Can AI Be an Inventor Under U.S. Patent Rules in 2026?

Direct Answer to the AI Inventorship Question

No. Under U.S. patent law, an AI system cannot be named as the inventor of a patent application. The USPTO requires the inventor disclosure to identify a natural person who conceived at least one claim of the claimed invention. In the DABUS litigation, Stephen Thaler sought patents naming an AI system called DABUS as the sole inventor, but the courts held that “inventor” refers to a natural person. The Federal Circuit affirmed that conclusion in 2025, and the broader rule is also reflected in the Patent Cooperation Treaty, under which an inventor must be a natural person. An AI-generated idea may therefore support a patent application, but a human must perform the legally required inventive work and be disclosed honestly.

Also worth reading: How Do You Build an AI White Paper Workflow That Produces Accurate, Reviewable Documents? · What Should an AI White Paper Checklist Include in 2026? · How Should Enterprises Govern AI-Generated Documents in 2026?

That does not mean every output from an AI tool is unpatentable. A human may qualify as an inventor if they contribute enough to the conception of the claimed subject matter, not merely select a result, request a variant, or file the paperwork. The practical problem is evidentiary: patent inventorship is assigned claim by claim, while AI tools may produce many outputs from opaque training data and user instructions. A business should not assume that the person who commissioned the work, paid for the model, or owns the resulting files automatically qualifies as the inventor. As of 30 September 2026, the safe rule remains to identify a qualified human contributor and prepare a written record before filing.

Why the Human-Inventorship Requirement Exists

The requirement protects more than an administrative form. U.S. patent law rewards people who contribute to the conception of an invention, and the inventor disclosure creates a statutory obligation of candor rather than a credit line based on project ownership. The inventor need not be the company, the engineer who implemented the invention, or the person with the most technical expertise. The relevant question is who conceived the claimed subject matter. That analysis may become difficult when a developer supplies a prompt, combines several tools, trains or fine-tunes a model, chooses training data, filters generated results, and then claims the selected output as an invention.

AI also changes how teams document conception. A conventional notebook can show sketches, experiments, revisions, failed configurations, and the point at which a person recognized a workable technical solution. A generative model may return an answer without exposing how it combined ideas, whether a feature was derived from protected material, or which instruction produced each element. That opacity does not transfer inventorship to the model, but it can make it harder to prove the human contribution and satisfy the duty of disclosure. The WIPO’s materials on AI and intellectual property similarly emphasize that questions extend beyond inventorship to ownership, data, confidentiality, and the legal treatment of AI-assisted work.

The USPTO has also addressed use of AI tools in prosecution through the 2024 AI-related patent practice guidance. That guidance focuses on identifiable human contributions and warns practitioners not to submit arguments or submissions that a person did not make. It does not create a general safe harbor for AI-assisted inventions. Although the primary intellectual-property concern is inventorship, applicants must still avoid false statements, unsupported factual assertions, and undisclosed material that could affect examination or validity.

What a Human Must Contribute to Qualify as Inventor

A person may qualify through conception of at least one claim, even if a computer executes, optimizes, simulates, or implements the work. A natural person might qualify by recognizing a non-obvious technical relationship, identifying the operative mechanism, selecting a specific solution from generated alternatives, or deriving a claim that goes beyond the instructions given to the AI. The contribution does not have to cover every element by itself, and an application may name several human inventors. The central test remains tied to the claimed invention rather than the commercial value of the entire project.

A human who merely enters “invent a new battery-management algorithm” into a general-purpose system is unlikely to establish conception merely by accepting the output. The person may also fail to qualify if they only ask for summaries, translations, formatting, or implementation code already dictated by a third party. Selection can require closer analysis. If a technical expert tests many machine-generated alternatives and conceives a specific arrangement that was neither requested nor apparent from the tool’s output, that person may have a stronger claim to inventorship. Legal advisers would examine the timeline, prompts, model configuration, source material, experiments, and contribution made by each person.

The table is a risk guide, not a bright-line test. Selection is sometimes inventive and sometimes not. Patent eligibility is also separate: even a properly named human inventor cannot obtain a patent for subject matter excluded by 35 U.S.C. § 101 merely because the invention used AI.

Disclosure, Patent Eligibility, and AI-Generated Code

Correct inventorship is only one part of a defensible application. An AI-assisted invention must still satisfy the requirements of novelty, non-obviousness, utility, written-description support, enablement, and patent-eligible subject matter. For software and AI inventions, examiners may ask whether the claims recite a practical technical improvement rather than an abstract idea implemented with generic computer components. The USPTO’s 2024 subject-matter eligibility guidance places particular weight on limitations that improve computer functionality or address a technical problem. Whether an AI system supplies wording, equations, source code, or a new process can therefore matter to both inventorship and eligibility.

AI-written code creates additional risks because an application must describe and enable the claimed invention without relying on knowledge that existed only inside a model or later became available publicly. The patent laws generally require the application itself to provide sufficient written description and enablement, although § 112 issues can arise later rather than during examination. Before filing, the team should verify every material parameter, identify dependencies, document data and model versions, test critical functions, and compare the claimed implementation with earlier code. A professional should not submit code containing third-party copyrighted material, credentials, private data, or restrictions that prevent lawful public disclosure.

Using an AI tool also raises prior-art questions. The fact that a model learned from public data may make some features combinations of known methods, even if the combination is unexpected. Confidential internal documents may reveal earlier disclosures if fed to an external service. The USPTO guidance recommends recognizing when an AI tool cannot distinguish confidential from non-confidential information and taking appropriate precautions. These issues may not establish that a human is not an inventor, but they can undermine validity or produce prosecution disputes later.

Inventorship, Ownership, and Employment Are Different Questions

The company that funds AI-assisted research does not necessarily own every resulting patent right, and a human inventor does not necessarily have the right to exclude competitors. Inventorship determines who made the required contribution; ownership depends on contracts, employment relationships, assignment rules, and applicable law. An employee’s patent-related inventions may belong to the employer under an agreement, an invention-assignment clause, or statutory rules specific to the jurisdiction and employment category. A founder should not use the phrase “our AI invented it” as a substitute for documenting who actually made the claim.

Contracting is especially important when several companies, universities, contractors, open-source contributors, or data providers participate. Agreements may allocate invention rights, licensing rights, model inputs, training data, evaluation results, improvements, and publication responsibilities. An assignment signed only after substantial conception can be less useful than a clear agreement reached before the project begins. The business should also determine whether its AI vendor claims rights in prompts, outputs, fine-tuned models, or feedback. Most enterprise services do not give a customer ownership of the model merely because the customer uses it, and the output allocation may depend on the service terms.

Inventorship reviews should be performed before an application is prepared because naming a person as inventor is a formal legal act, not a courtesy. A company may be the eventual owner, but it cannot cure inventorship by listing a nominal employee who performed no inventive work. Conversely, the individual who conceived a claim should not be omitted merely because the business funded the work. Professional responsibility for preparing or prosecuting the application belongs to a registered patent practitioner, not to the AI system.

A Practical Pre-Filing Process for AI-Assisted Inventions

The first step is to preserve an invention history rather than starting with a list of names. Keep dated prompts, system messages where available, model names and versions, retrieval sources, generated outputs, human edits, selected alternatives, experimental results, and meeting records. Mark which person made each creative or technical decision. This record will not automatically prove inventorship, but it can reveal whether a person merely approved a result or supplied claim-level conception. Access permissions, retention policies, and confidentiality reviews should be set before sensitive documents are uploaded to a consumer AI service.

The second step is to separate assistance from invention. A tool may be useful for brainstorming, code explanation, drafting, test generation, or converting an experimental description into formal language without becoming the source of inventorship. The team should document what the human knew before each use, what the machine returned, and what new technical contribution followed. A model that independently suggests a specific novel combination may create uncertainty because its full derivation cannot be inspected. That uncertainty is a reason for review, not a reason to list the model as inventor. If several people contribute, the patent group should map those contributions to the proposed claims.

The third step is a coordinated legal and technical review. A patent attorney or practitioner should assess inventorship, ownership, § 101 treatment, disclosure, conflicts, data provenance, and foreign-filing implications. A technical expert should verify that the claims correspond to a working or credible implementation. The pre-filing review should occur before the priority deadline because rushed disclosure can force the applicant to make unsupported statements or omit important material. Prompting and filing on the same day leaves little opportunity to reconstruct who conceived what and when.

Common Mistakes and When to Act Early

One common mistake is treating prompt authorship as equivalent to invention. Writing a long or sophisticated prompt can increase the chance of obtaining a valuable result, but prompt length is not a legal measure of conception. Another mistake is assuming that the model cannot own a patent and that the patent therefore belongs to the business. Both premises confuse different legal concepts. Businesses also make the opposite error: they list every engineer, executive, or employee, overlooking that inventorship is claim-specific and that adding an uninvolved person may constitute an improper inventor disclosure.

A second error is treating the patent system as a secrecy test. Information sent to an external AI provider may leave the organization and may be retained or used under the provider’s policies. Confidential information should be evaluated under applicable trade-secret obligations, contracts, and USPTO disclosure rules. A third error is allowing an AI tool to draft legal declarations without human verification. Disclosure to generative-AI tools can create prosecution risk when a filing relies on a generated analysis that the practitioner did not independently make or verify.

Early action is appropriate when a human contribution is limited, several people used multiple models, the tool trained on a proposed competitor’s data, or a third party supplied substantial technical content. A focused review is also sensible before public demonstration, sales, conference presentation, offer for sale, publication, or an international filing because some foreign jurisdictions have a narrow absolute novelty exception. The Paris Convention generally gives a 12-month period after an initial filing for certain applicants to seek corresponding protection elsewhere, but that period is not a license to postpone domestic and international patent decisions. If the commercial deadline is under 60 days, a patent specialist should be engaged immediately.

Cost, Alternatives, and Professional Advice

There is no mandatory USPTO fee for a preliminary AI inventorship opinion because that is not a standard USPTO service. Market estimates for a focused attorney analysis of a modest AI-assisted software or business-method invention commonly range from about US$1,500 to US$5,000, while more complex matters involving many claim sets, training data, or multiple parties can cost substantially more. Formal drafting, prosecution, office actions, and litigation are separate services with separate fees. A full software patent application may be quoted in the low five figures, but the quoted amount should be tied to technical complexity and search work rather than an assumed AI discount.

A lower-cost alternative is a human-led technical evaluation that records the principal contribution and then a short, claim-focused attorney review. This may be sufficient before deciding whether patent protection is economically sensible. The alternative is not to file a provisional application and defer the inventorship issue. Provisional applications can be useful for preserving an early date, but they still require honest identification of inventors and sufficient support for what is claimed later.

Whether to pursue protection depends on expected revenue, remaining development time, reverse-engineering difficulty, competitors, and the strength of the technical concept. Some AI improvements are better protected through trade secrets, especially for model weights, datasets, tuning processes, or operational thresholds that are not disclosed publicly. Others may fit copyright, contractual restrictions, or defensive publication strategies, although those alternatives do not create a right to exclude independent commercial use. Business-method and abstract AI claims may also be more vulnerable than claims directed to a concrete technical improvement. Legal advice should therefore assess both patentability and whether filing is the best investment.

Bottom-Line Rule for Businesses

As of 30 September 2026, AI cannot be an inventor under U.S. law or the PCT system, and an application naming only a machine would not be a valid way to secure patent rights. AI use does not automatically eliminate patentability, but a natural person must make the legally relevant inventive contribution. The mere ownership of an account, payment of a subscription, supervision of a research project, or copying of generated material into an application is not necessarily enough.

The defensible approach is simple but demanding: document every significant human contribution, evaluate the claims rather than the project as a whole, check ownership and confidentiality, verify technical support and patent eligibility, and identify evidence of any machine-generated material that could affect disclosure. Do this before filing and preferably before the first confidential or public disclosure. This is a legal issue that can affect the patent itself, not merely a housekeeping detail, so an unresolved question about a human’s contribution should lead to a claim-specific review by a registered patent practitioner.