# How Should Companies Review AI-Assisted Patent Inventorship Under Revised USPTO Guidance?

iprs.cloud · September 26, 2026

> What Revised USPTO Guidance Requires for AI-Assisted Inventions A company should review each material contribution to an invention and determine...

## What Revised USPTO Guidance Requires for AI-Assisted Inventions

A company should review each material contribution to an invention and determine whether a natural person made a significant inventive contribution to at least one claim. The USPTO’s revised guidance, announced in February 2024, applies its existing inventorship rules to AI-assisted inventions rather than creating a special rule that names the model as an inventor. Under 35 U.S.C. § 100(f), an inventor must be a natural person, while 35 U.S.C. § 115 and the associated regulations require inventors to be identified through their contributions to the claims. The operative unit of review is therefore the claim, not the output as a whole, the software release, or the business objective stated in a prompt.

**Also worth reading:** [How Do AI Patent Inventorship Rules Affect Human Inventors in 2026?](https://iprs.cloud/knowledge/how_do_ai_patent_inventorship_rules_affect_human_inventors_in_2026.php) · [How Should AI Teams Navigate Patent Eligibility, Inventorship, and Filing in 2026?](https://iprs.cloud/knowledge/how_should_ai_teams_navigate_patent_eligibility_inventorship_and_filing_in_2026.php) · [What are the current AI patent inventorship requirements for global intellectual property filings?](https://iprs.cloud/knowledge/what_are_the_current_ai_patent_inventorship_requirements_for_global_intellectual_property_filings.php)

Use of AI by an inventor is not itself a bar to patenting the resulting subject matter. A human can qualify as an inventor when that person contributes an idea to the claimed invention even if the person did not reduce it to practice or use conventional means to express or implement the idea. Conversely, supplying a prompt, operating a model, selecting from generated alternatives, or asking an engineer to implement instructions ordinarily does not make the user the inventor of every feature suggested by the system. USPTO guidance also recognizes that a claim directed merely to an abstract idea, mathematical relationship, natural law, or otherwise unpatentable subject matter cannot qualify merely because a natural person generated text describing it.

The review must distinguish inventorship from patentable subject matter and from authorship. Inventorship asks who contributed to the claimed invention, whereas patentability asks whether the claims satisfy statutory requirements such as utility, written-description support, enablement, nonobviousness, and eligible subject matter. A technically important contribution can be an inventive contribution even if the overall patent is later rejected under § 101 or another provision. This distinction is important because a company should not assume that adding a human name to an AI-generated output resolves either issue.

As of September 26, 2026, the practical answer remains that AI tools may help prepare, test, simulate, optimize, or document an invention, but the inventorship record should identify the natural persons who made significant contributions to what the claims actually cover. The review should be completed before filing whenever the relationship between human contributions and AI output is uncertain. It should also be revisited when claims are amended, because narrowing or adding claim language can change the relevant inventive contribution.", "## Why Prompting, Editing, and Human Judgment Are Not Equivalent

The hardest part of an AI-assisted inventorship review is not deciding whether a person interacted with AI, but deciding what that person added to the legally relevant conception. The USPTO’s February 6, 2024, memorandum explains that people using AI-assisted tools must make a significant inventive contribution to at least one claim and provides examples in which a natural person may qualify despite substantial computer assistance. It does not decide inventorship merely by counting prompts, revisions, drawings, or hours worked. A long prompt can describe a complete invention without demonstrating that the prompt author originated the claimed concepts, while a short instruction to test a specific experimental idea may precede a major inventive contribution.

Prompt authorship should be separated from conception of the claimed subject matter. A person who asks a model to propose a material, geometry, control method, or circuit arrangement may contribute an inventive idea if the person framed a specific problem, brought in relevant technical knowledge, recognized the proposed solution, and caused that solution to be incorporated into at least one claim. Merely describing a desired result is not enough when the only output is a general suggestion. It is also not enough for a person to adopt a result without understanding how it relates to the claimed invention, because inventorship is determined by contribution rather than by unexplained receipt.

Editing generated text does not automatically establish inventorship. Correction of grammar, formatting, or obvious errors normally does not amount to a significant inventive contribution. Material modifications may do so, particularly when the editor changes technical features, adds new relationships, or supplies a missing conception that becomes part of a claim. Selection can also matter, but selecting one of many conventional options is not likely to establish much, whereas selecting a generated arrangement because it solves a technical problem in a particular way may support inventorship. The record should explain the reason for the selection rather than relying on the fact that a person chose it.

A defensible process compares each claim with contemporaneous evidence of human contribution. Suitable evidence may include dated laboratory notes, design records, issue tickets, source-code history, test results, messages, presentations, and explanations from participants. A witness statement can help but should not substitute for contemporaneous records where those records exist. The company should preserve failed AI outputs as well as accepted ones, because they may show whether the human directed the conception, selected among conventional variants, or received an independently critical idea from the system.", "## How to Conduct a Claim-by-Claim Inventorship Review

The first step is to identify the contemplated claims, not merely the product description or internal invention disclosure. Claim construction during prosecution can differ from ordinary product terminology, and a contribution to a feature omitted from every claim may not affect inventorship. Counsel should create a mapping between each claim element and the people who contributed to that element, with separate treatment of elements derived directly from AI output. If the application includes several independent claims, the review may identify different inventors; the USPTO may permit multiple inventors, but every named inventor must have made a significant contribution to at least one claim.

The second step is to reconstruct the human contribution from reliable records. For each person, reviewers should ask what problem the person identified, what constraints they supplied, what concepts they proposed, and what they rejected or changed. A useful record distinguishes an original idea from an instruction to use known methods. It also distinguishes experimental direction from automation. Dates matter, so the team should preserve the chronology of model access, generated proposals, human comments, experiments, and formal claim drafting.

The third step is to evaluate legal significance rather than effort. A person may spend substantial time testing an invention without originating the claimed concept and may still not be an inventor. Conversely, a brief contribution involving a central claim limitation can be legally significant. The review should document why each contribution is or is not significant rather than using a minute count. Company management titles, such as chief technology officer, do not establish inventorship, and an employee’s status as the project owner does not replace the contribution analysis.

The fourth step is to correct the record before a problem develops. If the wrong person is named, a missing inventor is omitted, or an AI-related detail was concealed, counsel should evaluate the available correction mechanisms and the risk associated with incorrect inventorship. A material inventorship error may affect enforceability against the incorrect parties, while failure to correct the wrong inventorship can create exposure under § 256. The practical goal is not to add every person who touched the project, but to name every natural person required by the claims and omit those who made no qualifying contribution.", "## Comparison of Inventorship Review Approaches

Several review approaches are available, and each has a different balance of cost, speed, evidentiary strength, and suitability. A prompt-only review is fast but cannot reliably identify the origin of technical ideas. A contribution review is more demanding, while a claim-centered forensic review is usually appropriate only when inventorship is disputed or the technology relies heavily on AI. The table compares the main alternatives rather than treating one method as universally correct.

| Feature | Prompt-and-output review | Structured contribution review | Claim-centered forensic review |
| --- | --- | --- | --- |
| Main evidence | Prompts, chat history, generated responses | Contribution records, notebooks, tickets, design history | All prior evidence plus claim mapping and witness accounts |
| Typical time | Hours to several days | Several days to two weeks | Several weeks for a complex matter |
| Best use | Early triage of low-risk use | Routine product and patent review | Contested inventorship, complex AI claims, or possible § 256 issue |
| Main weakness | Mistakes interaction for conception | Can miss undocumented oral contributions | More expensive and operationally demanding |
| Human decision required | Yes | Yes | Yes, with legal analysis of each claim |
| Likely cost | Internal staff time | Internal review plus limited counsel | Counsel, technical experts, and sometimes e-discovery support |

A structured contribution review is the normal middle ground for a B2B company using AI in research. It can be performed before drafting begins, reducing the risk that a product team names an engineer solely because the engineer ran the experiment. A forensic review becomes sensible when a former contractor alleges inventorship, when two teams dispute who supplied a core concept, or when the claims are unusually broad and the contribution of each technical module is unclear.
The options are not substitutes for legal judgment. Automated classification, prompt scoring, or an AI-generated summary can organize evidence, but it cannot decide inventorship conclusively. The system should not make a final determination based on stylistic features, job titles, or the number of edits. Companies should keep the source evidence and allow patent counsel to apply the governing rules. This division of labor is particularly important for IP registry and workflow products, where auditability and reproducible records may be as valuable as the apparent efficiency of the tool.", "## Common Mistakes That Create Inventorship Risk

One common mistake is naming the person who submitted the prompt as the inventor of every generated feature. The error assumes that control over the software equals conception of the invention. If the prompt merely asks a model for solutions to a previously defined technical problem, the resulting ideas may not have been conceived by the prompt author. Another mistake is naming the person who performed the final mechanical conversion of an idea into a drawing or code, when that person merely implemented a design supplied by someone else. Implementation can contribute when it reveals a new inventive concept, but routine execution generally does not.

A second mistake is treating every meaningful idea as requiring a claim-based inventorship analysis only at filing. Claim scope can change during prosecution, and a person who contributed to a limitation later deleted from the claims may no longer need to be named for that subject matter. That does not mean the person’s work is irrelevant to the application; it means the inventorship analysis must follow the claims. Inventorship is also distinct from ownership, so employment agreements, assignment clauses, and contractor agreements should be reviewed separately. A company can own a patent while disputing the identity of an inventor.

A third mistake is preserving only the final prompt and deleting intermediate records. A final prompt may conceal earlier human steering, and generated outputs may not show which technical insights came from a person. The team should retain access logs and version histories where business and privacy policies permit, while applying appropriate access restrictions to confidential source code, trade secrets, personal data, and unpublished patent material. Records should be organized by date and claim element so that a later auditor can reconstruct the decision without relying on memory.

The fourth mistake is assuming that an AI disclaimer cures the problem. Saying that a tool assisted the work does not identify the inventorship contribution. The correct response is a documented analysis of the human contribution, followed by accurate naming and appropriate correction if needed. Concealing AI use can also create separate ethical, contractual, or regulatory concerns in some contexts, even though AI assistance itself is not generally disqualifying under USPTO inventorship rules.", "## When to Act and What the Review May Cost

The review should begin when a project moves from exploratory AI use toward a potentially patentable technical improvement. That may occur before an invention disclosure is approved, when a new formulation is narrowed to a few candidate approaches, or when a company decides to file a continuation or a new application based on additional experiments. Waiting until the application is about to be filed compresses the time available to locate contributors, reconstruct the chronology, and correct an inaccurate disclosure. For a fast-moving product, a preliminary review within two to four weeks may be sensible, although a complex claim set can require more time.

Cost depends mainly on the number of claims, the number of contributors, the quality of records, and whether experts are needed. A small internal review using an existing invention-disclosure process may require only several hours of engineer time and limited counsel review. A moderately complex AI-assisted invention may take three to ten business days and involve patent counsel, a technical reviewer, and records support. A contested matter with former employees, source-code disputes, or several competing technical theories can require weeks of work, expert analysis, and potentially litigation-oriented discovery; firms may charge several thousand to tens of thousands of dollars or more for that work.

These figures are planning ranges, not USPTO fees or fixed quotations. The USPTO charges official filing, search, examination, and issue fees, but the AI-assisted inventorship review itself is usually professional labor rather than a separate government service. Companies should therefore budget for evidence collection, technical review, claim analysis, and document correction. A registry SaaS platform can reduce tracking effort, but the platform’s subscription price should not be confused with the cost of determining legal inventorship. A low-cost tool that produces a report is not automatically better than a more expensive process supported by qualified counsel.

The best time to act is before a confidential filing, public demonstration, publication, or offer for sale. Public disclosure may affect foreign filing opportunities under arrangements such as the Paris Convention, and a rushed U.S. filing can create avoidable foreign-filing loss. The team should also act before transferring rights to investors or customers, because inventorship findings may affect representations, assignments, and indemnity obligations. Speed matters, but not at the expense of a reliable record.", "## What a Defensible AI-Assisted Patent Workflow Should Preserve

A defensible workflow combines an invention disclosure, an AI-use record, a claim map, and a human contribution statement. The invention disclosure should state the technical problem, the proposed solution, the alternatives considered, and the experimental evidence. The AI-use record should identify the tool, model version when known, date, purpose, important prompts, and material outputs, while avoiding storage of information that should not be retained. The contribution statement should connect the human work to individual claim elements and explain why the contribution is significant.

Counsel should review the record at several gates rather than at the end. The first gate occurs when a disclosure is accepted, the second when claims are drafted, and the third when material amendments are made. At each gate, the team should check whether the inventors are still correct, whether the specification accurately describes the human contribution, and whether an omitted inventor could later challenge the application. A change in claim scope should trigger a focused inventorship review even if the original review was satisfactory.

The workflow should also separate legal requirements from internal policies. An internal rule may require two reviewers for high-value inventions, but internal approval does not determine who is an inventor under federal law. Similarly, a company may require disclosure of all generative-AI use, while the USPTO question remains narrower: which natural persons made a significant inventive contribution to the claims. Clear separation prevents a useful internal control from being mistaken for a legal conclusion.

For a B2B intellectual-property platform, the resulting record should be exportable, permission-controlled, and linked to the relevant application or registry record. Auditability can help counsel and product teams reproduce decisions, but it should not convert a platform’s workflow label into a determination of inventorship. The platform should support the review and preserve evidence; qualified counsel should make the legal judgment. That approach is slower than copying a name from a prompt history, but it is more credible when ownership, validity, or inventorship is later examined.", "## Practical Answer for Counsel and Product Teams in 2026

The direct answer is that companies do not need to avoid AI-assisted invention, but they do need to identify the natural persons who made a significant inventive contribution to the claims. A person may use AI extensively and still be an inventor if the person originated or materially shaped a claimed concept. A person may use AI minimally and still not be an inventor if the person merely requested generic output or implemented someone else’s design. The model itself is not a named inventor under the U.S. framework discussed in the cited guidance and case law.

The most reliable next step is a claim-centered review supported by dated evidence. Counsel should map claim elements to human contributions, separate conception from routine implementation, and identify unresolved disputes before filing. If the current disclosure names the wrong people, the team should assess correction promptly under the applicable rules, including § 256 where appropriate. If a third party may claim inventorship, preserving records and obtaining technical review is more useful than relying on an automated score or an informal assurance that the project manager owned the work.

The answer should not be presented as a promise that AI use is harmless in every jurisdiction. This analysis is focused on U.S. inventorship practice, and other countries may apply different legal tests or treat human contribution differently. Patentability, ownership, confidentiality, export controls, data protection, and professional-responsibility rules can also create separate obligations. Companies operating internationally should obtain jurisdiction-specific advice rather than assuming that a U.S. workflow settles global requirements.

The practical takeaway is modest but clear: treat inventorship as a documented legal determination, not as a field automatically populated by the person who ran the AI tool. The more central a human contribution is to a claim, the more important it is to explain what the person originated and why that contribution qualifies. That discipline gives counsel a stronger file, gives product teams clearer records, and avoids making a weak inventorship decision that could become expensive later.

## Quick answers

### Can a human be a patent inventor if AI generated most of the technical solution?

Yes, potentially, if the natural person made a significant inventive contribution to at least one claim. Extensive AI use is not itself disqualifying, but prompting, testing, or selecting output must be evaluated against the claim-specific conception analysis. The person should document what technical ideas they supplied, changed, or originated.

### Does being the project manager or software engineer make someone the patent inventor?

No. Inventorship depends on contribution to the claimed invention, not job title, employment, ownership, or effort. A project manager who merely coordinated the work and an engineer who implemented someone else’s design may not qualify for particular claims.

### When should a company review AI-assisted inventorship?

The review should occur before filing and whenever material claim amendments or new contributors enter the process. An earlier review preserves evidence and creates time to correct the record. It is especially important before a public disclosure, investment transaction, or assignment of patent rights.

### What happens if the wrong inventors are named?

A missing or incorrect inventor can create enforceability problems and may affect the rights of the incorrectly named party. Section 256 of 35 U.S.C. provides mechanisms for correcting certain inventorship errors, but the available correction depends on the facts. Counsel should evaluate the issue promptly rather than waiting for an office action.

### Can an automated SaaS tool determine patent inventorship by itself?

No. Software can organize prompts, records, versions, and claim mappings, but the final determination requires legal analysis of significant inventive contribution. A tool may reduce administrative work while qualified patent counsel applies the governing rules to the technical evidence.

Canonical: https://iprs.cloud/knowledge/how_should_companies_review_ai-assisted_patent_inventorship_under_revised_uspto_guidance.php
Markdown: https://iprs.cloud/knowledge/how_should_companies_review_ai-assisted_patent_inventorship_under_revised_uspto_guidance.php/index.md