AI Inventorship and Human Conception

Inventorship is the doctrinal question that sits between "I built something" and "I have a patent." It is where the developer hands off to the attorney. With AI tools now embedded in normal engineering workflows, the inventorship question is no longer theoretical. It shows up every time a team decides who goes on a filing.

This page covers three things in order: the doctrine (Thaler, the USPTO's revised guidance, the human-conception standard), the practical decision framework (when AI use is safe, when it requires documentation, when it risks invalidation), and the engineering practice (what to log, how to attribute, where structured-disclosure tooling fits).

This article presents published USPTO guidance, Federal Circuit case law, and practitioner analysis for educational purposes. It is not legal advice. Consult a registered patent attorney for filing decisions.

Four different questions often get collapsed into one

Before walking through the inventorship doctrine, it helps to separate the four questions that get tangled together whenever AI tools are part of the engineering workflow. Each has its own legal framework, and the answers do not always line up.

Ownership and inventorship are particularly easy to confuse. A founder may say "the company owns it," and that may well be true through assignment, but the attorney still needs to know which humans conceived the claimed invention. A contract can assign rights. It cannot turn a non-inventor into an inventor.

The core rule: only humans can be inventors

The Federal Circuit decided this in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022). Dr. Stephen Thaler attempted to name his AI system "DABUS" as the inventor on two patent applications. The court held that under 35 U.S.C. § 100(f), an inventor must be an "individual," and an "individual" in this context means a natural person. AI systems, however sophisticated, are not individuals within the meaning of the Patent Act.

The USPTO's November 28, 2025 Revised Inventorship Guidance for AI-Assisted Inventions (90 Fed. Reg. 54636, Document 2025-21457, Docket PTO-P-2025-0014) is the current authoritative statement of how the Office handles AI involvement in patent applications. It rescinded the February 13, 2024 guidance and reaffirmed the human-only inventorship rule. The revised guidance also makes clear that no new substantive standard was created. The Office's view is that the existing inventorship doctrine, applied with care, already handles AI-assisted invention.

MPEP § 2109 captures the same principle for examiners: inventorship requires a natural person, and Thaler controls. MPEP § 2157 addresses the procedural side, including how incorrect inventorship is corrected under 37 C.F.R. § 1.48 and the rare cases in which the issue rises to rejection under §§ 101 and 115.

AI assistance does not create an inventorship bar

This is the part of the doctrine practitioners and developers most often misstate. The USPTO has been explicit, and the revised guidance reiterates it: AI assistance does not categorically disqualify a patent. The question is whether a human or humans supplied conception of the claimed invention. AI use that is genuinely tool-like is fine.

The Office's framing in the revised guidance is that AI systems function like laboratory equipment, calculation software, or databases: they can be powerful, they can generate suggestions or outputs, but they remain tools. The dispositive question is what the natural-person inventors did with them.

The USPTO presumption, and no new disclosure duty

Two practical points from the USPTO's FAQ on the revised guidance that reduce panic among engineers and founders:

• No heightened standard. The guidance does not create a heightened standard for AI-assisted inventions. The ordinary inventorship analysis applies.
• No new disclosure duty. There is no new requirement to file an "AI use confession" with the USPTO. Applications are presumed correct as to the named inventors. The applicant's duty of candor (37 C.F.R. § 1.56) continues to apply, but the guidance does not impose new affirmative disclosure obligations specifically about AI tool use.

The ordinary rule remains: name the natural persons who conceived the claimed invention. If AI use later becomes relevant to whether the named humans actually conceived the claims, the record matters. That is the practical reason to document.

What human conception actually requires

Conception is the formation of "a definite and permanent idea of the complete and operative invention". The phrase comes from the long line of cases anchored on Hybritech, Inc. v. Monoclonal Antibodies, Inc., 802 F.2d 1367 (Fed. Cir. 1986). It is a question of fact, and it is the entire ball game for inventorship.

In an AI-assisted workflow, conception is preserved when:

• A human articulates the technical problem to be solved and the criteria for an acceptable solution.
• A human evaluates AI-generated outputs and selects which ones to incorporate, applying technical judgment.
• A human integrates AI outputs into a working concept of the complete and operative invention.

Conception is at risk when:

• A human prompts an AI tool with "design a system that does X" and accepts the output without substantive technical evaluation.
• The "definite and permanent idea" of the claimed invention first appears in an AI output rather than in human thought.
• The human's contribution is administrative (typing the prompt, selecting from a menu of options) rather than inventive.

The gray zone between these two categories is real and the page does not pretend otherwise. It is the area where careful documentation matters most, where attorney engagement is most useful, and where the Federal Circuit has yet to issue an AI-specific holding.

Where Pannu still matters and where it does not

Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998), established the joint-inventorship test. To be a joint inventor, a contributor must (1) contribute in some significant manner to the conception of the invention, (2) make a contribution that is not insignificant in quality when measured against the dimension of the full invention, and (3) do more than merely explain to the real inventors well-known concepts or the current state of the art.

The rescinded February 13, 2024 USPTO guidance applied Pannu directly to AI contributions in a claim-by-claim analysis, evaluating whether each AI tool's role rose to a Pannu-type contribution. The November 28, 2025 revised guidance restored Pannu to its ordinary role: joint inventorship among natural persons. The Office's position is that when one human uses an AI tool, Pannu does not apply to evaluate the AI's contribution, because AI systems are not persons and cannot be joint inventors. The question for one-human-plus-AI scenarios is the conception question above: did the human conceive the claimed invention?

This is an examination-guidance correction, not a Federal Circuit AI-specific holding. The Federal Circuit has not directly addressed AI-assisted conception. Pannu remains good law for joint inventorship among natural persons, and the modern Federal Circuit reminder of how it operates is HIP, Inc. v. Hormel Foods Corp., where the court reiterated the clear-and-convincing burden and held that the alleged contribution was too insignificant relative to the full invention to support joint inventorship.

Multiple humans plus AI: Pannu still applies among the humans

When two or more natural persons work together using AI tools, Pannu still applies between them. The AI tool fact does not change the analysis. The harder documentation question becomes: what record shows person A's conception versus person B's conception versus either of them rubber-stamping AI output?

Inventorship follows the claims, not the GitHub hero story

This is the most under-appreciated point in practice, and it is where the engineering view and the patent view often diverge. Inventorship is tied to the claimed invention, not to who built the product. 35 U.S.C. § 116 recognizes that joint inventors may apply jointly even when they did not physically work together, did not make the same type or amount of contribution, and did not contribute to every claim.

Three practical implications:

• The people who shipped the most code on the product are not automatically the inventors of every patent claim.
• A feature that seems central in the product may never appear in the claims, while a small implementation detail suggested during development may become the patentable distinction.
• If claims are amended during prosecution, the correct inventor set may need to be revisited.

Separately: AI-related inventions can still be §101 eligible

Eligibility under §101 and inventorship under § 100(f) are different questions. They sometimes get tangled because both involve AI. The USPTO's Subject Matter Eligibility Examples 47-49, issued July 17, 2024, illustrate §101 eligibility analysis for AI-related claims. Example 47 covers artificial-neural-network anomaly detection. Example 48 covers speech separation with deep neural networks. Example 49 covers a personalized medical-treatment scenario.

These examples remain current under the USPTO's subject-matter-eligibility guidance. They are useful as illustrations of how the Office analyzes §101 for AI-related claims. They do not decide who invented the claimed subject matter. The inventorship question on each example would be analyzed under § 100(f), Thaler, and the November 2025 revised guidance, not under Examples 47-49 themselves.

For the eligibility side of the cluster, see our Section 101 Declaration Strategy guide and our Software Patents After Alice overview.

A practical risk table for AI-use patterns

Engineering teams want a concrete answer to "what AI use is safe and what isn't." The honest answer is that risk depends on whether the AI's contribution becomes claim-relevant, not on whether AI was used at all. The following table reflects current practitioner consensus on documentation discipline by use pattern.

Risk and documentation guidance is descriptive of current practitioner consensus reflected in the USPTO November 2025 revised guidance and the AmLaw commentary on AI-assisted patent drafting. It is not a legal rule. Specific filings should be discussed with counsel.

A conception note template for engineering teams

The single most useful artifact an engineering team can produce, when they suspect they are building something potentially patentable, is a short conception note. The format below is workable. It is not legal advice; it is an engineering habit that gives counsel something to evaluate later.

Conception note

Problem:
What technical problem were we trying to solve? Why is it hard?

Constraints:
Latency, privacy, reliability, cost, compatibility, regulatory,
deployment, or other constraints we worked within.

Alternatives considered:
What other approaches did we evaluate, and why did we reject them?

AI involvement:
What tool was used (model, version, configuration)? What did it
suggest? Was the suggestion generic, boilerplate, or potentially
claim-relevant?

Human technical judgment:
Who selected, modified, rejected, or integrated the idea? Why was
that technically sound? What did the human contribute that a tool
could not?

Evidence:
Design doc, PR, commit hash, issue, prototype, benchmark,
architecture diagram, experiment log.

Claim relevance:
Which feature might matter in a patent claim?
      

The note does not have to be long. Three to five sentences per field is enough. The value is not in length; it is in capturing the conception event close to when it actually happened. Reconstructing this later, from memory and commit history alone, is much harder.

A practitioner intake checklist

For attorneys and patent agents engaging with engineering clients on AI-assisted invention, a short intake conversation early in the engagement saves substantial work later. Questions worth asking:

• Who first identified the technical problem the invention addresses?
• Who proposed the architecture or algorithm that may be claimed?
• Did an AI tool suggest any claim-relevant feature? Which tool, with what configuration?
• Who evaluated alternatives and selected the final approach? What was the technical basis for that selection?
• Are there multiple human contributors who may rise to joint inventorship under Pannu?
• Did any contractor, advisor, customer, or open-source contributor contribute to conception?
• Are there AI logs, prompts, outputs, design notes, or other records that should be preserved, withheld, or reviewed under counsel's direction?

The last point ties to the discovery dimension covered separately. AI prompts and outputs can become discoverable in later litigation, as the Conservation Law Foundation v. Shell Oil Co. magistrate order this spring illustrated. See our emerging framework on AI prompts in federal discovery for the discovery side of the same documentation question.

Where the evidence comes from

The conception note above describes the artifact. The question is how engineering teams produce it without slowing down. Two approaches that work:

• Manual discipline. One person on the team (usually the tech lead) writes a brief conception note for any design decision that might be patent-relevant. The bar is low: three to five sentences in each field. The notes live alongside design docs in the team's normal documentation system.
• Structured-disclosure tooling. A code scanner that surfaces technical artifacts (distinctive algorithms, architectural choices, measured behavior) at the drafting stage, with citations to specific files and commits, gives counsel a structured baseline to evaluate for inclusion in the specification at filing. The point is the same as the conception note: capture the technical record close to when it happens, in a form an attorney can later evaluate.

ObviouslyNot's scanner is built around the second approach. The output is structured for the kind of evidentiary chain a patent attorney can use when building the specification and, later, when defending inventorship if the question arises. See scan your codebase or our Section 101 Declaration Strategy guide for how the technical record feeds the prosecution record.

International alignment

The U.S. is not alone on this question. The major patent jurisdictions have aligned on "only humans" as inventors, even though they reached that result through somewhat different statutory and procedural paths.

• European Patent Office: J 8/20 (Board of Appeal, December 21, 2021) held that an AI system cannot be designated as inventor under the European Patent Convention.
• United Kingdom: Thaler v. Comptroller-General of Patents, Designs and Trade Marks, [2023] UKSC 49 (December 20, 2023) held that DABUS was not a "person" and could not be an inventor under the UK Patents Act 1977.
• Germany: Federal Court of Justice (Bundesgerichtshof), June 11, 2024, X ZB 5/22, held DABUS could not be named as inventor but allowed a filing approach naming the human applicant and identifying DABUS as a tool used in the invention process.
• One caveat on global consensus: South Africa accepted a DABUS-related filing. The reporting suggests this reflects a formality-driven patent system rather than a substantive holding that AI can be an inventor. The U.S., EPO, UK, and Germany are the substantive consensus.

Ownership is not inventorship

State legislators have begun to address ownership of AI-generated outputs. Arkansas Act 927 (2025) creates state-law ownership rules for certain AI-generated content and trained AI models. Iowa considered similar bills in 2025-2026; reporting indicates none made it out of committee. More states will follow.

These statutes address ownership: who holds the rights once an output exists. They do not address inventorship: who conceived a claimed invention for purposes of being named on a patent. The two questions are different. A contract or statute can assign ownership to a company; it cannot turn a non-inventor into a named inventor on a U.S. patent.

For founders working through cap-table mechanics, the practical message is the same as it has always been: get assignments in writing from every contributor (employees, contractors, advisors, anyone who touched the code), and accept that the named inventors on any patent must be the humans who conceived the claimed invention.

What this is not

• Not legal advice. Specific filings should be discussed with a registered patent attorney.
• Not a guarantee of validity. Even perfect documentation cannot save an invention that lacks human conception. The record helps counsel evaluate and support inventorship; it cannot create it.
• Not a recommendation to preserve every AI prompt. Preservation should be counsel-directed. AI prompts can implicate privilege, confidentiality, trade secret, and litigation considerations that a one-size-fits-all "save everything" approach mishandles.
• Not a fully settled doctrine. The "supervised AI suggestion" gray zone in the conception analysis is genuinely unsettled. The Federal Circuit has not issued an AI-specific holding. Practitioners should expect the doctrine to develop over the next several years.

What to watch next

• District-court and Federal Circuit cases testing AI-assisted conception records. None on the public docket as of June 2026, but the question is overdue for a holding.
• USPTO updates to AI-related examination materials, including possible refinements to the November 2025 guidance as practitioner feedback accumulates.
• State-law ownership statutes that may confuse founders but do not change patent inventorship. Arkansas Act 927 is the precedent; more states will follow.
• Discovery disputes over AI prompts, outputs, and prosecution/drafting records. See our emerging framework page for the framework currently developing.

Sources

Related from the resources