Artificial Intelligence and Copyright

By Hank FasthoffJuly 4, 2026

Artificial intelligence affects copyright at two different stages. The first concerns ownership of material produced with an AI system. The second concerns potential infringement when copyrighted works are copied for training or when a model produces material derived from protected expression.

Current law supplies firmer answers on authorship than on training. A nonhuman system isn’t an author under the Copyright Act, but human contributions that satisfy copyright’s requirements receive protection. Courts haven’t established a general rule that treats all generative AI training as fair use or infringement.

Copyright Requires Human Authorship

Section 102 protects original works of authorship fixed in a tangible medium of expression. Copyright begins with the author, and the Copyright Act treats authors as human beings.

In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), Stephen Thaler sought registration for an image identified as having been generated autonomously by his Creativity Machine. His application named the machine as the sole author and Thaler as the copyright claimant.

The D.C. Circuit affirmed the Copyright Office's refusal to register the image. It held that eligible work must be authored initially by a human being because the Copyright Act's ownership, duration, inheritance, transfer, and joint authorship provisions presuppose human authorship.

The court didn't decide whether Thaler could claim authorship based on his development or use of the system because he hadn't preserved that argument before the agency. It also declined to decide whether the Constitution independently requires human authorship. The Supreme Court denied review on March 2, 2026.

Copyright in Human Contributions

Use of an AI system doesn't disqualify an otherwise copyrightable work. The Copyright Office's January 2025 copyrightability report explains that protection depends on whether a human determined sufficient expressive elements of the resulting work.

Copyright protects original human authored expression fixed in a tangible medium, including qualifying text, music, images, and code incorporated into a work made with AI. A human’s selection, coordination, arrangement, or modification of material receives protection when it satisfies the applicable originality requirement, even if individual elements are unprotected.

Prompts alone ordinarily don't establish authorship under the Office's current approach. The report concludes that prompts generally convey instructions and unprotectable ideas without providing enough control over the expressive details produced by current generative systems.

A different result may apply when a human authored work remains perceptible in the output or when the user makes copyrightable modifications after generation. The analysis concerns the expression contributed by the human, not the time spent operating the system or the number of instructions submitted.

The Copyright Office's Registration Decisions

Applying these principles to Zarya of the Dawn, the Copyright Office reviewed a graphic novel that combined human written text with images generated through Midjourney. Its registration decision recognized copyright in the text and in the selection, coordination, and arrangement of the text and images.

Registration didn't extend to the individual Midjourney images. The Office concluded that the user's prompts didn't provide sufficient control over the expressive elements produced by the system.

The boundary is being challenged in Allen v. Perlmutter, No. 1:24-cv-02665-WJM (D. Colo.). Jason Allen argues that more than 600 prompt iterations and subsequent modifications support authorship of Théâtre D'opéra Spatial. The Copyright Office maintains that Allen didn't control the expressive elements generated by Midjourney and has filed a cross motion for summary judgment. As of August 31, 2026, the motions remain pending.

Ownership Depends on the Human Contribution

A company doesn't obtain copyright in machine generated expression merely because an employee used the company's account, the company paid for the system, or a provider's terms purport to assign the output. A contract may allocate rights between the parties, but it can't make uncopyrightable material copyrightable.

Human contributions require a separate ownership analysis. Under Sections 101 and 201, copyrightable work prepared by an employee within the scope of employment ordinarily belongs to the employer as a work made for hire.

Independent contractors follow narrower rules. A commissioned work qualifies as a work made for hire only if it falls within one of the statutory categories and the parties sign the required written agreement. Otherwise, the human author initially owns the contribution. A voluntary transfer of copyright ownership requires a writing signed by the owner or the owner’s authorized agent under Section 204.

Businesses using AI systems should therefore separate three questions. The first is which expressive elements came from a human. The second is who owns those human contributions under employment, work made for hire, and assignment rules. The third is what contractual rights the AI provider and user have granted each other.

Registration of Works Made with AI

The Copyright Office’s March 2023 registration guidance requires applicants to identify the human authored material they seek to register. Applicants must disclose and exclude from the copyright claim AI generated material that is more than de minimis, meaning more than minimal.

An application must identify the human author and describe the copyrightable contribution, such as text, editing, visual modifications, or selection and arrangement. An AI system must not be named as an author.

An applicant who omitted material information from an earlier application may use supplementary registration when appropriate. The Office evaluates the human and machine generated components based on the specific work rather than applying a separate registration rule to every use of AI.

AI Training Involves Several Uses

Developing a generative model may involve collecting works, converting them into machine readable formats, storing copies, processing them during training, and retaining datasets for later use. Each act requires its own copyright analysis.

The Copyright Office's May 2025 report on generative AI training concludes that several stages of model development implicate copyright owners' exclusive rights. Whether a particular act is excused as fair use depends on the works, source, purpose, output controls, and market effects.

The report doesn't treat all training alike. It states that noncommercial research or analysis without reproduction in outputs is more likely to qualify as fair use. Commercial copying of expressive works obtained from pirate sources to produce competing expressive content weighs against fair use, particularly when licenses are reasonably available.

The Office recommends allowing voluntary licensing markets to develop before Congress imposes a new statutory licensing system. Its report remains a policy analysis rather than a judicial decision.

Bartz v. Anthropic

In Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), the district court separated Anthropic's training use from its acquisition and retention of books.

The court held that using books to train Anthropic's language models was fair use on the record before it. It also treated conversion of purchased print books into replacement digital copies for an internal library as fair use.

A different ruling governed millions of books downloaded from pirate libraries and retained in a general purpose collection. The court held that building and keeping that library wasn't fair use merely because Anthropic later used some copies for training.

The parties settled the piracy based claims for $1.5 billion without a determination of liability at trial. Final approval and judgment followed on July 20, 2026. The settlement resolved specified past acquisition and copying claims without establishing an appellate rule governing AI training.

Kadrey v. Meta Platforms

In Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025), Meta obtained summary judgment on the named plaintiffs' training claim. The court found Meta's use highly transformative and concluded that the plaintiffs hadn't presented sufficient evidence of market harm.

That ruling didn't establish that generative AI training is categorically fair. In a March 2026 order, the court emphasized that Meta won because the named plaintiffs hadn't developed meaningful evidence on the effect of training on the market for their works.

Claims concerning Meta's alleged distribution of books through BitTorrent and contributory infringement remain separate. The litigation was active as of August 2026.

Thomson Reuters v. Ross Intelligence

Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-613-SB (D. Del. Feb. 11, 2025), concerns a non generative legal research system trained with material derived from Westlaw headnotes.

The district court rejected Ross Intelligence's fair use defense. It found that Ross used the material to build a competing legal research product and that the use affected an existing market for the headnotes.

The Third Circuit accepted an interlocutory appeal and heard argument on June 11, 2026. As of August 31, 2026, it hasn't issued a decision. No federal appellate court has decided whether generative AI training qualifies as fair use.

The OpenAI Copyright Litigation

Claims against OpenAI and Microsoft involving books, news articles, and other works are being coordinated in In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143 (S.D.N.Y.). The consolidated proceedings include claims brought by authors, publishers, and news organizations.

In an April 2025 opinion, the court allowed core direct and contributory infringement claims brought by news organizations to proceed while dismissing or narrowing other claims. That ruling addressed the sufficiency of the pleadings, not whether training was fair use or whether the plaintiffs would prevail.

The multidistrict litigation remains active. Discovery orders issued in 2026 address training datasets, model records, and output logs, but the court hasn't entered a merits judgment establishing a general training rule.

Training and Outputs Present Different Questions

Input infringement concerns acts such as obtaining, copying, storing, and processing protected works during model development. Those acts may be authorized, infringing, or excused as fair use depending on the particular record.

Output infringement concerns what the system produces. An output infringes when it copies protected expression without authorization or an applicable defense and satisfies the governing infringement standards. Similar subject matter, ideas, facts, methods, or artistic style alone don’t establish infringement.

A favorable ruling about training doesn't immunize infringing outputs. Likewise, an output that doesn't infringe doesn't establish that every copy made during acquisition, dataset construction, or training was lawful.

A business that selects and publishes an infringing output may face direct liability even if it didn't know the model had reproduced protected material. Knowledge may affect remedies and other claims, but lack of knowledge doesn't supply a general defense to direct infringement.

Managing Ownership and Infringement Exposure

Businesses using generative AI should document the human contribution to important works. Drafts, edits, source files, selection decisions, and revision histories help identify which expressive elements came from employees or contractors.

Employment and contractor agreements should address ownership of human authored contributions, use of AI systems, disclosure obligations, confidential information, and responsibility for reviewing outputs. An assignment should cover the copyrightable contribution without assuming that a contract supplies protection to machine generated material.

Vendor terms require separate review. Relevant provisions include rights in inputs and outputs, use of customer material for training, retention, confidentiality, security, infringement warranties, indemnity, defense control, liability limits, and procedures for responding to third party claims.

Businesses should conduct a source and similarity review of higher risk outputs before publication. The review should account for the type of material, the specificity of the prompt, recognizable characters or passages, requested imitation of an existing work, and the commercial use planned for the output.

Copyright owners should preserve registrations and ownership records for valuable works. For a United States work, registration is generally required before filing an infringement action, and the timing rules in Section 412 affect eligibility for statutory damages and attorney's fees.

Artificial intelligence doesn't replace the ordinary copyright analysis. Ownership begins with identifiable human authorship, and infringement depends on the particular copying, protected expression, authorization, defenses, and market evidence involved.