Direct Answer: AI Output Cannot Be Named as an Inventor
As of October 2, 2026, no AI-generated output qualifies as a human inventor merely because an AI system selected a technical solution, generated claims, or produced the evidence submitted to a patent examiner. In the United States, an inventor must be a natural person, and inventorship is determined by the human contribution to the conception of the claimed invention. AI-generated experimentation, drafting, analysis, or claim language generally does not by itself establish inventorship; however, it does not automatically disqualify a patent if a natural person made and can substantiate the required inventive contribution.
Also worth reading: Can AI Be an Inventor, and How Should Patent Inventorship Be Determined for AI-Assisted Inventions in 2026? · How Do AI Patent Filing Controls Affect Inventorship, Disclosure, and Filing Strategy? · What is the current USPTO guidance on AI patent inventorship as of 2024, and how should inventors and practitioners comply with these requirements?
The controlling U.S. authority includes the Supreme Court’s decision in Thaler v. Perlmutter and the Federal Circuit’s decision in Thaler v. Vidal, which involved DABUS, an AI system configured to produce patent-eligible inventions autonomously. Those decisions rejected the proposition that an “inventor” under the Patent Act may be a machine. The USPTO’s August 2024 guidance on AI-generated inventions then instructed examiners to ask who caused the claimed invention and whether that human supplied enough contribution to be an inventor under Pannu v. Iolab and Hess v. Hartzfeld. Accordingly, evidence of AI inventorship evidence should focus on a documented human contribution—not proof that the AI was creative.
What Counts as Evidence of a Human Inventorship Contribution?
Inventorship attaches to the conception of at least one claim, not necessarily to the commercial idea, the request made to an AI tool, or every later prosecution step. Evidence should identify what was known before the AI operated, what instruction or technical input the human supplied, and how that person evaluated, selected, and revised the AI’s output into a claimed solution. Contemporaneous notes, design files, lab records, source-code commits, engineering calculations, model-selection records, drawings, employee assignments, and dated correspondence can show the path from human contribution to the filed claim.
The human contribution must be more than a high-level prompt such as “invent a battery with 20% greater capacity.” A useful record may explain why a chemist selected a particular electrolyte, how the person recognized an unexpected voltage improvement, or why an engineer rejected several AI-proposed structures. For software inventions, it might document a nonroutine algorithm, data relationship, architecture, or performance improvement conceived by the human and implemented with coding assistance. If two people collaborated, inventorship may include each person only for claim limitations each actually conceived. Inventorship is narrower than authorship and is not awarded to a person merely because they owned the hardware, funded the project, supervised the work, or performed routine implementation dictated by another inventor.
Why AI Prompts, Logs, and Patents Are Usually Not Enough Alone
A patent application, AI chat transcript, prompt history, or repository snapshot can be probative, but it may not establish who conceived the claimed features. Models can produce plausible combinations without a retrievable record explaining why a particular combination was selected. A long prompt also does not prove that the prompt author conceived the claimed invention; it may express a result already supplied by the model or another person. Conversely, a concise prompt may support inventorship if surrounding records show that the person already possessed and applied a specific inventive concept.
The examiner’s central inquiry is whether the application identifies a natural person who contributed to the conception of the claim. The guidance does not impose a general requirement that the human contribution be greater than the AI’s contribution, nor does it create a numerical threshold expressed as a percentage. Under the USPTO’s 2024 update, a person must contribute to more than the provision of instructions for using an AI tool where that person’s contribution is tied to an inventive concept. The practical burden is therefore factual: create a record capable of connecting a person’s pre-existing knowledge and judgment to specific claimed limitations. Evidence created after the fact should be candid about when it was made and should not be retrofitted to manufacture a narrative.
Comparing Human-Only, AI-Assisted, and Human-Claimed Automation
There are several possible workflows, but they should not be treated as legally interchangeable labels. The table below compares the usual inventorship position, the quality of supporting evidence, and the principal risk for each workflow.
| Feature | Human-Only Invention | AI-Assisted Invention | Human Claimed Invention Generated Autonomously by AI |
|---|---|---|---|
| Usual inventorship | Natural person who conceived claimed features | Natural person who made a qualifying inventive contribution | No independently qualifying human inventor |
| AI role | None | Search, drafting, simulation, coding, testing, or analysis | System conceives claimed solution without sufficient human contribution |
| Best evidence | Lab notebook, dated designs, calculations, correspondence | Prompt context plus technical records showing human selection and revision | Outputs, logs, and system architecture generally do not replace a human inventor |
| USPTO treatment | Conventional examination | Examined using human-inventorship principles | Likely rejected or objected to for lack of a proper inventor |
| Principal risk | Incomplete or inconsistent records | Unsupported assertion that the human merely curated output | Attempt to name an AI system as inventor or overstate minimal human input |
A Practical Documentation Method for AI-Assisted Patent Work
Start before prompting the system by recording the technical problem, the human’s proposed solution, relevant data, and the expected mechanism. Preserve exact prompts, model name and version, access date, temperature or configuration if material, outputs, rejected alternatives, and the reason for selecting the final option. This record should distinguish existing technical knowledge from knowledge discovered only after testing. Where an AI proposes a potentially inventive relationship, document what the human did with it: whether the person supplied missing constraints, performed a validating experiment, recognized a functional effect, or selected and refined a nonroutine feature.
Translate the narrative into claim-oriented evidence without artificially changing the invention after filing. A matrix may connect background records, human contributions, experiments, and issued claims, but the claims themselves remain the legal measure. Screenshots can be useful, although native files with timestamps and system metadata are often stronger; an image of a conversation may not show deletions, edits, or the author. Confidential source code, personal data, and third-party material should be protected before filing. Patent disclosure should avoid submitting a prompt as proof of ownership or priority when no corresponding foreign filing exists, because U.S. patent applications generally require their own filing within 12 months of any relevant provisional filing to claim the provisional date.
Audit inventorship before execution and again after examiner questions. The inventors should review every claim, not merely the system-level description, because a person may have conceived the core device but not a later-added algorithm, module, or manufacturing feature. Correcting inventorship after noticing an improper name has different consequences from correcting a misspelled name: a named inventor who did not invent may need to be removed through correction, while a true inventor omitted from an application may be added in appropriate circumstances. Counsel should determine which procedure applies because fees and enforcement risk differ.
Common Mistakes in Recording and Proving AI Contributions
The most common error is treating AI use as a binary condition: either no AI was used, or the AI is the inventor. That framing ignores the claim-specific human contribution on which U.S. inventorship depends. Another error is collecting only polished outputs after the invention is complete. A person may remember the conceptual chain but be unable to prove timing, and an examiner can compare the asserted contribution with contemporary records. Excessive prompts also do not prove authorship if the prompt states the entire solution, but they can undermine the claim that the human conceived it.
Inventors sometimes conflate experimentation with conception, patentability with inventorship, or significance with inventorship. Testing many AI-generated alternatives may be ordinary diligence rather than conception of each result. A dramatic commercial benefit does not establish who invented the claimed features, and an AI’s sophisticated implementation does not transfer legal authorship from the engineers who designed the system. Employers also should not assume that an employee automatically owns the relevant inventive contribution or that the employee is the proper inventor in every case. Employment, commissioning, and assignment terms should be reviewed separately from the inventorship analysis.
Finally, do not describe the application as “invented by AI,” “AI-generated,” or “inventor: DABUS-like system” without explaining the legal role of the human. Public descriptions can trigger ownership, disclosure, export-control, or prior-public-use questions even when they were not intended as patent disclosures. Use terminology such as “AI-assisted,” “human-directed,” or “computer-implemented” only when those terms accurately describe the process.
When to Act and What It May Cost
Create the evidence file at the beginning of the project, not when a dispute or examiner objection arrives. A practical checkpoint occurs before the first nonprovisional filing, with another review after major claim amendments and before a patent grant. In international work, prepare country-specific records because inventorship standards can differ. Germany and many other jurisdictions recognize employees as inventors, while the United States uses a natural-person, claim-specific standard; one global declaration should not be assumed to resolve every jurisdiction’s requirements.
Costs depend on the complexity and dispute level. Basic recordkeeping can cost little beyond the engineer’s time and suitable repository storage, while exporting model logs, reconstructing source-code history, conducting a formal contribution audit, or preparing a declaration may require $1,500 to $10,000 or more. A high-volume enterprise program can be cheaper per invention once templates and automated provenance tools exist. Preparing a properly supported U.S. application commonly involves substantial attorney fees in addition to USPTO fees, and those prosecution costs should not be confused with the cost of proving inventorship. The 2026 USPTO filing and examination fees should be checked against the current fee schedule because they change over time.
If inventorship is contested, the relevant remedy may be a correction, a duty of disclosure analysis, ownership litigation, or a defense to enforce an improperly named patent. Prompt action does not mean disclosing privileged strategy indiscriminately, but delaying the issue can magnify credibility and damages problems. The safest course is to obtain a documented review from registered patent counsel before naming, removing, or adding an inventor in response to AI-related evidence.
Bottom-Line Standard for Reliable Evidence
The strongest evidence of AI inventorship evidence is an auditable chain showing that natural persons conceived the claimed limitations and that AI performed an identifiable supporting role. That chain should contain dated technical records, human explanations of selection and revision, relevant experiments or implementations, and a claim-by-claim comparison. Logs help only when their legal significance can be explained; neither a large volume of prompts nor a dramatic AI demonstration establishes inventorship by itself.
For an applicant, the key question is not “Did AI help?” but “Which natural person contributed to the conception of each claimed feature, and what contemporaneous evidence supports that contribution?” For a challenger, the key question is whether the record instead shows that the claimant had no qualifying human contribution or that a named inventor merely supplied broad instructions. Applying the current U.S. natural-person rule makes the examination rigorous and technology-neutral: advanced tools may be used freely, but the Patent Act still requires an accountable human inventor.