Who owns AI-generated output, and what should my terms say?

Copyright in AI output after the Copyright Office's 2025 report and Thaler v. Perlmutter, what model providers' terms do and don't give you, and how to allocate output rights with your users.

Short answer

In the United States, material generated purely by AI isn't protected by copyright, because copyright requires a human author. Content is protected only to the extent a person contributes enough original expression, for example by writing parts of it, creatively modifying it, or creatively selecting and arranging AI material. Prompts alone are usually not enough. So "who owns the output" is mostly a contract question, and contracts have limits: a provider can assign whatever rights it has in an output, but it can't create copyright where the law gives none.

What the law says

• Human authorship is required. In Thaler v. Perlmutter, decided March 18, 2025, the D.C. Circuit held that the Copyright Act requires a human author, so a work created autonomously by an AI system can't be registered. The Supreme Court denied review on March 2, 2026, leaving the decision in place without ruling on the merits. The decision doesn't bar protection for original human contributions to AI-assisted works.
• The Copyright Office's position. Its January 29, 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability, concluded that AI outputs are protected only to the extent they contain sufficient original human expression. Prompts alone generally don't make someone the author. Protection can come from human-authored material in the work, creative human modifications, and original selection, coordination, or arrangement. AI-generated parts don't become protected just because they sit inside a human-authored work.
• Registration. Under the Copyright Office's March 2023 registration guidance, applicants must disclose more-than-minimal AI-generated material and generally exclude it from the authorship they claim. Using AI as an assistive tool doesn't always require a disclaimer.
• Patents. Under Thaler v. Vidal (Fed. Cir. 2022), an AI system can't be named as an inventor. In November 2025, the USPTO rescinded its February 2024 AI-specific inventorship guidance. Its current guidance treats AI as a tool: only people can be inventors, and AI-assisted inventions are judged by the ordinary conception and inventorship standards, with no separate or heightened test.
• Other countries differ. The United Kingdom's Copyright, Designs and Patents Act 1988, section 9(3), treats the person who made the arrangements necessary to create a computer-generated work as its author. The provision remains in force, but how it applies to generative AI, including who made the "arrangements," is unsettled. If you sell abroad, don't assume the US answer applies.

What your model provider's terms do and don't give you

• Ownership, as between you and the provider. Major providers' business and API terms generally give customers the rights in outputs, by assigning whatever rights the provider holds. That is an allocation of rights by contract. It doesn't make an output copyrightable, original, or free of third-party rights.
• No uniqueness. Outputs generally aren't unique. Two customers can receive identical or similar outputs, and output-rights clauses don't stop a provider from generating similar material for someone else.
• Indemnity has conditions. Some providers offer intellectual-property indemnity for qualifying commercial or API use, with exclusions. Common exclusions apply when you knew or should have known the output was likely to infringe; disabled or ignored citations, filters, or safety features; modified the output or combined it with other products in certain ways; lacked rights to your inputs; or faced a trademark claim. Don't tell users or customers that outputs can't infringe, or that you are always covered.
• Restrictions on competing models. Providers commonly restrict using their services or outputs to build or train competing AI systems, but the terms differ by provider. Some restrict how you use the service; others restrict how you use the outputs.

Read the specific contract, service, plan, and version you're on before making any representation about ownership or indemnity.

What your terms should say

• Who owns outputs, as between you and users. A common approach: users own their inputs and the outputs generated for them, to the extent any rights exist, and grant you a license to operate the service.
• No promise of protection or uniqueness. Don't tell users outputs are copyrighted, exclusive, or original.
• Inputs and training. Say whether you use inputs or outputs to train models. See Can I train or fine-tune a model on my users' data?
• Responsibility. Users must review outputs before relying on or publishing them. Disclaim warranties of accuracy and non-infringement, or limit them to what your provider actually gives you.
• Use restrictions. Pass through your provider's restrictions, and prohibit unlawful uses of outputs.
• Indemnity. Decide deliberately whether you offer any, and keep it within what your provider covers, including its exclusions.

For more on AI terms generally, see What should the terms of service for an AI product cover?

Checklist an agent can run against the codebase

• List every feature that generates text, images, audio, video, or code for users.
• Find the output-ownership clause in the terms of service, or note that there isn't one.
• Identify the model provider, service, and plan behind each feature, and link its output, indemnity, and competing-use terms.
• Check whether any setting disables provider citations, filters, or safety features that an indemnity depends on.
• Search the app's copy and marketing for claims that outputs are "yours," "unique," "original," or "copyrighted."
• Check whether code-generation features filter or flag output that matches public code, and whether license information is shown.
• Check whether user inputs or outputs are stored for training, and compare with the terms.

Sources

• Supreme Court docket No. 25-449, Thaler v. Perlmutter (certiorari denied March 2, 2026)
• U.S. Copyright Office, Copyright and Artificial Intelligence reports
• U.S. Copyright Office, Registration Guidance: Works Containing Material Generated by Artificial Intelligence (March 2023)
• USPTO, Artificial intelligence resources and current inventorship guidance
• UK Copyright, Designs and Patents Act 1988, section 9
• OpenAI Services Agreement
• Anthropic legal terms
• Google Cloud Service Specific Terms

If this applies to you

An app launch legal package includes terms of service with output, input, and training clauses that match your provider's terms. For a single new generation feature, a pre-build feature check covers that change.