What the USPTO Order Actually Says

Brian E. Mitchell is a registered U.S. patent attorney. In Magpul Industries Corp. v. Mission First Tactical Group, Inc., a patent infringement case in the Eastern District of Pennsylvania, he was responsible for the patent owner’s portion of a Joint Claim Construction Chart. According to the USPTO’s final disciplinary order in Proceeding No. D2026-16, Mitchell used one generative AI tool to assist with proposed constructions of claim terms and then used a second AI tool to review the document.

The problem was not merely awkward drafting. The AI-generated sections included citations to intrinsic patent evidence—the specification, figures, and prosecution history—and the parties later identified substantial errors. The final order states that many citations, quotations, and parentheticals referred to portions of the record that did not exist or were inaccurately attributed. Mitchell reviewed the chart, found additional errors, and circulated a corrected version the following day.

Law360’s reporting from the court hearing gives the scale: the disputed chart addressed 10 terms across two patents, used 22 citations and quotations from the written descriptions and 26 citations to prosecution history, and only two of the 22 specification citations were real while none of the 26 prosecution-history citations were real. Those numbers come from the hearing reporting; the USPTO’s stipulated order itself uses the more general formulation that the chart contained substantial erroneous citations.

The Hallucinations Were in the Patent Record Itself

Most lawyers now know that generative AI can invent a case name, a quotation, or a statute. Patent work adds another category of risk: the model can fabricate the location of evidence inside the very documents the practitioner uploaded.

A model may confidently say that column 8, lines 35–42 supports a proposed construction, that Figure 4 shows a particular relationship, or that an applicant made a specific statement in a prosecution response. Those references feel unusually trustworthy because they are precise and because the source documents are already part of the matter. But precision is not verification.

This is directly relevant to ordinary patent workflows. When analyzing an Office Action, practitioners increasingly ask AI to identify where a prior-art reference discloses a limitation, where an amended claim finds written-description support, what an earlier response said, or which drawing corresponds to a component. These are exactly the tasks where an invented paragraph number, figure reference, or file-history quotation can quietly contaminate the analysis.

Why a Second AI Did Not Solve the Problem

Using a second model is often useful. Independent model outputs can expose inconsistencies, surface missed issues, and provide a cheap second opinion. But a second opinion and source verification are different functions.

If one model says paragraph [0065] supports an amendment and another model agrees, the agreement does not make paragraph [0065] say what the models claim. The verification step is opening paragraph [0065] and checking the actual text. Two models can share the same failure mode, repeat the same mistaken extraction, or simply accept a plausible citation because the surrounding analysis appears coherent.

The USPTO’s existing AI guidance follows the same basic logic: AI use is not prohibited, but existing professional obligations still apply. The tool does not absorb the practitioner’s duty to make a reasonable inquiry before submitting work product.

A Plausible Conclusion Can Still Rest on a False Citation

This distinction matters in patent analysis. Suppose an AI concludes that reference D1 discloses feature A. After reading D1 as a whole, the practitioner may agree that the conclusion is defensible. But the AI also writes: “Paragraph [0048] expressly discloses feature A.” If paragraph [0048] does not contain that disclosure, the citation remains false even if the broader conclusion may ultimately be supportable elsewhere.

Professional work is not evaluated only by whether the final answer feels approximately right. The evidentiary chain has to be real. For claim construction, Office Action responses, invalidity analysis, FTO, and AI-generated claim charts, a false citation can be especially damaging because later reviewers often rely on the citation instead of re-reading the entire source.

What the USPTO Disciplined—and What It Did Not Find

The final order publicly reprimanded Mitchell and concluded that the stipulated conduct violated four provisions of the USPTO Rules of Professional Conduct: 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 order is also important for what it does not say. It does not make a finding that Mitchell intentionally deceived the court. It records that he promptly acknowledged and corrected the errors, accepted responsibility, cooperated with the OED investigation, and expressed contrition. The district court did not impose sanctions for the inaccurate AI-generated citations, and the order states that the client suffered no prejudice.

IPWatchdog has described the matter as the first USPTO disciplinary order in which generative AI use itself formed the factual predicate for the violations. That is a media characterization based on the published disciplinary record, not a historical finding stated by the USPTO in the final order.

A Better Patent-AI Workflow

The answer is not to prohibit AI or redo every AI-assisted task manually from zero. AI can still handle a large share of first-pass work: reading, summarizing, comparing, drafting claim charts, locating candidate support, suggesting amendments, and identifying issues for review.

What cannot be skipped is source-level confirmation of the propositions that carry the conclusion. A practical patent workflow can use four simple rules:

  • If AI says an amended feature is supported by the specification, open the cited specification passage.
  • If AI relies on prosecution history, open the actual Office Action, amendment, interview summary, or response.
  • If AI says a prior-art reference discloses a claim limitation, verify the cited passage, figure, and context in the reference itself.
  • If AI relies on a case, statute, regulation, or agency rule, verify the primary authority and confirm that the proposition attributed to it is accurate.

The human review should be risk-weighted rather than ceremonial. The most important citations are the ones that determine claim scope, amendment support, prior-art mapping, infringement, validity, or a recommended business action.

What AI Evaluation Teams Should Learn from This Case

This case also matters beyond law-firm ethics. It illustrates a central problem in evaluating legal and patent AI systems: scoring only the final conclusion is not enough.

A model can reach a plausible answer while inventing the evidence trail. That means a serious evaluation rubric should separately score at least three things: whether the conclusion is substantively supportable, whether each material citation resolves to the claimed source, and whether the source actually supports the proposition for which it is cited.

For patent-domain AI, that distinction is particularly important because source material is often highly structured—claim numbers, paragraph numbers, figures, priority records, family data, prosecution papers, and specific prior-art passages. A robust benchmark should test traceability, not merely fluency.

Takeaway

AI can review AI. It can compare drafts, challenge reasoning, and catch many mistakes. But using another model does not convert an unverified citation into evidence.

The durable lesson from the Mitchell order is narrower and more useful: before a patent professional signs, submits, or relies on an AI-assisted work product, the citations that actually carry the conclusion must be checked against the original sources.

Sources & Further Reading