In re Mitchell: Why the Second AI Didn’t Save Him

Using one associate to prepare a claim construction chart and a second to check it is ordinary practice. In In re Brian E. Mitchell, Proceeding No. D2026-16, a patent attorney did the same with two generative AI tools: one to draft proposed claim constructions and a second to review them. Erroneous citations still reached the district court, and the U.S. Patent and Trademark Office’s (USPTO’s) Office of Enrollment and Discipline (OED) publicly reprimanded him for that.

IPWatchdog previously reported what happened. This piece is about why. A claim construction chart cites a patent by column and line, which makes it close to the worst thing a patent attorney could hand a model to check. The stipulated order does not reproduce the prompts. What follows is not a reconstruction. It is how this failure can happen, why a second model is poorly placed as a check, and, before the close, where the Mitchell matter fits and where it does not.

What a Check Is

Consider what the second associate does. Say the chart cites the specification at column 7, lines 22 through 31. She opens the patent. She turns to that column. She reads the lines, compares them against what the chart says they support, and either confirms the citation or flags it. That is what a check consists of: not agreement with the drafter, but the cite resolved against the source.

Asked to check, a model does something different. It produces a written assessment, and like everything it writes, that assessment draws on two things: what sits in its working space, and what it learned in training.

The working space is everything in front of the model when it answers: the developer’s instructions, which you never see; your question; any document whose text was placed into the session; and anything a search tool fetched. Text in that space is available the way a document open on your desk is available. The model can read it and quote it. But the desk is only so big, and products differ in how they get a document onto it. Some pull in the pieces that look relevant, and some show the model page images. The tool doesn’t always tell you what it left out, and the model answers either way, with the same conviction.

Whatever isn’t on the desk, the model supplies from what it absorbed during training. That is not a stored copy of everything it read. It works more like memory. Think of a novel you read years ago. You remember the plot twist and the ending, but not that the twist came on page 312. Memory keeps the meaning and drops the coordinates: the page in a novel, the column and line in a patent.

Now consider the request an attorney would naturally write: here is the claim chart; tell me whether the citations are correct based on the specification. That phrase describes the task. It does not deliver the document. Unless the specification’s text is in the working space, or the tool goes and gets it, the model answers from recollection, and recollection does not hold coordinates.  What a model in that position can reliably tell you is that a cite is properly formed, not that it is right.

The Hard Case: Text Versus Layout

The obvious fix is to give the model the patent. That helps, but less than you would think. The coordinate must survive into what the model receives, and coordinates come in two kinds.

Some are textual. A published application’s paragraph numbers are printed in the body, an office action’s numbered paragraphs were typed by the examiner, and “FIG. 3” appears in a specification as text. Extraction carries them through intact. Others are typographic, recording only where the words landed on the printed page. The column and line numbers of an issued patent are layout, not text. Copy the text out of an issued patent PDF, as many tools do before a model reads it, and look at the result. The words arrive. The column and line numbers either disappear or turn up as stray numbers mixed into sentences.

A tool that keeps the page layout, or lets the model look at the page image, can in principle find column 7 and count down to line 22. Whether yours does is the question, and nothing in the model’s answer tells you. A model can have the entire specification in front of it, be asked for the cite, and supply one from recollection, even if it read the document all along.

The rest of the record has its own version of the problem. Office actions and responses are separately paginated. Drawings are images: “FIG. 3” is text, but what it depicts is not. Older file-history papers were scanned and reached a model only through OCR, if at all. Hence the caution the USPTO issued with the order: “the pitfalls of AI-generated citation error are not confined to extrinsic sources … but may extend to the intrinsic evidence of patent and trademark applications and their file wrappers.”  The second associate’s advantage has nothing to do with intelligence. Her copy of the patent has the numbers printed around the text. The model’s may not.

Why the Second Model Didn’t Catch It

If a second model gets the patent in a form that keeps the coordinates, finds column 7, and reads the lines, it has checked the cite. If it reads the first model’s chart, or the specification’s words without their addresses, it has only judged whether the cite is plausible, and a plausible cite looks exactly like a correct one.

A cite read from the document and a cite produced from recollection arrive in the same format and the same confident register. An attorney unsure of a pin cite looks it up. A model states a sound recollection and an invented one identically and cannot go and look unless something in the system looks for it.

Nor is a second model as independent as it appears. The second associate’s review matters because she has the patent in front of her, numbers and all. A second model may have the patent too, but if the column and line numbers didn’t survive, it has no way to confirm the address. It can only judge whether the address looks plausible, and a cite drafted by a model is plausible. The number of models is the wrong measure. Independence comes from the source, including the address.

What the Mitchell Record Shows

The order does not say what the models actually did. But the court file shows what happened, rules out one of the problems above, and confirms two others.

What happened is the thing the word “hallucination” is for. Opposing counsel counted 47 quotations in the chart that do not exist. They were not misremembered; they were written, and they fit Mitchell’s proposed constructions. His construction of “generally perpendicularly” was 90 degrees “subject to standard manufacturing tolerances”; the chart quotes the applicant telling the examiner that the term covers deviations “acceptable within standard manufacturing tolerances.” No such remark exists.

The coordinates were not the problem. A lost column number gives a real passage the wrong address. These passages have no address because they have no source.

The file history went unread. At a hearing, Mitchell told the court that he uploaded the intrinsic record and asked the first model for supporting citations. Yet, of the 47, 26 were attributed to the file history, and none of them exists. Whether the file history fell off the desk or never reached it is unclear. That left a gap, and a hallucination filled it with what he had asked for.

The checks never reached the source either. Mitchell asked the first model to confirm that the quotes were verbatim and to double-check its work; it consulted the same desk. He then gave the results to a second model, which, in his word to the court, “compounded the problem.”

Before You Sign

The order puts the onus on the practitioner, not the tool. Using AI was not a violation itself. In the Mitchell case, OED found four violations of the USPTO Rules of Professional Conduct, and the first two, competence under 37 C.F.R. § 11.101 and diligence under § 11.103, turn on how the tools were used. The competence violation rests first on failing to understand the risks inherent in AI-based research and drafting tools, including mistakes and hallucinated citations. The diligence violation rests first on failing to use those tools reasonably. That finding assumes reasonable use is possible. So, what does it look like?

It starts with pulling apart three questions that “review this” lumps together. Every citation in an AI-assisted filing raises all three, and each calls for a different kind of check.

• Does the cite exist and say what the filing claims? This is a lookup, not a judgment. Open the source and read it: the patent page with its column and line numbers visible, the drawing itself rather than the specification’s mention of it, the right prosecution filing at the right page. Check against the source document, not the chart and not a second model.
• Does the passage support the proposition?  A citation can be real and the quotation accurate, and the passage still may not carry the weight the paper puts on it. Does the prosecution history excerpt establish the disclaimer the chart asserts?  Does the specification define the term, or merely describe an embodiment?  A model earns its place here if the passage is in its working space and the prompt asks for a passage-by-passage comparison rather than an overall impression.
• Is the conclusion right?  This one belongs to a person with relevant experience. A second model can help as an adversary: ask it for the strongest argument against the construction. It is less useful as a vote. Frontier models are trained on much of the same material, so on a question with a confident conventional answer, two of them tend to agree for the same reasons.

Apply the three questions to Mitchell’s chart, and an inversion appears. The third question requires the most legal judgment. The first requires none, and the first is the one that produced the order. Mitchell was not disciplined for a wrong construction; the order records his position that the constructions were grounded in the intrinsic record. He was disciplined for citation errors that anyone with the patent open could have caught.

And if an error gets through, correct it immediately and completely. When opposing counsel flagged two claim terms, Mitchell reviewed the whole chart, found the problem ran further, and circulated a revised chart the next day. The Office made the same point by quoting the Ninth Circuit in Lnu v. Blanche: “read everything cited in a court filing—whether drafted by generative AI or not—and disclose quickly and transparently generative AI hallucinations that are inadvertently included in court filings.”

What Would Have Saved Mitchell

The lesson is not that AI cannot review AI. It can, for the right question. Mitchell had two models reviewing the chart. What he lacked, on his own account, was a sufficient check against the source documents. That check needed no expertise and no second opinion, only the patent, open to the right page.

Image Source: Deposit Photos
Author: BiancoBlue
Image ID: 831471608