# What Evidence Proves Human Inventorship in AI-Assisted Patent Applications?

iprs.cloud · October 2, 2026

> The Direct Answer to AI Patent Inventorship Evidence Evidence of human inventorship in an AI-assisted patent application should show that one or more...

## The Direct Answer to AI Patent Inventorship Evidence

Evidence of human inventorship in an AI-assisted patent application should show that one or more natural persons made the claims to the invention and contributed to conception of the claimed subject matter. That evidence can include dated laboratory notebooks, source-code commits, design files, model and prompt records, employee assignments, invention-disclosure forms, test results, and witness declarations. The strongest file connects each inventor to a specific claim element, explains the person's role, and shows when the conception occurred. An AI system's ability to generate output does not itself establish a legally recognized inventor under current U.S. practice, because an “inventor” in a patent must be a natural person. The USPTO's 2023 AI inventorship guidance states that AI assistance does not bar a patent application from mentioning or relying on AI, but an application must identify the human inventor or inventors and explain how AI was used.

**Also worth reading:** [How Should Companies Review AI Patent Inventorship Before Filing?](https://iprs.cloud/knowledge/how_should_companies_review_ai_patent_inventorship_before_filing.php) · [What Is the Best PCT National Phase Strategy for Patent Applications in 2026?](https://iprs.cloud/knowledge/what_is_the_best_pct_national_phase_strategy_for_patent_applications_in_2026.php) · [How Does AI Patent Inventorship Compliance Software Help Teams in 2026?](https://iprs.cloud/knowledge/how_does_ai_patent_inventorship_compliance_software_help_teams_in_2026.php)

The practical burden is claim-specific rather than document-specific. A generic statement that a scientist “used machine learning” is usually weaker than a record showing that the scientist selected the technical objective, supplied essential experimental parameters, recognized the problem, and conceived the operative solution. Human authorship of code or a prompt is evidence, but neither automatically proves inventorship because the legally relevant act is conception of the claimed invention. Organizations should preserve evidence before disclosure, filing, or litigation, while recognizing that a tidy paper trail cannot cure a naming error or invent facts that never existed.

## Why Current U.S. Patent Law Requires Natural-Person Inventorship

The controlling authority is the Supreme Court's decision in Thaler v. Vidal, No. 21-2347, decided by the Federal Circuit in 2022. The court held that the patent statutes identify an inventor as a natural person and do not permit an “inventor” to be an AI system merely because the system produces a claimed invention. The decision did not decide that every AI-assisted invention is unpatentable. Its holding concerns who can be named as the inventor, so a human may still qualify when that person performed the required acts of conception and is properly identified under 35 U.S.C. § 115.

Section 115 requires applicants to make a statement regarding inventorship that is made in good faith and to correct the naming of an inventor without deceptive intent. The USPTO has separately stated that improper inventorship may be addressed through correction, and extreme or intentional misconduct can produce more serious consequences. Inventorship is not judged by title, seniority, coding volume, or project ownership. A supervisor who did not contribute to conception is not an inventor merely because the supervisor managed the team, while a junior engineer can be an inventor even if the company owns the resulting patent rights.

The human-contribution standard is especially difficult where generative AI appears to have produced an output in seconds. Speed does not answer the legal question. A person who conceives a result merely by asking an AI system to produce it raises difficult questions under Thaler, and the person must still be identified and supported as the human inventor. By contrast, a person who frames a specific inventive concept, provides essential inputs, tests alternatives, and recognizes a solution is in a stronger position. Advice of counsel should focus on reconstructing the actual inventive history rather than attaching conclusory labels to a workflow.

## What Kinds of AI Patent Inventorship Evidence Are Strongest?

The best evidence identifies a natural person, a date, and a contribution to conception. Dated laboratory notebooks and electronic laboratory notebooks can show the sequence from problem formulation to successful conception. Source-control records may establish authorship and timing, but a commit message rarely explains the technical contribution by itself. Useful records pair commits with pull requests, issue discussions, design reviews, or declarations that map a change to particular claim limitations. Hash-timestamped records can improve integrity, but they do not replace substantive explanation.

Prompt histories may be relevant when they show the human's selection of objectives, constraints, data, or evaluation criteria. They are less persuasive when a generic prompt merely asks a model for an invention. Training-data selections, parameter adjustments, feature engineering, and failure analysis can be important in machine-learning inventions, but engineers should avoid treating every technical action as inventorship. Inventorship turns on contribution to conception, not every activity surrounding development. An employee who only curated data, deployed code, or managed budgets may have valuable employment evidence without being an inventor.

Witness testimony can explain omitted details, but it is strongest when it is consistent with contemporaneous records. A declaration written years later should address specific records rather than use retrospective memory to reverse-engineer a preferred answer. Invention disclosures, patent committee memoranda, prosecution instructions, and inventor declarations can organize the record, although later documents generally carry less weight than records created during development. Companies should maintain a chain from conception through enablement, filing, and assignment, while keeping confidential AI vendor materials under appropriate confidentiality and license terms.

| Evidence feature | Stronger record | Weaker record |
| --- | --- | --- |
| Human contribution | Specific conception of claim elements | A statement that AI “helped” |
| Timing | Contemporaneous, dated records | Undated or reconstructed notes |
| Technical link | Maps each person to a claim limitation | Lists everyone on the project |
| AI use | Shows task, inputs, human judgment, and validation | Records only a final prompt or output |
| Corroboration | Notebook, code, experiments, and witness agree | One retrospective declaration |

## How to Reconstruct and Document an AI-Assisted Invention
The first step is to identify the proposed claims, because the evidence must be assessed against the invention as claimed rather than an abstract marketing description. Create a claim-contribution record for each person and explain whether that person conceived a limitation, solved a stated problem, or merely performed routine implementation. This is not a search for the largest possible inventor list. Overdesignating can itself create a § 115 problem, and omitting a true inventor can leave validity, ownership, or assignment disputes unresolved.

The second step is to preserve the earliest human contribution: problem statements, sketches, constraints, prompt iterations, model-selection decisions, experimental results, and failed approaches. Preserve metadata where available, including UTC timestamps, commit identifiers, file hashes, and version histories. Retain source code and data under a defensible retention schedule, but do not destroy or alter records after a dispute is foreseeable. Legal holds may be needed when a notice, claim, investigation, or threatened enforcement action creates preservation duties.

The third step is to use the USPTO's AI-focused examination resources and an attorney or patent professional to evaluate the application. The USPTO's 2023 guidance recognizes that patent applications may discuss AI-related subject matter and asks for information sufficient to support the role of the human inventor. Examination may also consider whether a claim is patentable under § 101, whether it has adequate written description and enablement, and whether the naming of inventors is correct. Passing an inventorship review is separate from establishing patent eligibility; an application can name a proper inventor and still face § 101, disclosure, or prior-art objections.

## Alternatives to Calling the AI System the Inventor

Organizations generally have three defensible routes: name the human contributors, correct the record if a filing named the wrong person, or abandon an application that cannot ethically and factually be supported. Some companies also frame the output as an improvement to an existing human-conceived invention, which may produce a supportable case for a natural-person inventor. That framing must match reality and cannot be used to disguise the absence of a human contribution.

A human-only naming approach is appropriate where a person conceived the claimed subject matter and AI performed search, drafting, optimization, or routine analysis. A correction approach is appropriate when the original application omitted a person or misidentified the inventor without deceptive intent. A trade-secret or controlled-publication strategy may be preferable where public filing would cause competitive harm, but disclosure deadlines and contractual duties still require review. A foreign-filing strategy cannot avoid U.S. inventorship rules when the human contribution or naming issue is material, and separate inventorship laws may apply abroad.

| Situation | Proper response | Main risk |
| --- | --- | --- |
| AI assisted drafting or search | Name the human who conceived the claim | Calling the model an inventor |
| Human contribution is uncertain | Conduct a documented claim-by-claim review | Overclaiming an inventor's role |
| Wrong inventor was named | Consider correction under § 115 | Ignoring a known error |
| Human conception cannot be shown | Seek counsel before filing; reassess strategy | Filing an unsupported declaration |
| Confidential competitive disclosure matters | Consider delayed or alternative protection | Missing a filing or disclosure deadline |

## Common Mistakes in AI-Related Patent Records
A frequent mistake is treating model sophistication as proof of human inventorship. The fact that a team trained a large model, selected a dataset, or ran a million experiments may be commercially impressive, but it does not automatically show conception of every claimed feature. Another mistake is treating the first person who typed a prompt as the sole inventor, even when several colleagues contributed indispensable technical ideas. Conversely, treating every employee as an inventor dilutes credibility and can create avoidable correction proceedings.

Companies also err by preserving only screenshots. Screenshots may omit metadata, version history, collaborators, and the context needed to establish conception. Deleting experimental branches or rewriting notebooks to make the story cleaner is worse than an incomplete record. Prompt logs should be retained in a form that permits verification, subject to security restrictions and third-party terms. A vendor's refusal to provide records is a diligence problem, not a reason to invent an explanation.

The legal mistake is confusing inventorship with ownership. The human inventor may assign rights to a company, university, or customer, but assignment does not make the assignee the inventor. Similarly, a patent application's references to AI are not automatically admissions that the claims are unpatentable. The USPTO has expressly distinguished AI use from the question of who performed the inventive act. A disciplined record should state what AI did, what the human did, and which legal requirement each fact addresses.

## Timing, Cost, and When to Act

Evidence should be assembled during development, not after a notice or lawsuit. A reasonable operating window is to capture records at each major inventive milestone, such as problem definition, first conception, enabling experiment, and filing approval. For a high-volume AI product, an internal evidence template and automated repository retention can reduce the time needed for later review. High-value, novel, or internationally protected technology merits a more detailed claim mapping process because the same invention may require different inventorship analyses in different jurisdictions.

Private legal review commonly costs from roughly $2,500 to $15,000 for a focused inventorship and filing-readiness review, while a more extensive contested inventorship analysis, expert reconstruction, or cross-border review can reach $25,000 or more. Official USPTO fees are separate from legal fees, and AI analytics or evidence-management software may add subscription or project costs. No universally valid price exists because a straightforward assisted invention differs from a multi-inventor dispute involving lost records and expert testimony. Organizations should obtain a scoped estimate covering the claims, people, documents, jurisdictions, and anticipated dispute risk.

Act immediately when a patent application is due within 60 to 90 days, a competitor alleges inventorship misconduct, an invention is about to be publicly disclosed, a licensing transaction is being negotiated, or a former employee contests ownership. The strongest reason to act early is not administrative convenience: a missing witness, expired notebook access, deleted repository, or departing engineer can make later proof harder. Conversely, urgency does not justify naming a person without a factual basis. A short pause to document the real contribution is usually better than a filing built on a convenient but inaccurate story.

## A Defensible Record for Counsel and Product Teams

For iprs.cloud users, the relevant issue is not whether AI appears in the patent narrative; it is whether the registry and prosecution record can reliably support the human inventive history. A product team should preserve technical events, counsel should map them to claims, and business stakeholders should track assignments and filing deadlines. That division prevents software logs, legal conclusions, and commercial promises from being confused with one another. It also supports later patentability checks, freedom-to-operate work, invalidity analysis, and evidence-of-use studies without pretending that analytics alone resolves inventorship.

A defensible file typically answers four questions. Who conceived the claimed subject matter? What concrete act demonstrates that contribution? When did it occur? Which records independently support the answer? AI prompts, model versions, training runs, and code commits may support those answers, but only when connected to a human's conception. A registry should therefore treat provenance, timestamps, permissions, and audit trails as evidence features rather than cosmetic metadata.

The conclusion is conditional but clear. If AI performed routine assistance and a natural person conceived the claimed invention, the application can ordinarily proceed with careful naming and disclosure. If the AI is proposed as the inventor, current U.S. law does not support that position. If the human contribution cannot be reconstructed, the correct response is to seek advice, correct what can lawfully be corrected, and consider non-patent or confidential protection. Evidence cannot manufacture inventorship, but well-preserved evidence can make the actual inventorship far more defensible.

## Quick answers

### Can an AI system be listed as a patent inventor?

Under current U.S. law, an inventor must be a natural person. The Federal Circuit's 2022 decision in Thaler v. Vidal rejected the attempt to name an AI system as the inventor, although the decision did not prohibit every form of AI-assisted patenting.

### Do ChatGPT prompts prove patent inventorship?

A prompt can be evidence, but it proves inventorship only when connected to a natural person's conception of the claimed subject matter. A generic prompt asking for an invention is weaker than records showing the human defined the problem, supplied essential technical constraints, selected a solution, and validated it.

### Who is an inventor on an AI-assisted patent?

The inventor is a natural person who contributed to conception of the claimed invention, not necessarily the project owner, employer, supervisor, or every person who performed engineering work. Multiple people may be inventors if each made an inventor-like contribution to at least part of the claimed subject matter.

### What happens if the wrong inventor is named?

A naming error may be addressed through correction procedures, particularly where the correction is made without deceptive intent. Intentional misconduct or an uncorrected material error can create validity, ownership, and regulatory consequences, so counsel should review the facts before filing.

### Is using AI evidence of patent ineligibility?

No. The USPTO's 2023 guidance distinguishes AI use from patent eligibility and inventorship. An application may discuss AI-assisted work, but the claims must still satisfy applicable requirements, including adequate disclosure and patentable subject matter, and the human inventor must be correctly identified.

Canonical: https://iprs.cloud/knowledge/what_evidence_proves_human_inventorship_in_ai-assisted_patent_applications.php
Markdown: https://iprs.cloud/knowledge/what_evidence_proves_human_inventorship_in_ai-assisted_patent_applications.php/index.md