AI and Copyright in the US: Who Owns AI-Generated Work, and What Is Risky

Generative AI has collided with copyright law from two directions at once. On the input side, AI models are trained on enormous amounts of text, images, music and code, much of it copyrighted. On the output side, people are producing articles, illustrations and software with AI help and want to know whether they own the results. US law has answered some of these questions clearly and left others to courts that are still deciding. Here is what is settled, what is not and what to do in the meantime. This is general information, not legal advice.

The foundation: copyright protects human authorship

US copyright protects original works of authorship fixed in a tangible form. It arises automatically when a work is created; registration with the US Copyright Office is not required for protection but is required before suing for infringement of a US work, and it brings other advantages.

The key principle for AI is that copyright requires a human author. The Copyright Office has long refused to register works produced by machines or nature without human creative input, and federal courts have agreed. In Thaler v. Perlmutter, the courts upheld the Office's refusal to register an image that its creator said was generated autonomously by an AI system with no human author, with the D.C. Circuit affirming in 2025.

What the Copyright Office says about AI-assisted work

The Copyright Office has published registration guidance and a multi-part report on copyright and AI, collected at copyright.gov/ai. Its main positions on outputs:

• Prompts alone are generally not enough. Typing a description, even a detailed one, does not give the user sufficient control over the expressive elements the AI produces, so the output is not protected as the prompter's work.
• Human contributions are protected. Text you wrote, edits you made, and your creative selection and arrangement of material can be protected, even when combined with AI-generated elements.
• Expressive inputs count. If you feed in your own drawing or text and it remains perceptible in the output, your original contribution is still yours.
• Disclosure is required. When registering, applicants should identify AI-generated material that is more than minimal and claim only the human-authored parts.
• AI as a tool is fine. Using AI to assist, such as brainstorming, outlining or editing, does not strip protection from human expression.

An early, widely cited example was the graphic novel Zarya of the Dawn. The Office allowed registration for the author's text and her selection and arrangement of the images, but not for the individual AI-generated images themselves.

Practical ownership scenarios

The business consequence: if an asset is purely AI-generated, competitors may be able to copy it, and you may have limited ability to stop them. For logos, characters and flagship creative work, meaningful human authorship matters.

The training question: is it fair use?

AI developers generally argue that copying works to train a model is fair use, a defense that weighs four factors: the purpose and character of the use (including whether it is transformative and whether it is commercial), the nature of the work, how much was used, and the effect on the market for the original.

Authors, artists, news organizations, music publishers and photo agencies have filed many lawsuits in US federal courts challenging that view. Early rulings have not produced a single, clear answer; different judges have weighed the factors differently depending on the facts, including how the training material was obtained and whether outputs compete with the originals. Some disputes have ended in licensing deals or settlements. The Copyright Office has also examined training in its report, concluding that some training uses may be fair and others may not, depending on the circumstances.

Expect this to remain unsettled until appeals courts weigh in, and possibly until Congress acts.

Can AI outputs infringe?

Yes. If an output is substantially similar to a copyrighted work the model was trained on, publishing or selling it could infringe, regardless of whether you intended it. The risk is higher when you prompt for a specific artist's style plus a specific character, ask for song lyrics, or request text "in the exact words of" a book.

Trademarks and publicity rights are separate concerns. AI images that include a company's logo or a recognizable person's likeness can create problems under trademark law and state right-of-publicity laws, even where copyright is not at issue.

Practical steps for businesses and creators

1. Read the terms. Check who owns outputs, whether your inputs are used for training and whether the provider offers indemnification for business customers.
2. Keep humans in the creative loop for anything you need to own: logos, brand characters, books, signature designs.
3. Document your process. Save drafts, edits and notes showing human contributions in case ownership is ever questioned.
4. Avoid prompting for specific protected works, living artists' names, characters or lyrics in commercial projects.
5. Run a similarity check on important outputs, such as a reverse image search for visuals.
6. Disclose AI material when registering with the Copyright Office.
7. Address AI in contracts. Agreements with freelancers and agencies should say whether AI may be used, require disclosure and allocate responsibility for infringement.
8. Respect your own confidentiality. Uploading unpublished manuscripts or client designs to a tool that trains on inputs may compromise them.

A note for software developers

Code raises the same questions in a particular form. AI coding assistants can occasionally reproduce snippets that closely match publicly available code, which may be under open-source licenses with conditions such as attribution or sharing changes. Some assistants offer filters that block suggestions matching public code, or show references when a match is found. Turn those features on where available, run your usual license-scanning tools and keep human review of AI-generated code in place, just as for any third-party contribution.

Watching what comes next

The basics, that copyright requires a human author and that existing infringement rules apply to AI outputs, are stable. The open questions concern training, licensing markets and whether new legislation will address either. For authoritative updates, the Copyright Office's AI page is the place to start, and for a specific project, a short conversation with an intellectual property attorney is cheap insurance compared with a dispute later.