Guide · AI and intellectual property
AI and IP Law: How U.S. Patent, Copyright and Trademark Law Treat AI
Where U.S. law stands on AI inventors, AI authors, training data, patent eligibility, voices and likenesses, and trade secrets. Each development is dated and sourced, and the table below separates what is settled from what is still open.
Last reviewed October 21, 2026. For education and general information; not legal advice.
On this page
Can an AI be an inventor?
No. Under U.S. patent law only natural persons can be inventors. An invention made with the help of AI can still be patented, but only if a human made the contribution that counts as inventing: conception of the claimed invention. Since November 2025 the USPTO applies the same inventorship test whether or not AI was used.
In practice
- Name the people who conceived what is claimed. Do not name an AI system, and do not leave a human off because AI helped.
- Keep records of who framed the problem, chose among outputs and arrived at the claimed solution. Inventorship questions are factual.
Key developments
- Aug. 5, 2022
Thaler v. Vidal (Fed. Cir.)
The Federal Circuit held that the Patent Act's 'inventor' means a natural person, affirming the USPTO's refusal of applications that named the AI system DABUS as the sole inventor. The Supreme Court denied review on April 24, 2023.[1]
- Feb. 13, 2024
First USPTO guidance on AI-assisted inventions
The USPTO said AI-assisted inventions are not categorically unpatentable and asked whether each named inventor made a significant contribution, applying the Pannu joint-inventorship factors.[2]
- Nov. 28, 2025
Revised guidance rescinds the 2024 approach
The USPTO rescinded the February 2024 guidance in its entirety and withdrew the Pannu-factor approach for AI-assisted inventions. AI systems are treated as tools. The same legal standard for inventorship applies to every invention, and the Pannu factors apply only when deciding whether several natural persons are joint inventors.[3]
Is training AI on copyrighted works fair use?
It depends on the use, and the law is still developing. One federal appeals court has held that copying a competitor's content to build a competing, non-generative tool was not fair use. Two district courts in California found training generative models on books to be fair use on their records, while treating pirated copies and market harm as separate problems. No court of appeals has yet ruled on training a generative model.
In practice
- Know where training data came from and on what terms. Pirated sources were treated separately from the training itself.
- The closer a product comes to substituting for the works it learned from, the weaker a fair-use defense becomes.
- If the content is available for license, that licensing market can count against fair use.
Where we teach it
- Ch. 3 The Four Factors Meet the Machine
- Ch. 4 Two Judges, One District
- Ch. 5 When the Machine Speaks
- Ch. 6 Where Liability Actually Lands
Key developments
- Feb. 11, 2025
Thomson Reuters v. ROSS (D. Del.)
Judge Bibas, sitting by designation, held that Westlaw headnotes were original enough to be protected and that ROSS's use of them to train a competing legal-research tool was not fair use. ROSS's tool returned existing judicial text rather than generating new text. He certified the questions for interlocutory appeal.[9]
- May 9, 2025
Copyright Office report, Part 3: Generative AI Training (pre-publication)
The Office said building training datasets implicates the reproduction right, that fair use must be decided case by case, and that market harm is often the decisive factor. Its reports state the Office's views; they do not bind courts.[5]
- June 23, 2025
Bartz v. Anthropic (N.D. Cal.)
Judge Alsup held that training large language models on books was fair use, as was scanning print books Anthropic had bought. Downloading and keeping millions of books from pirate libraries was not excused, and that issue was set for trial.[12]
- June 25, 2025
Kadrey v. Meta (N.D. Cal.)
Judge Chhabria granted Meta summary judgment on fair use for training on the 13 plaintiff authors' books, because they had not developed evidence of market harm. He stressed the ruling was limited to that record and that market dilution by AI-generated works could defeat fair use in another case. Claims about Meta's torrenting of books continue.[14][15]
- July 20, 2026
Bartz settlement approved
The court gave final approval to a $1.5 billion class settlement covering works Anthropic downloaded from pirate libraries. The release does not cover claims about AI outputs or conduct on or after August 25, 2025, and Anthropic did not admit liability.[13]
- Sept. 29, 2026
Thomson Reuters v. ROSS (3d Cir.)
The Third Circuit affirmed: the headnotes are copyrightable and ROSS's use was not fair use. It is the first federal appellate ruling on fair use in AI training. The court treated the use as commercial and minimally transformative, recognized harm to a market for licensing content as training data, and set generative AI aside.[10][11]
Are AI inventions patent-eligible?
They can be. AI inventions are examined under the same 35 U.S.C. 101 framework as other software, the Supreme Court's Alice/Mayo test as set out in MPEP 2106. A claim that recites an abstract idea, such as a mathematical calculation, is still eligible if it integrates the idea into a practical application, for example a specific improvement to how a computer or model works. USPTO guidance since 2024 has moved in applicants' favor, but none of it changes the statute or the courts' test.
In practice
- Describe the technical improvement in the specification and recite the features that produce it in the claims.
- Where a rejection turns on facts, evidence such as benchmarks in a Rule 132 declaration can answer it. See our explainer on subject matter eligibility declarations.
Key developments
- July 17, 2024
USPTO AI subject matter eligibility update
The USPTO restated the Alice/Mayo framework as applied to AI, discussed recent Federal Circuit decisions, and added new AI examples. It said whether an invention was made with AI assistance is not part of the eligibility analysis.[16]
- Aug. 4, 2025
Examiner reminders for software and AI
A memo to the software and business-method technology centers told examiners not to stretch the 'mental process' grouping to limitations that cannot practically be performed in the human mind, to evaluate claims as a whole, and to reject only when ineligibility is more likely than not.[17]
- Nov. 4, 2025
Ex parte Desjardins made precedential
The Appeals Review Panel vacated a 101 rejection of a machine-learning training claim, faulting the analysis for reading the claim at too high a level of generality. The decision was designated precedential and, on December 5, 2025, added to the MPEP's eligibility guidance by advance notice.[18][19]
- Dec. 2025 – Sept. 2026
Subject matter eligibility declarations
The USPTO issued best practices for Rule 132 declarations that put evidence of a technological improvement behind an eligibility argument (December 4, 2025, superseded April 30, 2026), then a flowchart for how examiners evaluate them (September 29, 2026).[20][21]
Trademarks, voices and likenesses
Federal trademark law already reaches AI-generated content that falsely suggests a brand or a person endorses something (15 U.S.C. 1125(a)). There is no federal right of publicity. Protection for a person's name, image and voice comes from state law, and some states have updated it for AI. A federal digital-replica bill has been proposed but not enacted.
In practice
- Using a real person's voice or likeness in AI output needs their consent; which state's law applies can matter.
- Brand owners can use false endorsement and infringement claims against AI content that misuses their marks.
Where we teach it
- Ch. 10 Brands in the Machine
- Ch. 12 Your Face, Your Voice
- Ch. 13 The Federal Vacuum and the State Patchwork
Key developments
- July 1, 2024
Tennessee ELVIS Act takes effect
Signed March 21, 2024, the law added voice, including a simulated voice, to Tennessee's publicity-rights statute and reaches those who knowingly make an unauthorized replica available.[22]
- July 31, 2024
Copyright Office report, Part 1: Digital Replicas
The Office found existing law does not adequately cover realistic digital replicas of voices and likenesses and recommended a new federal law.[6]
- June 2026
NO FAKES Act advances, not enacted
Reintroduced in May 2026, the federal digital-replica bill was approved by the Senate Judiciary Committee in June 2026. It had not been enacted as of this page's last review.[23]
Federal false endorsement: [25]
Trade secrets and AI
Trade secret law protects information only while its owner takes reasonable measures to keep it secret (18 U.S.C. 1839(3)). AI raises two practical questions rather than new law: whether putting confidential information into a third-party AI tool is consistent with reasonable secrecy measures, and how to protect model weights, training data and prompts that are kept confidential.
In practice
- Set rules for what confidential information may go into which AI tools, and check the tools' data terms.
- Treat model weights, curated datasets and system prompts as potential trade secrets: limit access and use NDAs.
No AI-specific statute or appellate ruling has changed trade secret law; the general rules apply. [24]
What's settled and what's still open
As of October 21, 2026. "Settled" means a binding appellate ruling or statute and no pending review; "Open" means courts or agencies have not given a final answer.
Learn it properly
IP Law in the Age of AI
An 18-chapter self-paced course on everything above, from authorship and fair use to publicity rights and the practitioner's own duties. $297 self-study or $397 with the certificate track; also included in Wysebridge ONE.
Try a lesson first, free.
Thomson Reuters v. ROSS and fair use in AI training: the four factors, the training-data market and three practice hypotheticals.
Frequently asked questions
Keep reading
Sources
- Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022); cert. denied Apr. 24, 2023 (IPWatchdog report)
- USPTO, Inventorship Guidance for AI-Assisted Inventions, 89 FR 10043 (Feb. 13, 2024)
- USPTO, Revised Inventorship Guidance for AI-Assisted Inventions, 90 FR 54636 (Nov. 28, 2025)
- Copyright Office, Registration Guidance for Works Containing AI-Generated Material, 88 FR 16190 (Mar. 16, 2023)
- U.S. Copyright Office, Copyright and Artificial Intelligence (Parts 1–3)
- Copyright Office, Part 1: Digital Replicas (July 31, 2024)
- Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025) (Goodwin summary)
- Thaler v. Perlmutter, No. 25-449, cert. denied Mar. 2, 2026 (SCOTUSblog case file)
- Thomson Reuters v. ROSS Intelligence (D. Del. 2025): case page and rulings (BakerHostetler)
- LawNext: Third Circuit rules for Thomson Reuters against ROSS (Sept. 2026)
- IPWatchdog: Third Circuit affirms fair use ruling against ROSS (Sept. 30, 2026)
- Bartz v. Anthropic, order on fair use (N.D. Cal. June 23, 2025) (Akin summary)
- Authors Alliance: Bartz v. Anthropic settlement receives final approval (July 21, 2026)
- Authors Alliance: Kadrey v. Meta fair use ruling (June 26, 2025)
- Kadrey v. Meta: remaining torrenting claims schedule (Mar. 2026)
- USPTO, 2024 Guidance Update on Patent Subject Matter Eligibility, Including on AI, 89 FR 58128 (July 17, 2024)
- USPTO examiner memo on eligibility for software and AI (Aug. 4, 2025) (BakerHostetler summary)
- USPTO memo on Ex parte Desjardins (precedential Nov. 4, 2025)
- USPTO updates subject matter eligibility guidance in the MPEP (Dec. 5, 2025)
- USPTO memo, Subject Matter Eligibility Declarations: flowchart and best practices (Sept. 29, 2026)
- USPTO subject matter eligibility page
- Manatt: Tennessee's ELVIS Act expands publicity rights for voices (2024)
- Sen. Blackburn: NO FAKES Act passes Senate Judiciary Committee (June 2026)
- 18 U.S.C. 1839 (trade secret definitions)
- 15 U.S.C. 1125 (false designation of origin, false endorsement, dilution)
Last reviewed October 21, 2026. Court rulings and agency guidance change; check the source before relying on any entry.
Advise on AI with the law in hand.
Eighteen chapters on authorship, inventorship, training data, liability and publicity rights.