By Bijimol P, Manager – Legal
Artificial intelligence has not rewritten copyright law. It has forced courts and regulators to apply familiar principles of authorship, ownership, infringement, and fair use to technology that can generate expressive content with limited human involvement.
The rise of generative AI has brought several distinct legal questions. Can an AI-generated work receive copyright protection? Does a prompt make the user an author? And does ownership of an output depend on whether copyrighted material was lawfully used to train the model?
Current U.S. law provides relatively firm answers to some of those questions, while others remain the subject of rapidly developing litigation. The clearest starting point is also the oldest one: copyright requires human authorship.
Copyright Still Starts with a Human Author
Section 102 of the Copyright Act protects “original works of authorship fixed in any tangible medium of expression.” 17 U.S.C. § 102(a). The United States Supreme Court in Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340, 345 (1991), has explained that originality requires independent creation plus at least a minimal degree of creativity.
Generative AI raised a more basic question: must the author be human?
The D.C. Circuit answered “yes” in Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025). In Thaler, Dr. Stephen Thaler sought registration for an image that he represented as autonomously created by his “Creativity Machine,” which he identified as the sole author. The Copyright Office rejected the application, and the court affirmed, holding that the Copyright Act requires eligible works to be authored “in the first instance by a human being.” Id. at 1051.
The decision was important, but narrow. Thaler did not hold that using AI defeated copyright protection. It addressed a work presented as entirely machine-generated and therefore did not decide how much human involvement makes an AI-assisted work copyrightable.
That distinction tracks the U.S. Copyright Office’s approach. Its 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability,[1] concludes that copyright continues to protect human expression in works created with AI assistance. Purely AI-generated material, or material over which the user lacks sufficient creative control, does not receive protection. Whether human contributions amount to authorship requires a case-by-case analysis.
What Humans Can Still Own in an AI-Assisted Work
The better rule, therefore, is not that “AI works cannot be copyrighted.” Copyright may protect the human contribution to an AI-assisted work.
A person may claim copyright in original expression that the person contributes to an AI-assisted work. That may include independently written text, expressive revisions, substantial modifications, or sufficiently creative selection, coordination, and arrangement. Section 103 reflects the same principle for compilations and derivative works: copyright extends to the new material contributed by the author, not automatically to unprotectable underlying material.[2]
Prompts present a harder issue. A sufficiently original prompt may itself contain copyrightable expression, but copyright in the prompt does not automatically extend to the model’s response.[3] The critical question is whether a human made and controlled the creative choices reflected in the final work.
Copyright ownership also differs from contractual rights. Platform terms may allow a user to publish or commercialize an output even when some machine-generated portions lack copyright protection. Licenses, work-made-for-hire rules, and ownership of a digital file therefore address different legal interests.[4]
Ownership And AI Training Are Different Questions
Whether an AI output qualifies for copyright protection does not determine whether a developer infringed copyright while building or training the model. Copyrightability asks whether a new work contains protectable authorship and infringement asks whether someone exercised a copyright owner’s exclusive rights without authorization or a valid defense. Fair use under 17 U.S.C. § 107 provides one such defense and requires a fact-specific analysis of purpose, nature, amount used, and market effect.
Two 2025 decisions illustrate both the importance and the limits of fair use in AI training.
In Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025),[5] the court held that Anthropic’s use of books to train its Claude models qualified as fair use on the record before it. The court viewed the training use as highly transformative and emphasized that the plaintiffs did not allege that Claude produced infringing copies of their books. But it separately held that downloading and retaining pirated copies in a permanent library did not receive the same fair-use protection. That distinction is central. A court may find the use of works for training fair while separately scrutinizing how those works were acquired or retained. The remaining Bartz claims later settled, and the court granted final approval of the class settlement on July 20, 2026.
Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025), reached another fair-use result favorable to an AI developer, but with an important qualification. The court held that Meta’s use of copyrighted books to train Llama was fair use on the evidentiary record before it. At the same time, Judge Vince Chhabria recognized that generative AI could potentially cause legally relevant market harm by producing large volumes of competing works. The plaintiffs failed to establish that theory adequately in Kadrey, but the court’s analysis reinforces that market effects may prove decisive in a different case with a stronger record.
Together, Bartz and Kadrey resist any categorical rule that AI training is always fair use.
Why the New York Times Litigation Still Matters
In The New York Times v. Microsoft/OpenAI,[6] several copyright claims survived dismissal in 2025, including certain claims under § 1202 of the Digital Millennium Copyright Act. A ruling at that stage did not establish liability. It merely allowed adequately pleaded claims to proceed. The procedural posture, however, changed after the Supreme Court’s 2026 decision in Cox Communications, Inc. v. Sony Music Ent., 607 U.S. 583 (2026), which held that providing a generally available service with knowledge that some users employ it to infringe does not, without more, establish contributory copyright liability. In light of Cox, the court on August 6, 2026, dismissed with prejudice certain contributory-infringement claims based on a “material contribution” theory against OpenAI and Microsoft and declined to permit new contributory-liability theories against Microsoft at that late stage of the litigation.
The broader litigation continues to test training practices, alleged copying, model outputs, and fair-use defenses on facts distinct from Bartz and Kadrey.
So Who Owns AI Output Today?
U.S. copyright law still begins with a human author. Thaler confirms that a machine cannot itself hold the status of copyright author. That rule does not exclude AI-assisted works from protection. Human-written text, meaningful revisions, creative modifications, and sufficiently original selection or arrangement may still receive copyright protection. The question is which expressive elements originated with the human creator.
Prompts occupy a narrower role. They may contain protectable expression themselves, but they do not automatically make the user the author of whatever the model produces. Creative control over the final expression remains the key consideration. Contractual rights must also remain separate from copyright. A user may have broad permission to exploit an output under platform terms even when federal copyright does not protect every element.
Training presents another question entirely. Bartz and Kadrey show that courts may find particular training uses fair while still examining acquisition methods, retention practices, and market effects separately. The continuing New York Times litigation reinforces the same point. AI copyright disputes turn heavily on the specific facts and theories presented.
For creators and businesses, the practical lesson is clear. Where copyright exclusivity matters, meaningful and demonstrable human authorship matters as well. For AI developers, lawful sourcing and careful handling of training material can matter independently of what the model ultimately produces.
[1] U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 29, 2025)
[2] 17 U.S.C. § 103(b).
[3] https://www.copyright.gov/ai/ai_policy_guidance.pdf?
[4] 17 U.S.C.S. § 201
[5] https://www.copyright.gov/fair-use/summaries/Bartz-v-Anthropic-PBC-787-F-Supp-3d-1007-ND-Cal-2025.pdf
[6] https://docs.justia.com/cases/federal/district-courts/new-york/nysdce/1%3A2023cv11195/612697/1449
For more practical litigation drafting tips, pleading and legal research insights, follow LawCompany.