AI-Generated Works and Copyright: What Is Protected in 2026

Creators, product teams and in house counsel keep asking the same question: can you own copyright in something a generative AI tool produced? In the United States the answer is now settled at the level that matters most. Purely AI generated output is not protected by copyright. Where a person contributes real creative expression, that human contribution can be protected, and it can be registered.

This guide explains the rule, the registration process that follows from it, and how to build a content pipeline that produces work you can actually own.

The rule, and where it comes from

Copyright requires human authorship. In *Thaler v. Perlmutter* (D.C. Cir. 2025), the D.C. Circuit upheld the Copyright Office's refusal to register a work generated autonomously by an AI system, holding that a human author is required and that the machine could not be named as the author.

The Copyright Office applies the same test at registration. Works created solely by an AI system are not registrable. Where a work mixes human and AI generated material, the Office registers the human authored parts and excludes the rest, case by case.

You have to disclose AI generated content in the application. If the traditional elements of authorship in a work were produced by a machine, the Office will refuse registration for those elements. Leaving them out of the application is worse than declaring them, because an inaccurate application can undermine the registration later.

Prompting alone is not authorship. Writing a prompt, however detailed, has not been enough on its own. What counts is the expressive material a person actually contributes: the selection, the arrangement, the edits and the transformation.

The same direction of travel shows up elsewhere in IP. USPTO guidance on AI assisted inventions takes the position that only a natural person can be named as an inventor, with the analysis turning on the human contribution. The tests differ between patents and copyright, but the principle is consistent.

Outside the United States, other copyright offices apply a human authorship test too, though the wording, the evidence they expect and the treatment of computer generated works vary from country to country. We confirm the rule in each country before filing rather than assuming the US position travels.

How to secure protection when you use generative AI

The whole game is establishing and documenting human authorship. Six steps.

1. Design for human authorship

Decide up front what the human creative contribution is: the concept, the story, the brand voice, the visual system. Assign named people to the selection, arrangement, editing and transformation of any AI output. The more your team shapes the final expression, the stronger your position.

2. Treat AI output as raw material

Model output is a starting point, not a deliverable. Avoid shipping unedited, uncurated output. Transform it through human judgment, and record why one option was chosen and another discarded.

3. Keep authorship records as you go

• Log prompts, models, versions and parameters against each asset.
• Save intermediate drafts, tracked changes and layer stacks for text, image, audio and video.
• Keep before and after snapshots that show the human transformation.

These records are what turn "we edited it" into something you can actually put in front of the Copyright Office, or in front of a court.

4. Clear your inputs and read the tool's terms

Check what rights you hold in any data you control, and read the model provider's terms for what you may do with the output commercially. Training and input questions are still developing, so track official guidance rather than headlines.

5. Register the human authorship

US registration runs through the Copyright Office's electronic system, and our walkthrough of the eCO application and group registration process covers the mechanics step by step. For an AI assisted work:

• Identify the human authored components precisely.
• Disclaim the AI generated material that is not yours to claim.
• Add a short description of the human contribution, for example "original text, selection and arrangement; edits and transformations to AI generated imagery".
• Pay the filing fee and answer any Office questions about the scope of human authorship.

If you would rather not handle the application yourself, our copyright registration service prepares and files it, including the AI disclosure.

6. Contract for clarity

• Employment and contractor agreements should assign rights in human authored material and require disclosure of the AI tools used.
• Platform and vendor agreements should address warranties and indemnities for training data and output rights.
• Enterprise terms, where you can get them, are usually tighter than the consumer defaults.

Common pitfalls

• Assuming AI output is automatically yours. In the US, AI only material is not copyrightable, so copyright alone will not stop a competitor reproducing it.
• Not disclosing AI generated portions. An omission risks refusal now and an invalidity argument later.
• Treating the prompt as the work. The expressive contribution has to be the human one.
• Shipping unedited output. Without meaningful human selection, arrangement or transformation, there may be nothing to register.
• Ignoring chain of title in the inputs. Even where your edits are protectable, upstream data problems can trigger takedowns or claims.
• Confusing patent and copyright rules. Both emphasize human contribution, but the tests and the remedies are different.
• Assuming one country's answer applies everywhere. Check locally before you file or publish in a new market.

Recommendations

Build human authorship into the pipeline. For text, have human editors rewrite, reorganize and fact check, and keep tracked changes. For images and video, require human compositing, color work and sequencing, and keep the project files. For code, keep commit history showing which logic was human designed.

Document at the asset level. Prompt logs, model versions and selection rationale should live next to the asset, not in someone's memory.

Register accurately and on a schedule. File for the human authored components, describe them properly, and time filings around product launches so the registration exists before you need it.

Set an internal policy. Any public claim that something is "AI generated" should trigger a quick review of what you can and cannot assert about it.

Segment by risk. Marketing collateral usually justifies heavy human editing and registration. Product documentation needs human drafting standards. Data products are strongest when human analysis sits alongside any AI summary.

Keep watching. The Copyright Office continues to publish guidance on AI and registration practice, and the case law is still developing. Review your policy at least once a year.

The bottom line

Human authorship decides what you own. Purely machine generated expression is not protected in the US, and only the human authored parts of a hybrid work are. Keep the documentation tight, keep the contracts explicit, keep the registration disclosures accurate, and confirm the local rule in each country where the work will matter.

Sources

• U.S. Copyright Office, Copyright and Artificial Intelligence
• U.S. Copyright Office, Registration
• USPTO, Artificial Intelligence