Do patent attorneys have to disclose that they used AI? The confidentiality and candour rules explained

By Alius Noreika

Key takeaways

• There is no general duty to tell the USPTO, the EPO or the UK IPO that an AI tool helped draft a filing.
• The duty of candour creates a narrow but real exception: if AI use is material to patentability, it must be disclosed.
• The signature on the paper — not the software behind it — is where liability attaches. Trusting AI output is not a reasonable inquiry.
• Obligations to the client are stricter than obligations to the office, and often require informed consent before confidential material goes into a third-party model.
• Courts are moving faster than patent offices, and several judges now demand certifications that patent offices do not.

The question comes up in almost every firm that has started experimenting with generative models: if a machine helped write the specification, does anyone have to say so? The answer that patent professionals keep arriving at — no, but — is correct, and the “but” is doing far more work than most summaries admit.

The short answer: no general duty to tell the patent office

The controlling US document is the USPTO guidance on AI-based tools in practice, published in the Federal Register on 11 April 2024 at 89 FR 25609. It is explicit on two points. There is no prohibition on using AI tools to draft documents for submission, and there is no general obligation to disclose that they were used. The Office repeats the point later in the notice: practitioners are not required to inform the USPTO that an AI tool was used in drafting a paper unless the USPTO specifically asks.

That guidance is interpretive rather than substantive rulemaking. It announced no new rule and created no new duty. What it did was map the existing rulebook — candour, signatures, confidentiality, export control, competence — onto a new category of tool, and warn that the rulebook is quite capable of catching careless AI use without any AI-specific provision.

The European position landed in the same place by a different route. The EPO Guidelines for Examination that entered into force on 1 April 2026 added a dedicated General Part, Section 5 on the use of artificial intelligence. Its organising principle is that parties remain responsible for the content of applications and submissions regardless of whether an AI tool helped prepare them. No disclosure statement is contemplated. The same revision also confirms the EPO’s own use of AI, including AI-assisted minute-taking for videoconference oral proceedings.

Where the exception bites: candour and materiality

The narrow exception sits inside the duty of candor and good faith, which for US practitioners runs through 37 CFR 11.303 and, in patent examination and reissue, through 37 CFR 1.56(a). The Office’s position is that if the use of an AI tool is itself material to patentability as defined in 37 CFR 1.56(b), it must be disclosed. Materiality is the trigger, not the technology.

The paradigm case is inventorship. If a drafting tool introduces alternative embodiments that no named inventor conceived, and the applicant then seeks to claim them, information about the interaction with the system — including inputs and outputs — may become material and disclosable. This is not a hypothetical failure mode. Broad, generative expansion of a disclosure into unclaimed embodiments is exactly what these tools are good at.

One wrinkle deserves attention, because a great deal of 2024-vintage commentary is now out of date. The April 2024 guidance framed its materiality example against the USPTO’s February 2024 inventorship guidance, which applied the Pannu factors to AI-assisted inventions. That framework is gone. The revised inventorship guidance for AI-assisted inventions published on 28 November 2025 at 90 FR 54636 rescinded the earlier version in its entirety and withdrew the Pannu-based approach for AI. Artificial intelligence systems are now treated as instruments analogous to laboratory equipment or research databases, and the traditional conception standard applies uniformly. Pannu survives only for allocating joint inventorship among multiple natural persons.

The practical consequence is subtle rather than dramatic. The disclosure trigger has not disappeared; it has been re-anchored. The question is no longer whether an AI contribution was “significant” under a bespoke test, but whether a human being actually conceived the claimed subject matter with the definiteness the case law demands. Where the honest answer is no, candour is engaged.

The signature is the real checkpoint

Most of the enforcement risk in AI-assisted patent work has nothing to do with disclosure at all. It sits in 37 CFR 11.18(b), the certification a person makes by signing or presenting any paper to the Office. Paragraph (b)(1) certifies that statements of the party’s own knowledge are true. Paragraph (b)(2) imposes a duty of reasonable inquiry covering improper purpose, legal warrant, and evidentiary support.

The USPTO’s formulation is unambiguous: relying on the accuracy of an AI tool is not a reasonable inquiry. To make the certification, the signer must have reviewed and verified the paper. Related mechanics follow from the same principle — a signature must be personally inserted by a natural person, an AI system cannot sign, cannot hold a USPTO.gov account, and cannot be sponsored as support staff.

Information disclosure statements are a specific trap. AI can populate a PTO/SB/08 form and can find references in the first place, but the signer certifies review of each listed reference. Dumping a large, unfiltered machine-generated list on an examiner risks being characterised as a paper presented for an improper purpose, because it can cause unnecessary delay or needless cost. The duty of disclosure under 37 CFR 1.56(c) cannot be delegated to a computer system.

Choosing tools with the verification burden in mind

Because verification is non-delegable, tool selection is a compliance decision rather than a procurement one. Specialized AI patent drafting tools can help patent professionals analyze invention disclosures, prepare initial claim sets and specification sections, and maintain consistent technical terminology throughout an application. The problem is too important for generic output, the language is specialized, and every generated draft still requires careful professional review.

Verification also has to reach beyond citations. Where a specification or drawings are AI-assisted, technical accuracy and compliance with 35 U.S.C. 112 must be checked before filing, because post-filing corrections risk new matter. Prophetic examples need to be clearly distinguishable from working examples. Where a priority document was drafted with AI assistance, technical errors carried into the US filing can jeopardise the priority claim.

Duties to the client are the stricter half

The disclosure question that most firms should actually be worrying about is not what they owe the office but what they owe the client. Here the answer is not “nothing”.

The American Bar Association’s Formal Opinion 512 on generative AI tools, issued 29 July 2024, works through competence, confidentiality, communication, candour, supervision and fees. Two conclusions matter most. First, a client’s informed consent is required before information relating to the representation is entered into a self-learning tool, and boilerplate language buried in an engagement letter does not qualify as informed. Second, disclosure to the client is required in defined circumstances — on request, where the engagement terms require it, and where the output will influence a significant decision in the representation.

State guidance is generally consistent and sometimes firmer. The Florida Bar’s Advisory Opinion 24-1 recommends obtaining the affected client’s informed consent before using a third-party generative tool where confidential information would be disclosed. In California, the State Bar has been considering amendments to the Rules of Professional Conduct that would fold AI-specific principles directly into binding rules, including for agentic systems.

Formal opinions are persuasive rather than binding, and a US patent practitioner is answerable to both the USPTO Rules of Professional Conduct and a state bar. Where the two diverge, the stricter obligation governs in practice.

Confidentiality is the harder problem

Disclosure is a paperwork question. Confidentiality is an existential one, because in patent practice a leak is not merely a compliance failure — it can destroy novelty before the application is ever examined.

Under 37 CFR 11.106(a) a practitioner must not reveal information relating to the representation without informed consent or another applicable basis, and paragraph (d), added in 2021 to track ABA Model Rule 1.6, requires reasonable efforts to prevent inadvertent or unauthorised disclosure. The USPTO’s specific warning is that AI systems may retain user input, that operators may use it for further training or share it with third parties, and that confidential material used in training may surface in outputs delivered to others. Supervising practitioners are responsible under 37 CFR 11.501 to 11.503 for staff and junior colleagues using these tools.

Then there is export control, which is easy to overlook and hard to unwind. Under 37 CFR 5.11 a foreign filing licence is required before exporting technical data for foreign filing purposes, and it does not authorise exporting subject matter abroad to prepare applications for filing in the United States. Where an AI service runs on servers outside the US, data entered into it may constitute an export; even domestically hosted systems can raise deemed export issues under 15 CFR 734.13(b) where non-US persons are involved. The Office’s instruction is to understand a tool’s terms of use, privacy policy and cybersecurity practices before adopting it.

Europe and the UK: same answer, different plumbing

European practitioners reach comparable conclusions through professional conduct rules rather than a single agency notice. The epi guidelines on the use of generative AI require members to ensure adequate confidentiality of prompts, datasets and other content transmitted to models, and to remain responsible for the work product. In the UK, IPReg’s artificial intelligence guidance identifies which parts of its regulatory arrangements bear on AI use, against a Core Regulatory Framework that already obliges attorneys to keep client affairs confidential. CIPA and CITMA have added their own practical material for members and clients.

Layered on top, EU practitioners face GDPR obligations wherever personal data is processed, and potential duties under the EU AI Act depending on whether the firm is acting as a provider or a deployer of a given system.

Courts are stricter than patent offices

The disclosure picture changes once a dispute reaches a tribunal. New York adopted the first comprehensive statewide rule: 22 NYCRR Part 161 on the use of artificial intelligence technology, adopted 25 March 2026 and effective 1 June 2026, applies to all Unified Court System courts. Its posture is accountability rather than disclosure — AI use is permitted and no system-wide disclosure duty is imposed — but its Appendix A supplies an opt-in model rule under which a signature certifies careful review and the absence of fabricated cases, statutes or other material. Because adoption is per court, practitioners still have to check.

Individual judges are less forgiving, and hundreds maintain standing orders requiring disclosure, certification or verification. The consequences are not theoretical for patent litigators. In Lexos Media IP LLC v. Overstock.com in the District of Kansas, Senior Judge Julie A. Robinson sanctioned five attorneys on 2 February 2026 under Rule 11 after briefing in a patent infringement case was found to contain non-existent cases, fabricated quotations and citations to real decisions that held the opposite of what was claimed. Monetary sanctions totalled roughly $12,000, with the largest fine falling on the attorney who admitted incorporating unverified ChatGPT output, and liability was spread across colleagues who signed without checking. The court declined to allow a corrected brief.

Some tribunals do require an affirmative statement. Quebec’s Tribunal administratif du travail, for example, directs litigants and representatives to add a line at the head of any document indicating that AI generated at least part of its content. The lesson is that “no disclosure required” is a statement about a particular forum, never a general rule.

Who requires what

Frequently asked questions

Does using AI affect who is named as inventor? It can, but not because AI is a co-inventor — it cannot be one. Under the November 2025 revised guidance, the question is whether a natural person conceived the claimed invention. Claims covering machine-generated embodiments no human conceived are the risk.

Can a client insist on knowing whether AI was used? Yes. Formal Opinion 512 treats a client request as a circumstance requiring disclosure, and engagement terms can require it independently.

Is an enterprise tool with a no-training setting enough? It is necessary rather than sufficient. Confidentiality obligations also reach retention, subprocessors, server location and breach exposure — and, in the US, export control.

The bottom line

Patent attorneys do not, as a rule, have to announce that they used AI. What they cannot do is treat that absence of a disclosure duty as an absence of obligation. The signature is the certification, the client’s confidences are not the firm’s to donate to a model, and candour attaches to what the filing says rather than to how it was produced. Offices that decline to demand disclosure are not being permissive. They are pointing out that the existing rules were always about responsibility, and that responsibility is exactly what a generative tool cannot accept on a practitioner’s behalf.

Sources

• USPTO, “Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the United States Patent and Trademark Office,” 89 FR 25609 (11 April 2024): https://www.federalregister.gov/documents/2024/04/11/2024-07629/guidance-on-use-of-artificial-intelligence-based-tools-in-practice-before-the-united-states-patent
• USPTO, “Revised Inventorship Guidance for AI-Assisted Inventions,” 90 FR 54636 (28 November 2025): https://www.federalregister.gov/documents/2025/11/28/2025-21457/revised-inventorship-guidance-for-ai-assisted-inventions
• ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512 (29 July 2024): https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
• New York State Unified Court System, 22 NYCRR Part 161: https://www.nycourts.gov/rules/part-161-use-artificial-intelligence-technology
• New York State Bar Association on Part 161: https://nysba.org/effective-june-1-2026-the-new-york-state-unified-court-system-has-adopted-a-new-rule-regarding-the-use-of-artificial-intelligence/
• epi, “Guidelines: Use of Generative AI in the Work of Patent Attorneys”: https://information.patentepi.org/issue-4-2024/epi-guidelines-use-of-generative-ai.html
• IPReg, Artificial Intelligence guidance: https://ipreg.org.uk/pro/practice-development/artificial-intelligence
• Marks & Clerk on the 2026 EPO Guidelines and AI: https://www.marks-clerk.com/insights/latest-insights/102mn4e-2026-epo-guidelines-how-the-epo-is-approaching-ai-in-practice/
• Mathys & Squire, preview of the 2026 EPO Guidelines for Examination: https://www.mathys-squire.com/insights-and-events/news/european-patent-office-previews-the-2026-guidelines-for-examination/
• ABA Journal on the Lexos Media sanctions proceedings: https://www.abajournal.com/news/article/judge-orders-patent-attorneys-to-explain-ai-hallucinated-citations
• Reporting on the 2 February 2026 sanctions order: https://thedailyrecord.com/2026/02/06/federal-judge-fines-attorneys-for-ai-generated-brief-made-up-cases/
• The Florida Bar on Advisory Opinion 24-1 and Formal Opinion 512: https://www.floridabar.org/the-florida-bar-news/aba-issues-first-ethics-guidance-on-a-lawyers-use-of-ai-tools/
• State Bar of California, proposed rule amendments on AI: https://www.calbar.ca.gov/public/public-meetings-comment/public-comment/public-comment-archives/2026-public-comment/proposed-amendments-rules-professional-conduct-related-artificial-intelligence
• IPWatchdog on practitioner obligations after the USPTO guidance: https://ipwatchdog.com/2024/05/02/tips-for-using-ai-tools-after-the-usptos-recent-guidance-for-practitioners/