Table of Contents
A patent illustrator’s work used to end at the drafting table. Now a complete figure set can land in your inbox twenty minutes after someone pasted an invention disclosure into a chat window. AI-generated patent drawings are not prohibited, and the USPTO has never required that a figure be drawn by hand. But the moment an AI system touches a sheet, a specific set of rules attaches to the human being who signs and files it. Those rules cover signatures, candor, client confidentiality and export control, and not one of them is satisfied by the tool’s own output. Here is what actually applies.
What AI-generated patent drawings are, and what they are not
The phrase covers several different things that carry very different risk. At one end is a model that vectorises a photograph or cleans up a hand sketch into line art. At the other is a generative system that reads an invention disclosure and invents the figure itself, deciding which components to show, how to section them, and which reference numerals to attach.
Those are not the same activity. Tracing a sketch changes the rendering of something the inventor already conceived. Generating a figure from prose can introduce structure that the inventor never described, which is where the real exposure sits. Most commercial tools sit somewhere in between, and the vendor’s marketing copy will rarely tell you which.
It matters because the Office does not regulate the tool. It regulates the filing and the person who signs it. There is no box on an application data sheet asking whether a figure was machine-made, and there is no rule that says a generated drawing is inadmissible. The obligations land elsewhere, and they land hard.
- Vectorisation and cleanup — lowest risk; the underlying disclosure is human.
- Automated view generation from CAD — moderate; geometry is real but sectioning and shading conventions frequently miss.
- Generative figure drafting from text — highest; the model can add structure, and that becomes a new matter and written-description problem.
- Automated numbering and lead-line placement — moderate; errors here are mechanical but they are also the most common source of objections.
The April 2024 guidance is the document that governs this
On 11 April 2024 the USPTO published Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the United States Patent and Trademark Office, applicable from that date. It remains in force. It is worth reading in full, because it is the only USPTO document that speaks directly to figures produced with AI assistance.
The guidance says that where the specification and/or drawings of an application are drafted using AI tools, practitioners must take extra care to verify the technical accuracy of the documents and their compliance with 35 U.S.C. 112. That is an explicit statement about drawings, not a general platitude, and it is the sentence most vendors do not quote.
The guidance also warns about prophetic examples. A model asked to illustrate an embodiment will happily draw one that was never built and never described. MPEP 2164.02 distinguishes prophetic from working examples, and a figure that blurs that line creates an enablement argument for an examiner or, later, for an opponent.
Do not confuse this with the inventorship guidance. The USPTO issued Inventorship Guidance for AI-Assisted Inventions on 13 February 2024 and then rescinded it in its entirety with revised guidance on 26 November 2025. The revised guidance keeps the core rule: only natural persons can be named as inventors, because 35 U.S.C. 100(f) defines an inventor as “the individual” who invented the subject matter. AI systems are tools.
Nobody but a human can sign, and the rules say so twice
37 CFR 1.4(d)(1) requires that patent correspondence bear a person’s signature, and 1.4(d)(2)(i) requires that an S-signature be personally signed or inserted by that person. 37 CFR 11.18(a) says the same thing for practitioners. The April 2024 guidance is blunt about the consequence: it is not acceptable for correspondence to carry the signature of an AI tool or any other non-natural person, and one person may not type another’s signature even when told to.
The certification in 37 CFR 11.18(b) is where the weight sits. Signing a paper certifies that statements of your own knowledge are true, and 11.18(b)(2) imposes a duty of reasonable inquiry. The guidance states plainly that simply relying on the accuracy of an AI tool is not a reasonable inquiry. If a generated figure shows a component that does not exist in the disclosure, the signer wears that.
The access rules are stricter than most firms assume. A USPTO.gov account belongs to one individual and may not be shared. An AI system may not obtain an account, and a practitioner may not sponsor an AI tool as sponsored support staff. Exceeding authorised access to the patent electronic systems is not merely an administrative problem — the guidance points at 18 U.S.C. 1030.
The duty of candour under 37 CFR 1.56(a) rounds this out, and the guidance is explicit that it cannot be transferred to another person or to a computer system. If you know that a figure or a claim originated with a tool in a way that bears on inventorship, that knowledge is yours to act on.
Confidentiality and the export-control trap
This is the risk that gets least attention and deserves the most. 37 CFR 11.106(a) prohibits a practitioner from revealing information relating to the representation of a client absent informed consent, and 11.106(d) requires reasonable efforts to prevent inadvertent or unauthorised disclosure. Pasting an unpublished disclosure and a set of figures into a third-party model is a disclosure decision, whether or not anyone treats it as one.
37 CFR 11.101 requires competence, which the rule frames as keeping abreast of the benefits and risks of the technology used to handle client matters. In practice that means reading the tool’s terms of use, its data-retention policy and whether prompts are used for training, before the first disclosure goes in — not after a client asks.
Then there is the export problem. 35 U.S.C. 184(a) requires a foreign filing licence before an application is filed abroad within six months of a US filing, and 37 CFR 5.11 frames the licence as authorising export of technical data for the preparation of a foreign application. The guidance warns directly that AI tools may run on servers outside the United States, so data entered into them may be exported, potentially in breach of export administration and national security regulations.
15 CFR 734.13(b) treats release of controlled technology to a foreign person as an export. A drawing is technical data. If the model that cleans up your figure runs in a datacentre abroad, you have made a transfer, and the fact that it happened inside a web form does not change the analysis. Ask where the inference runs before you ask what the subscription costs.
Formalities do not bend: 37 CFR 1.84 and 1.152 still govern the sheet
No part of 37 CFR 1.84 was relaxed for machine output. Drawings must be in black ink under 1.84(a)(1); colour in a utility application requires a petition under 1.84(a)(2), the fee set by 1.17(h) and a statement in the specification. Design applications are the exception — colour drawings there do not need that petition. Photographs are not ordinarily permitted under 1.84(b) and are accepted only where they are the sole practicable medium.
The checks generated output fails most often are mechanical. 1.84(l) requires lines that are durable, clean, black, sufficiently dense and dark, and uniformly thick and well-defined; models routinely produce anti-aliased greys that reproduce badly. 1.84(p)(5) requires that reference characters not mentioned in the description do not appear in the drawings, and that characters mentioned in the description do appear — a bidirectional check that an automated numbering pass will break the moment the text is edited.
Design applications are harder still. 37 CFR 1.152 allows broken lines to show visible environmental structure but not hidden planes, permits solid black surface shading only to represent the colour black or colour contrast, and bars combining photographs and ink drawings as the formal drawings of one application. Generative tools mix rendering styles happily, and that single habit will sink a design filing.
If you want the formalities in isolation, our walkthrough of 37 CFR 1.84 patent drawings covers each subsection, and automated patent drawings deals with the narrower question of which of those checks automation gets wrong. Design-side dashes are unpacked in design patent broken lines.
Europe: Rule 46 EPC is gone, and colour is now allowed
Half the advice circulating about European drawing formalities is citing a rule that no longer exists. Rule 46 EPC, which set out the form of the drawings, was deleted with effect from 1 February 2023, along with Rule 49(3) to (12) EPC. The presentation requirements moved into a Decision of the President of the EPO published in the Official Journal, which lets the Office adjust them without amending the Implementing Regulations.
What survived is the brief description. Rule 42(1)(d) EPC still requires the description to contain a brief description of the figures. If a tool or a template tells you that Rule 46 EPC governs the form of your sheets, the tool is working from a stale corpus, which is exactly the failure mode you should expect from a model trained on old practice notes.
The bigger change is colour. From 1 October 2025 the EPO accepts drawings filed electronically in colour or greyscale, under the current Decision of the President and the accompanying Notice in the Official Journal. There is no conversion to black and white, and the figures must be contrast-rich and legible at 300 dpi. That is a genuine opening for rendered output, and it is covered in detail in EPO color patent drawings.
The PCT did not follow. PCT Rule 11.13(a) still requires drawings to be executed in durable black lines without colourings. A colour figure that the EPO will accept on a direct European filing is still non-compliant on an international application, so a single generated set cannot be used unthinkingly across both routes.
Who owns the output, and the honest answer about copyright
Clients ask whether they own a generated figure. The honest answer is that copyright is probably not the protection they are imagining. In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), the D.C. Circuit held that the Copyright Act requires a human author, affirming the Copyright Office’s refusal to register a work naming an AI system as sole author. The Supreme Court denied certiorari on 2 March 2026, so the human-authorship rule stands.
In practice this rarely bites a patent filing. A figure published in an issued patent is doing disclosure work, not licensing work, and the patent claims — not copyright in the line art — are the asset. Where it does matter is marketing collateral, product manuals and litigation exhibits derived from the same files, and in vendor contracts that purport to assign copyright the vendor may not hold.
Inventorship is the sharper question. Since a figure can supply written-description support under 35 U.S.C. 112(a), a generated drawing that introduces structure nobody conceived is not a drafting shortcut — it is an unexamined contribution sitting in your application. Read it as new matter until you have confirmed otherwise against the disclosure, and remember that drawings can also function as prior art against you, which we cover in patent drawings as prior art.
Contractually, the practical move is to require the vendor to disclose whether generative models are used, where inference runs, whether prompts are retained or used for training, and to warrant that deliverables are free of third-party material. Most drawing vendors will answer those four questions in writing. The ones that will not have told you something.
A pre-filing review for AI-generated patent drawings
Reviewing AI-generated patent drawings is not the same as proofreading them. You are looking for invented structure first and formalities second, because the formalities are fixable after filing and invented structure usually is not. Work through this before the sheets go into Patent Center.
- Reconcile every element against the disclosure. Any component in a figure that the inventor did not describe is a new-matter problem, not a style question.
- Run the reference characters both ways. Every numeral in a figure must appear in the description and every numeral in the description must appear in a figure, per 37 CFR 1.84(p)(5).
- Check line quality at print size. 1.84(l) wants solid black, uniformly thick lines; zoom to 100% and look for grey edges and broken strokes.
- Confirm the colour position. Utility colour needs a petition under 1.84(a)(2) and the 1.17(h) fee; design colour does not; PCT still wants black lines under Rule 11.13(a).
- Check design conventions separately. Broken lines only for visible environmental structure, solid black shading only for black or contrast, and never photographs mixed with ink drawings in one application under 1.152.
- Verify the European citations. If anything in the file refers to Rule 46 EPC as live law, the source is out of date.
- Record where the data went. Note the tool, the hosting region and the retention terms, so the 11.106 and 35 U.S.C. 184 analysis exists on the file rather than in someone’s memory.
None of this argues against using the tools. It argues for treating generated sheets as a draft from an unqualified assistant: useful, fast, and never filed without a human who is prepared to certify it under 37 CFR 11.18(b).
Have Your Generated Figures Reviewed Before You File
PerspireIP prepares and reviews figure sets for USPTO, EPO and PCT filings, including sheets that started life as machine output. We check them against 37 CFR 1.84 and 1.152, against the current EPO presentation requirements rather than deleted ones, and against the disclosure itself — so that what you file is supported by what your inventor actually described. See our patent drawing services or contact us with the files you already have.
Frequently Asked Questions
Are AI-generated patent drawings allowed at the USPTO?
Yes. No rule prohibits them, and the USPTO does not ask how a figure was made. What applies is the April 2024 AI guidance plus the ordinary formalities in 37 CFR 1.84 and 1.152, and the certification duties of the human who signs the filing.
Do I have to disclose that a figure was made with AI?
There is no general disclosure requirement for the tool itself. But the duty of candour under 37 CFR 1.56(a) applies to information material to patentability, and the USPTO’s guidance gives the example of needing to disclose where AI-drafted content raises a question about whether a named inventor made a significant contribution.
Can an AI tool sign or file a patent document?
No. 37 CFR 1.4(d)(1) requires a person’s signature and 11.18(a) requires the practitioner to sign personally. The April 2024 guidance states that a signature of an AI tool or other non-natural person is not acceptable, and AI systems may not hold or be sponsored for a USPTO.gov account.
Is uploading an unpublished application to an AI tool an export?
It can be. The USPTO guidance warns that AI tools may use servers outside the United States, so data entered may be exported. Under 15 CFR 734.13(b) release of controlled technology to a foreign person may be deemed an export, and 35 U.S.C. 184 governs foreign filing licences.
Does Rule 46 EPC still apply to European drawings?
No. Rule 46 EPC was deleted with effect from 1 February 2023 and the form requirements now sit in a Decision of the President of the EPO. The brief description of the figures remains a Rule 42(1)(d) EPC requirement, and colour or greyscale drawings have been accepted for electronic filings since 1 October 2025.
Can a generated figure create a new matter problem?
Yes, and this is the main risk. Drawings can supply written-description support under 35 U.S.C. 112(a), so a figure showing structure the inventor never described adds subject matter to the application. Reconcile every element against the disclosure before filing.