Can AI-Assisted Patent Inventorship Be Handled Correctly in 2026?

The Short Answer: AI Cannot Be Named, but Human Contributions Matter

Yes, but AI-assisted patent inventorship must be handled through a careful identification of human contributions. As of September 30, 2026, United States patent law does not recognize an AI system, generative model, or autonomous software agent as a patent inventor. Under 35 U.S.C. § 115 and the USPTO’s case law applying the “natural person” requirement in 35 U.S.C. § 118, the inventor named on a U.S. patent application must be a natural person. This means the application should identify the people who contributed to the claimed subject matter, not simply the person who directed an AI tool or submitted the application. A company may own a patent through assignment, but the application must still name qualifying human inventors. The practical question is not whether AI participated; modern inventions often involve software, machine learning, and automated testing. The question is which natural person made a significant contribution to each claimed invention and whether that contribution can be substantiated.

Also worth reading: How Much Should You Budget for an AI-Assisted Technical Writing Project? · How should AI-assisted processes be compared with human baselines using a representative test set? · How can founders use AI-assisted startup validation techniques to reduce risk before launch in 2026?

AI can nevertheless assist with invention, research, drafting, simulation, testing, and optimization without becoming a named inventor. The USPTO’s February 2024 revised guidance, developed after the Thinventor litigation, emphasizes that AI-assisted inventions are not categorically excluded from patentability. Instead, guidance and examination policy focus on the human contribution to each claim. A prompt entered by a human may be relevant evidence, but the mere existence of a prompt does not automatically establish inventorship. Likewise, a person who asked an AI system to “invent” a solution does not necessarily become the inventor of every output the system produces. Conversely, a technically sophisticated user who selected a proposed feature, recognized its technical effect, and supplied missing conception details may qualify. Inventorship is a legal attribution based on conception of the claimed invention, not a reward for project management or commercial sponsorship.

Why AI Output Creates an Inventorship Problem

An invention is conceived when a human forms the complete mental picture of at least one claim, according to the standard USPTO formulation drawn from cases such as Burroughs Wellcome and Pannu v. Iolab. AI breaks that familiar sequence because a model can generate combinations that were not previously recorded in its training data and may produce the proposed solution within seconds. The output can still be evaluated for novelty, non-obviousness, eligibility, and enablement, but those patentability questions differ from the question of who invented it. A generated output is not automatically patentable merely because it is novel, and it is not automatically an invention attributable to the person who supplied the prompt. The generated text may also be incomplete, incorrect, or unsupported by the underlying model. Human review is therefore needed to determine whether the output is technically meaningful and whether the final claim represents a human conception rather than an unexplained machine result.

Inventorship must also be assessed claim by claim. If a patent application contains a method performed by a processor, a trained model, and a controller that determines a technical operation, different people may have contributed different claim limitations. Naming only the business owner, principal investigator, or software developer may therefore understate or misstate the legally required inventorship. The USPTO’s revised guidance uses “claim-by-claim” analysis because inventorship is not determined by counting ideas or by assigning equal credit to a team. If a human contributes only an optimization parameter directed to a specific technical result, that contribution may be significant enough for the relevant claim. If another human contributes the overall control architecture, that person may be an inventor for a different claim. Inventorship is a legal determination, so counsel should document the basis for each conclusion and revisit it when claims are amended during prosecution.

The important threshold is not a numerical percentage such as 50 percent or 90 percent AI use. The USPTO guidance does not impose a general percentage threshold for AI contribution, and no rule says that a claim becomes non-patentable once AI involvement becomes high. The legal focus is instead on whether a natural person made a significant contribution to the conception of the claimed subject matter and whether the human contribution was more than mere attention to a machine-generated result. That distinction can be difficult to prove because the model’s operation is proprietary, changing, and sometimes impossible to reproduce. A record showing the person’s experimental work, feature selection, parameter changes, failure analysis, and technical reasoning can be more persuasive than a generic statement that the person used a particular platform. The documentation should connect each human act to a limitation in a specific claim, without implying that AI itself owns or contributes legal inventorship.

How Human Inventorship Is Evaluated in AI-Assisted Work

USPTO guidance distinguishes between contributions to conception and contributions that do not ordinarily rise to inventorship. A natural person generally may qualify by supplying the problem, the inventive concept, a specific solution, or a combination of known elements that produces a new and non-obvious arrangement. In an AI-assisted setting, that person might define the technical objective, select a data set, establish constraints, identify a missing component, and recognize why a generated approach solves the technical problem. However, those acts must be tied to the claimed invention rather than described only as high-level supervision. Telling a system to produce “a better cooling system” is usually not enough to establish conception of every feature later inserted into a claim. The person should be able to explain what was conceived, why the alternatives were rejected, and how the selected elements cooperate. The assessment is especially demanding when a generated solution is so unexpected that the user appears merely to have accepted it.

A person who solely requests an output, forwards a model-generated answer, or asks a colleague to file the resulting application ordinarily should not be listed as an inventor based solely on those acts. A person who owns the AI account, paid for the subscription, or entered business requirements likewise does not become an inventor through those activities alone. By contrast, a person who experimentally modifies the proposed solution, supplies a non-routine technical insight, or orders and combines particular features in a way that completes the claimed conception may have a qualifying role. The same principle applies to AI-assisted business methods, computer-implemented processes, and chemical or mechanical inventions: automation does not change the human requirement, but it changes what evidence is needed to show who exercised conception. The relevant inquiry is substantive, not dependent on whether the final work was typed manually or generated automatically.

Prompt language can be evidence, but it is rarely a complete answer. A detailed prompt may show that a human articulated a technical concept, yet a prompt can also merely request an outcome without demonstrating conception of the specific claim. The evidential value depends on the prompt’s specificity, the stability of the model, the nature of later selection, and whether the human understood and deliberately confirmed the result. The same prompt may produce different outputs on different runs or after a model update, making the record harder to reconstruct. Counsel should preserve dated prompts, outputs, system settings, model identifiers, notebooks, design documents, test results, and communications with technical staff. Those materials do not replace a claim-focused analysis, but they help establish what the human knew at the relevant time.

A Practical Workflow for Patent Teams

The first practical step is to identify every human and AI tool used in the inventive process before drafting the application. This includes general chatbots, coding assistants, optimization software, simulation packages, automated testing systems, and domain-specific models. For each tool, the team should record its version, operating date, provider, and role in the work. The team should then map human activities to potential claim limitations, distinguishing conception, implementation, testing, and project administration. A worksheet that records “John used ChatGPT on March 4” is not enough; it should explain that John identified a threshold problem, selected a particular multi-stage control sequence, changed two operating ranges, and recognized the resulting technical effect. This mapping allows counsel to evaluate the application before names are committed to a declaration. It also helps identify missing inventors before an oath or declaration is signed under penalty of perjury.

The next step is to reconstruct the final claimed invention independently of the AI narrative. Patent claims should be clear, supported, and written in terms a patent examiner can understand. A generated description that relies on undefined model behavior, inaccessible training data, or conclusory statements about technical improvement may be inadequate even if the human contribution is properly identified. The specification should explain the human-conceived features, their relationship, and the technical problem solved. Where an AI model is part of the claimed apparatus, the application should disclose enough about the model and its operation to support the claimed use, subject to applicable disclosure and trade-secret considerations. The application should not imply that the AI is an inventor or that a model performed a legally cognizable act of conception. It should instead present the relevant human contribution accurately and describe the machine’s technical role.

Inventorship review should continue through prosecution because claim amendments can change the analysis. A claim may initially attribute conception to one engineer but later acquire a limitation suggested by a different employee or generated by an external tool. The prosecution team should compare each amended claim with the original contribution record and obtain additional declarations where needed. Under USPTO procedures, correcting inventorship can involve a declaration or oath that identifies the correct inventors and explains the error, while the USPTO may issue notices or allow correction under the applicable rules. A correction is not merely administrative: inventorship is material to the validity of the patent, and knowingly inaccurate inventorship may create serious consequences. Teams should therefore schedule an inventorship checkpoint before filing, after major claim revisions, and before allowance.

Comparison: Human-Directed Work Versus Prompt-Only Acceptance

The following comparison is useful for organizing facts, although it does not create a bright-line safe harbor. The legally decisive factor is the human contribution to the claimed subject matter, not the label applied to the process.

Neither column guarantees a patent or a valid inventorship determination. A highly documented process can still describe an obvious or ineligible claim, while a concise process may produce a patentable and properly attributed invention if the record is legally adequate. The table is therefore a triage tool rather than a substitute for attorney review. It also should not be used to reverse-engineer a name onto an application after the fact. The correct approach is to preserve facts and determine inventorship from the actual human contribution.

Common Mistakes in AI-Assisted Patent Practice

One common mistake is treating the prompt author as the inventor automatically. This approach confuses access to a generative tool with conception of the claimed invention. Another mistake is naming the person who supervised the project, directed the budget, or owns the company, even when that person had no role in forming the technical solution. A third mistake is listing every employee who touched a project, which can overstate inventorship because routine implementation, coding, testing, and administrative work do not necessarily contribute to conception. The opposite error also occurs: omitting a person who supplied the specific inventive concept because a different employee entered it into an AI system. Inventorship is based on the human contribution, not on who had access to the model or who appeared first in a project chronology.

Teams also make the mistake of assuming that AI makes an invention unpatentable. The USPTO’s guidance does not categorically exclude inventions made with AI assistance, and software-related inventions remain subject to the existing eligibility framework. Separate from inventorship, a claim may still fail because it is abstract, lacks a practical application, is anticipated, is obvious, or is insufficiently disclosed. A stronger disclosure strategy cannot cure those defects automatically. Conversely, a technically strong claim can be undermined by an incorrect inventorship declaration. The review should address both questions independently: whether the claim is patentable under the applicable substantive law and whether the named inventors are the natural persons who conceived the claimed subject matter.

A final mistake is failing to preserve an audit trail. Model versions, system messages, retrieval sources, randomness settings, and output quality can change over time, and a later reviewer may be unable to determine what the tool produced on the filing date. Dates, identities, model names, prompt content, output files, and human edits should be captured in an organized repository with access controls. The record must be collected without exposing confidential source code, personal data, unpublished patent strategy, or third-party trade secrets. A hash, timestamp, or authenticated experiment log may help establish chronology, but it does not guarantee legal inventorship. The purpose is to support a defensible factual record when questions arise from the USPTO, a co-inventor, an assignment dispute, or litigation.

When to Act and What It May Cost

Action should begin when a team has a potentially patentable technical result, not necessarily when the first AI prompt is entered. A pre-filing review is most useful before an application is submitted because correcting inventorship later can require additional prosecution, declarations, fees, and disclosure of the error. The review is particularly important when a company uses a shared AI account, relies on an external consultant, combines contributions from multiple departments, or anticipates an assignment and investor diligence process. Companies in pharmaceuticals, biotechnology, software, electronics, materials science, manufacturing, and advanced engineering may encounter these issues more often, but the need is not limited to particular industries. A business plan, white paper, or technical roadmap should identify which outputs may support patent filings and which information is better kept as a trade secret. Patent and trade-secret decisions should be made before public disclosure, since a public disclosure can affect available patent rights depending on the jurisdiction and circumstances.

Costs vary according to the number of inventors, technical complexity, AI systems used, and extent of the prosecution record. A focused inventorship review for a relatively simple application may cost several hundred to a few thousand dollars, while a multi-claim, multi-contributor AI-assisted invention may require several thousand to tens of thousands of dollars for legal analysis and supporting declarations. Patent drafting and prosecution fees commonly run into the thousands or tens of thousands of dollars, and an international portfolio can cost substantially more because each jurisdiction has different inventorship, documentation, translation, and filing requirements. No general fee fixes the risk of an incorrect declaration, and inexpensive automated inventor-identification services cannot replace professional advice. The relevant comparison is between the cost of a documented review and the potential cost of a validity challenge, an ownership dispute, correction proceedings, or lost filing opportunities. Budget should therefore cover evidence preservation and legal review, not only model access and filing fees.

For AI technical writing projects, the most sensible practice is to document human decisions alongside the white paper or business-plan drafting process. A writing team can describe the technology, assumptions, and technical differentiation, but it should not identify inventors merely because it composed a narrative. Technical authorship, patent inventorship, and legal ownership are separate questions. The same person who explains an invention in a business plan may not have conceived it, and the company that funds the work may own the application through assignment without being eligible to be named as an inventor. Before publication, counsel can assess whether the material supports patent filings, should be retained under confidentiality, or can be released. This is especially relevant where AI-assisted tools helped structure the technical work and where a later patent application may need evidence of the original inventive process.

The Defensible 2026 Position

As of September 30, 2026, the defensible position is that AI-assisted inventions are not automatically disqualified and that AI cannot be named as a U.S. inventor. Human inventorship remains required, and the analysis must connect each natural person to a significant contribution to the claimed subject matter. The absence of a fixed AI-use percentage means teams should not rely on a numerical threshold as a safe harbor. They should instead identify the human’s technical problem, inventive concept, feature selection, modifications, and recognition of the operative solution. The legal record should make that contribution understandable without overstating what the AI or the human accomplished. This approach follows the principles reflected in the USPTO’s revised AI-assisted inventorship guidance, the Thinventor decision, and commentary from firms such as Crowell & Moring, Holland & Knight, Morgan Lewis, Baker Botts, and Passle.

International treatment may differ, and a strategy built solely around U.S. practice should not be assumed to apply worldwide. Other offices may ask for different declarations, human-contribution records, or explanations of an applicant’s role. Section 256 and other mechanisms for correcting inventorship errors may also involve procedural conditions, deadlines, and discretionary or statutory requirements that require case-specific review. The best practice is to preserve evidence, avoid premature public disclosure, obtain advice before filing, and reassess inventorship whenever claims change. AI can be a useful technical assistant, but it does not replace the human responsibility for conception, candor, and accuracy. In a disputed inventorship matter, the decisive question is likely to be factual: what did each human person contribute, and can that contribution be tied to the claims?