AI-generated patent citation errors lead to a USPTO public reprimand
The United States Patent and Trademark Office (USPTO) recently issued a disciplinary order that deserves attention from anyone using generative AI in patent practice.
The case is In the Matter of Brian E. Mitchell, Proceeding No. D2026-16. In its final order dated July 27, 2026, the USPTO approved a settlement between OED and the attorney and ordered a public reprimand.
An attorney handling U.S. patent litigation used generative AI to prepare claim construction materials. Those materials contained numerous inaccurate citations to the patent specification, drawings, and prosecution history. The matter ultimately led to USPTO disciplinary proceedings.
With tools such as ChatGPT becoming more common in patent work, this case has implications beyond a single U.S. attorney's mistake.
A second AI review still missed the errors
Brian Mitchell represented the patent owner in a patent infringement action before the U.S. District Court for the Eastern District of Pennsylvania.
Claim construction, the process by which a court determines the meaning of terms used in patent claims, is a critical stage of patent litigation. The parties propose interpretations of claim terms and identify supporting passages in the specification, drawings, and prosecution history.
Mitchell used generative AI to prepare those materials.
That alone is unremarkable.
However, some AI-generated citations did not match the actual patent record. They pointed to passages that did not contain the attributed material or incorrectly identified drawings or parts of the prosecution history.
What makes the case especially instructive is that Mitchell did not simply submit the first AI tool's output.
He used a second generative AI tool to review the first tool's work.
The errors survived both tools.
The litigation parties identified inaccurate citations and quotations when examining the materials. Mitchell then reviewed the document again and found further errors.
The USPTO's concern was not the use of AI itself
It would be inaccurate to read this case as saying that attorneys in the United States face discipline merely for using AI.
The USPTO does not prohibit the use of generative AI in patent practice as such.
Using AI to organize information or prepare drafts is already becoming a familiar part of legal work.
The problem in this case arose at the next step.
The attorney did not adequately check the AI-generated material against the source documents.
If AI says a particular statement appears on a certain page or in a particular paragraph of a patent specification, the practitioner must open the specification and confirm it.
If AI says the applicant made a particular argument in a response during prosecution, that response must be checked directly.
This may sound obvious.
Yet it is a step that can be surprisingly easy to skip in practice.
AI answers often read fluently. The wording is plausible, and the tool may provide precise-looking citations. When a second AI tool confirms the same answer, it is tempting to believe that verification has already taken place.
This case exposes that risk.
Having one AI tool review another is not the same as a person checking the original source.
No court sanctions, but USPTO discipline followed
Another aspect of the case is worth attention.
After the inaccurate citations came to light, Mitchell corrected the errors promptly. He circulated a revised claim chart the following day and acknowledged the shortcomings in his review.
The federal court did not impose sanctions over the AI-generated citation errors. The order also records Mitchell's representation that the mistakes were quickly corrected and that his client suffered no prejudice in the action.
The USPTO's disciplinary assessment was separate.
Its Office of Enrollment and Discipline (OED) addressed the attorney's duties of competence and diligence, as well as the presentation of inaccurate material in a legal proceeding. The matter resulted in a public reprimand.
The final order approving the settlement expressly identifies violations of 37 C.F.R. §§ 11.101, 11.103, 11.804(c), and 11.804(d).
The absence of separate court sanctions does not eliminate a patent practitioner's professional responsibility.
The errors concerned the patent itself, not invented case law
For several years, reports of generative AI use in legal work have highlighted hallucinations involving nonexistent court decisions.
For patent practitioners, this case is arguably even more directly relevant.
The problem was not fabricated case law. It was inaccurate citations to the patent specification, drawings, and prosecution history.
Patent practitioners often ask AI to perform tasks such as these:
“Find where this element is described in the specification.”
“Find where the applicant adopted a narrow interpretation of this term during prosecution.”
“Find support in the prosecution history for this proposed claim construction.”
AI can be useful for these tasks.
It may identify candidate passages much faster than a person reading a lengthy specification or hundreds of pages of prosecution records from beginning to end.
The critical step is checking whether the identified location is actually correct.
Paragraph numbers, column and line references, figure numbers, Office Action page numbers, and the applicant's exact wording in an amendment must ultimately be confirmed in the original documents.
AI can help identify where to look. Its answer is not itself evidence.
Source verification can matter even more in patent work
Patent practice depends heavily on the underlying documents.
Whether a particular feature is disclosed in the specification may affect an added-matter assessment. A single statement made by the applicant during prosecution may affect claim construction.
The same applies to prior art.
An AI summary stating that a reference discloses a combination of features A and B cannot simply be accepted at face value.
The original reference may describe A and B in different embodiments, or the AI may have expanded the disclosure beyond what the document actually says.
A practical approach is to assign clear roles.
AI can assist with searching, organizing, comparing, and drafting.
However, a person must verify the final supporting material against the original sources.
That step is especially difficult to dispense with for specifications, prior art, prosecution records, and judgments that support a final opinion.
Verification procedures matter more than whether AI was used
This USPTO order need not be understood as a reason to restrict AI use.
Generative AI is likely to become more common in specification drafting, prior art searching, Office Action analysis, and claim chart preparation.
The practical question is not simply whether AI was used.
What matters is how much of the work AI performed and how a person verified its output.
In this case, one AI tool reviewed the work of another.
The result was still wrong.
Anyone using generative AI in patent practice should remember this example.
When AI says, “That is what the specification says,” the final step is still to open the specification.
Pine IP Firm continues to review developments, rules, and cases concerning generative AI in patent practice, alongside its work in domestic and international patent prosecution and patent disputes.
