Using AI Before the Patent Office
USPTO guidance, April 2024. Read from the primary document; quotations verified against its own text.
In April 2024 the USPTO issued guidance confirming that existing duties of candor, signature certification, reasonable inquiry and confidentiality govern the use of AI tools in matters before the Office, without prohibiting AI-assisted drafting and without creating a general requirement to disclose that AI was used.
This guidance is most useful for what it declines to do. It does not ban AI-assisted drafting and it does not require you to disclose that a tool was involved. What it does is close the gap someone might hope for: every existing duty still applies, and the signature on a paper certifies the whole paper regardless of what produced it. Human verification of AI output is described as necessary, which means the reasonable-inquiry duty is not satisfied by a tool's confidence. For a firm, the work this creates is process work, not disclosure work.
The Office "issues this guidance to inform practitioners and the public of the important issues" that professionals and innovators "must navigate while using Artificial Intelligence (AI) in matters before the USPTO". It is a notice applying existing rules, not a rulemaking. It creates no new obligation and grants no exemption. The Office issued it to say that nothing had changed, which is itself the message.
The Duties That Apply Unchanged
| Duty | What the guidance says | What it means when a tool drafted the text |
|---|---|---|
| Candor and good faith | Each individual associated with a proceeding "has a duty of candor and good faith in dealing with the Office" | Unchanged; anything material is disclosable whatever produced it |
| Reasonable inquiry | Rule 11.18(b)(2) "imposes a duty of reasonable inquiry", that "the paper is not being presented for any improper purpose, the legal contentions are warranted by law, the allegations and other factual contentions have evidentiary support" | The signer's inquiry, not the tool's |
| Verification | "human verification of the outputs of AI tools is necessary" | Necessary, not advisable; its extent is not defined |
| Disclosure of AI use | No general duty created | An absence, not a safe harbour |
Those duties combine into a certification that cannot be delegated. The tool may draft, and the signer certifies. Because verification is necessary rather than advisable, relying on unverified output is a failure of reasonable inquiry regardless of how the text was produced. That pairs directly with the electronic-signature rule, under which the named signer must personally make their own signature: one rule says who signs, the other says what the signature certifies.
What a Firm Policy Should and Should Not Build
- Build it around verification, not disclosure. Verification is the enforceable point. A disclosure obligation is the one firms most often over-implement, because the Office did not create one; ordinary materiality rules are untouched either way.
- Keep the signer identifiable per paper. Certification is non-delegable in substance.
- Treat confidentiality as a separate question from correctness. The guidance discusses the risks of putting client technical information into third-party systems; that discussion is not distilled on this page and is covered in the AI tool compliance checklist.
"Necessary" is stated and not measured: the guidance sets a requirement whose extent it does not define. That gap has now been filled a little from the other end, by enforcement.
The First Discipline Order
In In re Brian E. Mitchell, Proceeding No. D2026-16, the USPTO publicly reprimanded a registered patent attorney by settlement. The agreement was executed by the respondent on 20 July 2026 and by the OED Director on 21 July, and the final order issued on 27 July 2026, signed on delegated authority. We have found no earlier USPTO discipline order predicated on generative AI use; the order does not claim to be the first, and that characterisation is ours.
What happened is ordinary enough to be worth stating plainly. Preparing a joint claim construction chart, the practitioner "used one generative artificial intelligence ("AI") tool to assist with drafting proposed claim constructions" and "then used a second AI tool to review" it. The output "included citations to the intrinsic evidence of the patent, including the patent's specification, figures, and prosecution history. At least some of the AI-generated citations were erroneous."
| Rule | What it requires | How it was breached, in the order's words |
|---|---|---|
| 37 C.F.R. 11.101 | Competent representation | (1) "not fully understanding the actual or potential inherent risks (e.g., mistakes, hallucinated citations or quotations, etc.)" of the tools; (2) "not adequately reviewing the citations to the intrinsic record"; (3) "not fully understanding the requirement to conduct an inquiry reasonable under the circumstances when presenting a paper to a tribunal" |
| 37 C.F.R. 11.103 | Reasonable diligence | (1) "not making reasonable use of AI-based legal research and drafting tools in connection with representation of a client"; (2) the same failure to review the citations; (3) "not conducting an inquiry reasonable under the circumstances in order to certify that that the citations offered in support of the client's proposed claim construction were warranted by fact or law" |
| 37 C.F.R. 11.804(c) | No misrepresentation | "presenting inaccurate support, including false citations and quotations from the intrinsic record" |
| 37 C.F.R. 11.804(d) | No conduct prejudicial to the administration of justice | Same conduct |
Two details make this more useful than a cautionary tale. The second AI tool did not catch the first one's errors, so a model reviewing a model was not verification. And "[t]he Court did not impose sanctions due to the inaccurate AI-generated citations on Respondent or his client": the discipline came from the Office afterwards, on conduct a court had let pass.
The order does not set a threshold for how much verification is enough, and it was settled rather than litigated, so it decides nothing about a contested case. What it does supply is a concrete description of the failure: not understanding that the tool invents citations, and not checking the ones it produced against the record.
The order is a scanned image with no text layer. Quotations here were extracted by OCR and then verified by reading the rendered page images of the order itself on 29 September 2026.
What This Does Not Decide
- It does not say whether a particular firm's AI practice complies.
- It does not say whether a specific submission satisfied reasonable inquiry.
- It concerns professional duties, not patentability.
Educational, not legal advice. This page reports what an authority says. Applying it to a matter is work for a registered practitioner with the file in front of them.
Sources
- USPTO, Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the USPTO (Apr. 2024)
- Back to the cluster