USPTO Issues Its First AI-Predicated Discipline Order Involving Hallucinated Cites to Intrinsic Record

The U.S. Patent and Trademark Office’s (USPTO’s) Office of Enrollment and Discipline (OED) posted a final order to its website on August 27, publicly reprimanding a California patent attorney for submitting a claim construction chart containing citations that generative AI invented — not to case law, but to the intrinsic record of the patent in suit.

In re Brian E. Mitchell, Proceeding No. D2026-16, resolves by settlement. Mitchell executed the agreement on July 20, 2026, and the OED Director on July 21. Tricia Choe, Associate General Counsel for General Law, approved it on July 27 on delegated authority from Under Secretary of Commerce for Intellectual Property and USPTO Director John A. Squires.

It is, on the stipulated facts, the first USPTO disciplinary order in which generative AI use is the predicate for the violations — and the first to treat hallucinated citations to a patent’s specification, figures, and prosecution history as an ethical offense. More on the first point below, because the conventional account of OED’s AI docket is wrong.

What Happened

Mitchell, of San Francisco, is Registration No. 50,612 and a member of the California bar. He represented the patent owner plaintiff in Magpul Industries Corp. v. Mission First Tactical Group, Inc., No. 2:24-cv-5551 (E.D. Pa.).

Preparing the plaintiff’s portion of a Joint Claim Construction Chart, he used one generative AI tool to draft proposed constructions of certain claim terms. He then used a second AI tool to review the draft.

The AI-generated portions cited the patent’s intrinsic evidence: specification, figures, prosecution history. Per the stipulated facts, many of the citations, quotations, and parentheticals referred to portions of the record that did not exist or were inaccurately attributed. Once the defect surfaced, Mitchell reviewed the chart, identified further erroneous citations himself, and circulated a corrected chart the following day citing the intrinsic record already before the court.

He does not concede the constructions were wrong. The order records his position that they were grounded in the intrinsic record, and that what failed was checking the citations against the source documents.

The court imposed no sanctions. The client suffered no prejudice. The case later settled and was dismissed with prejudice.

What OED Found Anyway

Four rules: 37 C.F.R. § 11.101 (competence), § 11.103 (diligence), § 11.804(c) (misrepresentation), and § 11.804(d) (conduct prejudicial to the administration of justice).

The competence finding is the one to read twice. OED framed it three ways: not understanding the inherent risks of AI-based research and drafting tools; not adequately reviewing the citations to the intrinsic record in a paper submitted to a tribunal; and not understanding the requirement to conduct a reasonable inquiry before presenting it.

Note what the second AI tool did not accomplish. Mitchell ran a reviewing model over the drafting model’s output, and OED still found competence and diligence failures. The order does not call a verification model worthless — but it plainly does not treat one as discharging the duty.

Worth flagging for anyone tracking Section 11.804(c): the stipulated facts contain no finding of knowledge, intent to deceive, or scienter of any kind. Prompt correction, no court sanction, full cooperation, contrition — all recorded as mitigation. OED found misrepresentation regardless.

Jurisdiction, and a Striking Paragraph 4

Paragraph 4 of the stipulated facts states that Mitchell was registered as a patent attorney in 2002 but has never prosecuted a patent or trademark application and has never appeared before the Trademark Trial and Appeal Board (TTAB) or the Patent Trial and Appeal Board (PTAB).

His registration is, functionally, dormant. The conduct occurred in a federal district court, not before the Office. No tribunal sanctioned him. OED disciplined him anyway. That is not new in principle — In re Hicks, D2013-11, reached litigation conduct in the Eastern District of New York a decade ago. But patent litigators holding a registration number they never use should absorb the point. The number travels with you.

The Order as Policy Vehicle

The Official Gazette notice OED ordered published reaches well beyond this respondent, reminding practitioners of obligations under Sections 11.101, 11.103, 11.301, 11.303, 11.804(c), 11.804(d), and 11.18 — the last of which substantially parallels Rule 11.

Then comes the sentence that matters most. 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.

That is a genuine shift. Nearly every citation-verification habit the profession has built over the past three years runs against Westlaw and Lexis. Nothing in that stack catches a wrong column-and-line cite to a specification, a figure number that does not match the drawing, or a quotation attributed to a response that never said it.

What Matos Actually Said

Which brings us to the received account. In re Anthony Matos, Proceeding No. D2025-13 — signed by the same official, Tricia Choe, on March 6, 2025 — is routinely described as the USPTO’s first discipline case involving AI. Read it.

Matos, a New York and New Jersey attorney not registered in patent matters, filed a TTAB trial brief in Cancellation No. 92077944. The Board catalogued fabricated quotations, propositions attributed to cases that never discussed them, and at least two decisions that do not exist. Matos found the cases through internet searches surfacing blog posts and articles, and admitted he never read them because he lacked access to the reporters.

Paragraph 13 of the stipulated facts records his assertion that while he used a generative AI tool to learn about the mechanics of TTAB proceedings, he did not use any such tool to draft the brief. OED nonetheless required two hours of CLE on generative AI in legal practice, which is what gave the order its reputation. But the stipulated predicate was a failure to read, not a failure of a model.

Emil Ali of McCabe & Ali flagged the ambiguity at IPethics & INsights within a week of the order, asking outright whether Matos had used AI at all. The stipulated facts answer him.

Why the Sanctions Differ

Matos drew a reprimand plus CLE. Mitchell drew a reprimand alone. The distinction is not severity of error — it is what each man did once the error surfaced.

The petitioner’s rebuttal brief put Matos on notice, and he let nearly two months pass through oral argument without seeking to explain or correct the citations. Mitchell found additional errors himself and circulated a corrected chart the next day. That is the same axis the Ninth Circuit turned on in Lnu v. Blanche, No. 24-4790 (9th Cir. June 3, 2026), where the court said lesser sanctions might have followed had counsel simply disclosed and apologized.

One through-line survives both orders. Matos had never practiced trademark law; the cancellation was his first trademark matter. Mitchell has never prosecuted an application. In each, a tool bridged a gap in experience — and the gap is precisely what made the errors invisible to the person who signed.