Relying on the software to parse the complex material, Mitchell submitted the chart to the court without manually verifying the AI's outputs against the primary source documents.[1]

What makes this disciplinary action a landmark is the specific nature of the artificial intelligence's fabrication. The AI did not merely hallucinate a nonexistent court ruling or misquote a federal statute; it hallucinated citations directly to the patent's own intrinsic record. The software confidently generated references to specific figures, specification lines, and prosecution history documents that simply did not exist within the patent being litigated. This represents a dangerous evolution in AI errors, moving from easily verifiable external case law to the dense, bespoke technical evidence that forms the foundation of any patent infringement claim.[1]

Until now, the cautionary tales of artificial intelligence in legal practice have largely involved "extrinsic" sources. Throughout 2025 and early 2026, federal and state judges across the country levied thousands of dollars in fines against attorneys who submitted appellate briefs citing fake cases generated by ChatGPT. In one notable instance earlier this year, the USPTO disciplined another attorney, Anthony Matos, for filing a Trademark Trial and Appeal Board brief loaded with fabricated quotations and nonexistent decisions he found via unverified internet searches. Those cases established that citing fake law is a sanctionable offense, but they did not address the corruption of a case's underlying factual record.[1][2]

The Mitchell order significantly elevates the standard of strict liability for practitioners using these tools. According to the stipulated facts in the disciplinary order, Mitchell had never prosecuted a patent application before, and the generative AI tool was effectively bridging a gap in his technical experience. When the opposing counsel flagged the glaring discrepancies in the claim construction chart, Mitchell immediately reviewed the document, identified further erroneous citations himself, and circulated a fully corrected version to the court and opposing counsel the very next day. He did not attempt to cover up the error or blame the software in a defensive posture.[1]

Despite his rapid correction, the outcome was severe. The federal judge overseeing the Magpul case imposed no sanctions, and the client suffered no prejudice before the case was ultimately settled and dismissed with prejudice. Yet, the USPTO's disciplinary arm chose to act anyway, issuing a formal public reprimand. This decision underscores a critical and unforgiving ethical absolute: the tool's failure is ultimately the attorney's failure. Disciplinary boards are signaling that the mere act of submitting unverified, hallucinated material constitutes a breach of the duty of competence, regardless of the attorney's intent or subsequent apologies.[1][3]

"Practitioners are cautioned that the pitfalls of AI-generated citation error are not confined to extrinsic sources such as statutes, regulations, and case law, but may extend to the intrinsic evidence of patent and trademark applications and their file wrappers," the USPTO order explicitly stated. This language serves as a formal policy vehicle, warning the entire patent bar that the duty of candor and diligence applies to every single line of text an AI generates. It effectively puts practitioners on notice that delegating the review of technical file wrappers to an algorithm carries immense professional risk.[1]

The broader legal community is already grappling with the widespread fallout of unsupervised AI adoption. Industry trackers have identified over 600 cases nationwide where lawyers cited nonexistent authority due to AI hallucinations, prompting a crisis of confidence in legal tech. In response, legal technology experts and ethics panels are urging law firms to implement strict, multi-tiered verification workflows. They note that while AI is a powerful assistant for summarizing vast amounts of text and drafting preliminary arguments, it remains a fundamentally unreliable unsupervised author that requires a human-in-the-loop at every critical juncture.[2][3]

As generative artificial intelligence continues to evolve and integrate into daily legal practice, the regulatory focus is shifting from outright bans to enforcing rigorous human oversight. The Mitchell reprimand makes it abundantly clear that the efficiencies gained by using generative AI do not absolve attorneys of their fundamental duty to verify the accuracy of their submissions. In the high-stakes, technically demanding world of patent litigation, there is simply no substitute for reading the intrinsic record, and the USPTO has proven it will permanently mark the records of those who fail to do so.[1]