Who Qualifies as an Inventor When AI Helps Create a Patentable Invention?

Direct Answer to AI Patent Inventorship Rules

As of September 26, 2026, the generally applicable rule is that a patent must name one or more natural persons as inventors, and those people must have made a contribution to the claimed invention. An AI system, generative model, software platform, or autonomous machine is not itself an inventor under the patent laws of the United States, the United Kingdom, or the European patent framework. Naming an AI as the inventor is not a shortcut around filing fees, disclosure duties, or the requirement that the application identify a human rights holder.

Also worth reading: How does invention disclosure workflow automation actually work for corporate IP and product teams? · How do you integrate IP SaaS with Jira and GitHub for invention disclosure and patent filing workflows? · Can an AI Be the Named Inventor on an EPO Patent Application in 2026?

The rule does not mean that a human must perform every experimental step or write every line of code. A person can qualify as an inventor through a contribution to a claimed feature, even if colleagues performed the implementation and testing. The relevant inquiry is tied to the claims: did that person contribute to at least part of the inventive concept, and does that contribution amount to more than a merely routine or conventional activity? The contribution must also be evaluated under the applicable inventorship standard rather than authorship, employment, or who directed the project.

US law is particularly explicit. The USPTO ordinarily requires the inventor to be a natural person, and its examination guidance states that an AI cannot be a named inventor. In the Thaler v. Vidal litigation, the Federal Circuit held that an “inventor” must be a natural person, and the Supreme Court declined to review that decision in 2022. The USPTO’s 2024 Inventorship Guidance for AI-Assisted Inventions then explained how natural persons may be named when AI materially assists their inventive work. That guidance is a policy document, not a replacement for statutes, regulations, or case law, so counsel should verify whether it has changed by the filing date.

How Human and AI Contributions Are Evaluated

AI patent inventorship rules are claim-centered. Counsel first compares the final claims with the laboratory notebook, source-code history, design records, test results, prompts, model versions, and documented human decisions. A named inventor should have contributed to at least one claim, or at least to a feature incorporated into a claim, where the contribution is not routine. If AI proposed the operative solution and a human merely selected an obvious output or implemented instructions, inventorship may be doubtful even if the human supervised the entire project.

The USPTO’s AI guidance uses a 50% threshold for each claim to describe the point at which a person may be considered to have contributed to that claim, but this is not a universal “more than 50%” ownership test. The guidance considers whether a natural person made a significant contribution to every claim or at least one claim, and whether that person shaped the claimed subject matter in a way not already routine in the field at the time. Percentages elsewhere may be used for allocating economic rights, but inventorship is not determined by a majority vote or by an employee’s percentage interest in the results.

Prompting is not automatically enough. A prompt that supplies a highly specific solution may establish a contribution, while a generic request such as “optimize this process” may not. The person who formulated a nonroutine problem, selected important constraints, recognized an inventive result, or revised an output in a way that maps to a claim can have a stronger basis for inventorship. A person who supplied conventional parameters or approved a result without materially determining its claimed content may not.

No single software log proves inventorship. Records should connect a person’s dated intellectual contribution to particular technical features, but proprietary prompts and model details may be unavailable, incomplete, or difficult to export. Human declarations, laboratory notebooks, version-control commits, and contemporaneous emails often become more important when a dispute concerns which person shaped the claimed invention.

Jurisdictional Differences and Filing Consequences

The basic personhood requirement is not the only jurisdictional issue. The United States, United Kingdom, and Europe generally reject a machine acting alone as an inventor, but their treatment of human-assisted inventions, employee rights, entitlement, and the scope of routine assistance differs. The UK Intellectual Property Office requires an applicant and inventor to be a natural person, and it uses a “significant contribution” approach in AI-assisted cases. The European Patent Convention requires the inventor to be a natural person, and current practice likewise requires a properly identified human inventor.

A filing naming only an autonomous AI can fail formal examination or invite correction, but that is not the end of the analysis. The applicant may still be a corporation, university, or other legal person that owns the invention, while the inventor field identifies the natural person or persons entitled under local law. In many systems, inventorship and ownership are separate: the inventor receives an initial legal claim, while employment agreements, service inventions, assignment rules, funding terms, or statutory provisions can determine who owns the rights.

The DABUS proceedings illustrate why the distinction matters. Stephen Thaler proposed that an autonomous machine be treated as an inventor in several jurisdictions, including the United States and United Kingdom. The UK Supreme Court found in 2023 that the statutory term “inventor” referred to a natural person. The USPTO’s Thaler case produced a similar personhood result, while a parallel European proceeding was rejected because the applicant was not entitled in the way required by the European system. A valid business owner or assignee cannot simply be inserted as the human inventor unless that entity legally qualifies under the jurisdiction’s rules.

International applications also do not avoid the problem. A PCT application is governed by the selected national or regional phase, and a PCT applicant is normally a natural person or legal entity, not the inventor designation itself. Applicants should coordinate the human inventorship determination before filing because inconsistent inventor declarations can create formal, validity, or ownership complications. Patent prosecution is not the same as patent litigation, but defects in inventorship can surface during examination, post-grant proceedings, or ownership disputes.

Evidence and Documentation: Practical Steps

Start a dedicated inventorship record when the project begins, not when the application is drafted. Record the names of all technical contributors, the version of the AI tool used, the date of material outputs, the prompts or instructions supplied, the human selection and modification steps, and the experiments conducted after generation. Preserve laboratory notebooks, code commits, simulation files, test results, architecture diagrams, and emails in a timestamped repository. These records help counsel distinguish an inventive human contribution from a routine implementation detail.

Next, map the proposed claims to the evidence. For each claim, identify the human who contributed to the feature or combination, explain why that feature was not routine, and document the specific problem solved and technical result achieved. If several people contributed different portions, determine whether they should be joint inventors and whether any rights must be assigned. Inventors need not all have conceived the entire invention, but each should have contributed to at least one claim in a legally significant way.

Before filing, conduct a separate disclosure review. List the AI models, training sources where known, externally supplied code, datasets, APIs, and third-party components that influenced the invention. The USPTO’s guidance on generative-AI disclosure addresses the risk that applicants may unknowingly claim subject matter developed by others or fail to identify material AI contributions; however, the precise disclosure required depends on the technology, the claims, and the examiner’s questions. Adding every prompt to the specification is usually not a substitute for a targeted, well-supported case.

Companies should then execute or confirm assignments, employment agreements, confidentiality terms, and contractor provisions. A consultant who contributed to a claim may need to assign rights, while a person who supplied only routine administrative support may not need to be named. These are different analyses. The final declaration should reflect the actual inventive record, not the project manager’s seniority or the company’s desire to list its largest contributors.

Comparison of Inventorship and Related Rights

This table matters because companies sometimes conflate “inventor” with “owner.” A corporation can own a patent even when an employee is named as the sole inventor, and a university can be the applicant when an individual made the inventive contribution. Conversely, a company’s funding, access to the model, or control of the laboratory does not automatically make the company the inventor. The owner’s legal entity status does not eliminate the need for a qualifying natural person.

AI itself may also contain prior art or embody third-party rights. Patentability, inventorship, ownership, copyright, trade-secret protection, and data-use restrictions are separate questions. A successful inventorship record therefore should not be treated as a general clearance opinion. The same project can have a valid human inventorship determination, a patentability concern, and a contractual restriction on commercial use of an AI component.

Common Mistakes and Reasons Applications Fail

One common mistake is naming the AI tool because the model produced the headline idea. This ignores both the legal requirement for a natural person and the need to show what a human contributed. Another mistake is naming every person who touched the project, without analyzing whether the contribution affected a claim. Over-naming can produce objections and disputes, while under-naming can create a correction, invalidity challenge, or inability to enforce against a later owner.

A second error is treating a prompt as conclusive evidence. A prompt may contain the key inventive limitation, but the record should show how the person derived, selected, and tested that limitation. “The AI suggested it” is not a substitute for identifying a technically meaningful human contribution. The same error appears in the opposite direction: assuming that coding a model is automatically routine, even where the person designed a new architecture, altered the model, or discovered an unexpected technical effect.

Companies also make mistakes by relying on vendor assurances, hiding AI use, or describing software as wholly autonomous when humans selected the objective and supplied the inventive constraints. The USPTO guidance does not provide a blanket safe harbor for failing to disclose AI assistance, and patent applications must remain accurate about what was made and who did what. A generic assertion that an engineer “used AI” is not enough, but unnecessary disclosure of confidential commercial details can also create avoidable risk; counsel should balance materiality and proportionality.

Finally, assuming that filing first resolves ownership can be expensive. Inventorship disputes can affect assignments, employee confidentiality claims, university policies, and the ability to maintain a patent through every national phase. The correct sequence is to document contributions, evaluate legal entitlement, assess disclosure, and then file. If a deadline is near, a carefully reasoned interim declaration can preserve options, but it should not knowingly omit a material human contributor or misidentify the invention’s origin.

When to Act, and What It May Cost

A company should act before the first nonconfidential filing, public disclosure, publication, demo, sale, or offer for sale. Most jurisdictions have a one-year grace period for certain applicant-originated disclosures in the United States, but that rule is limited and does not cover every foreign disclosure or every disclosure by another person. Public disclosure can also destroy novelty or trigger statutory bars elsewhere. Because the public disclosure date is often a trigger for licensing and filing strategy, teams should consult counsel before presenting a result externally.

The cost depends on the complexity and the existing record. A low-volume applicant with clear contributor records may spend roughly $1,000 to $3,000 for a focused inventorship and disclosure review, while a technically complex AI project may cost $3,000 to $10,000 or more for claim mapping, contract review, and jurisdiction-specific analysis. Formal drafting and prosecution fees are separate and commonly run from several thousand to tens of thousands of dollars per family, depending on the technology, number of claims, and filing jurisdictions. Government filing, search, and examination fees vary substantially by office and entity size; small-entity or university discounts may apply, but they do not replace the inventorship analysis.

IPRaaS tools can organize evidence, contributor questionnaires, deadline tracking, and portfolio records. They cannot decide inventorship with certainty, and a dashboard that automatically labels the prompt author as inventor is not legal advice. The tool’s value is administrative visibility and reproducibility, while counsel remains responsible for legal judgment. A registry platform is most useful when it preserves the connection between people, technical versions, claims, assignments, and filing status across a portfolio.

The best time to seek advice is at the discovery stage, when a new algorithm, model configuration, or experimental result appears. Waiting until an examiner raises an inventorship objection leaves less room to reconstruct evidence and can complicate amendments. If a deadline is imminent, counsel can prioritize a rapid factual review, identify the largest evidence gaps, and reserve a full claim-by-claim analysis until the core documents are secured. Early action is not always mandatory, but deliberate delay increases cost and uncertainty.

Practical Decision Standard for Patent Counsel and Product Teams

The defensible working rule is: identify the natural person or persons who made a qualifying contribution to the claimed invention, document that contribution, disclose material AI use as required, and separately establish the owner’s entitlement. AI can create, suggest, simulate, optimize, or express technical subject matter, but the legal decision-maker should explain what the human team conceived and why the claimed result was inventive. This standard is more reliable than asking whether AI was “used,” because a person may be an inventor after using a conventional tool and may not be an inventor after relying entirely on an autonomous system.

The analysis should also be separated by jurisdiction. A filing strategy that works for a US application may not answer a UK inventorship question or a European entitlement requirement. Counsel should coordinate the claims, contributor record, disclosure position, and ownership documents before selecting countries. That does not mean every product needs the same package; it means the relevant evidence should be capable of supporting the jurisdictions in which the company intends to seek protection.

The 2026 practical baseline is therefore stricter and more evidence-based than the earlier “AI as inventor” debate. The question is no longer simply whether AI can be named; it is which people shaped the claims, what they contributed, and whether the application accurately explains the relevant technical and human contributions. Teams that treat the issue as a recordkeeping and claim-mapping problem are better positioned to protect software, hardware, biotech, and other inventions, but they should not treat a registry workflow as a substitute for legal advice or as a guarantee of patent validity.