The registration form makes you disclaim the part you paid for

The rule is a disclosure duty, not a verdict on your video

The Office's own nine page version of it, signed by Register of Copyrights Shira Perlmutter on 10 March 2023, states the principle in a line: copyright can protect only material that is the product of human creativity. Then it turns the principle into a filing instruction. When an AI technology determines the expressive elements of its output, that material is not protected by copyright and, in the Office's words, must be disclaimed in a registration application.

Must be disclaimed. Not may be refused, not is legally void: disclaimed, by you, in writing, on an application you sign. AI output being uncopyrightable is a fact about the output, and it costs you nothing until the day you file. The duty to disclose and exclude is a fact about your paperwork, and it arrives the moment you decide to register anything.

None of it arrived with generative video. Chapter 300 of the Compendium of U.S. Copyright Office Practices, its pages footered 28 January 2021, already refused a photograph taken by a monkey, a mural painted by an elephant and driftwood shaped by the ocean, on the ground stated at section 313.2: a work must be created by a human being. The 2023 statement added a field on a form, not a rule about machines.

The $65 form is the only one with a field to disclaim in

The Standard Application is $65 on the Office's fee page. The single application, for one author, one claimant, one work, not for hire, is $45. If your video carries more than de minimis AI content you do not get the choice: the Federal Register notice says the Office's other types of application forms do not contain fields where applicants can disclaim unprotectable material such as AI-generated content. The twenty dollars is not the story. The missing field is.

The Office publishes the words to type. Applicants must use the Standard Application and provide a brief statement in the Author Created field describing the authorship contributed by a human, and the guidance hands you the sentence: selection, coordination, and arrangement of the human-authored content created by the author, and the AI content generated by artificial intelligence.

The exclusion goes somewhere else entirely, in the Limitation of the Claim section, under the heading Material Excluded. One field describes what a person made. The other names what you are giving up.

Correcting it afterwards costs more than the certificate did. Supplementary registration is $100 filed electronically, against $65 for the original, and applicants who fail to update the public record risk losing the benefits of the registration. Two mechanisms, both named in the guidance: the Office may take steps to cancel a registration, and a court may disregard one under section 411(b) where the applicant knowingly gave inaccurate information that would have led to a refusal.

The 7,000 registrations exist because the AI was disclaimed

The Register's written testimony of 12 May 2026 puts it in one clause: the Office has to date registered more than 7,000 claims that include AI-generated materials where those materials have been disclaimed in accordance with our registration guidance. The disclaimer is not the price of a weaker certificate, it is the mechanism by which a certificate issues at all. Hold that against the 2023 policy statement's own figure for the Office's intake: roughly 500,000 applications a year.

Part 2 of the Office's report on copyright and artificial intelligence, published 29 January 2025, put the same count at hundreds of works. Fifteen months later it is more than seven thousand, so a figure lifted from an article written in between is true of the report before last.

Part 2 also answers the question a video product actually has. The inclusion of elements of AI-generated content in a larger human-authored work does not affect the copyrightability of the larger work as a whole, and its worked example is a film with AI-generated special effects or background artwork. The D.C. Circuit went the same way on 18 March 2025 in Thaler v. Perlmutter: all work must be authored in the first instance by a human being, and that requirement does not prohibit copyrighting work made with the assistance of artificial intelligence. Certiorari was denied on 2 March 2026.

A film with generated effects in it is copyrightable. The effects are not, and the form is where you say which is which.

Iterating harder does not make the output yours, and Japan counts the iterations anyway

Zarya of the Dawn is a registration the Office granted and then reopened. Eighteen pages, one of them a cover, registered as VAu001480196 with an effective date of 15 September 2022, and unpicked by letter on 21 February 2023 for a single reason: the registration did not disclaim its Midjourney-generated content. The letter says the Office intends to cancel the original certificate and issue a new one covering only the expressive material the author created.

The letter quotes the applicant on how those images were made. Kashtanova described trial and error in which she provided hundreds or thousands of descriptive prompts to Midjourney until hundreds of iterations created as perfect a rendition of her vision as possible. The Office found no human authorship in the images regardless, because prompts function closer to suggestions than orders. Teams treat prompt iteration as the labour that earns the right. In the United States it earns nothing.

Japan reads the same facts differently, which matters when somebody quotes a foreign guideline at you. Guidelines from May 2024, quoted in Part 2, decide copyrightability case by case on the content of the prompts, the number of generation attempts, the selection from multiple outputs and subsequent human corrections. Generation attempts are a factor there and none here. What survived in Zarya was the text plus the selection, coordination and arrangement of its elements, and footnote 138 of Part 2, quoting the Compendium at section 312.2, puts a floor under that: a selection of two or three elements is necessarily de minimis.

Licensed footage comes off the claim too, and for a different reason

Generated frames are not the only layer that comes off the claim. Chapter 600 of the Compendium, at section 618.2, lists four kinds of unclaimable material a registration does not extend to: previously published material, previously registered material, material in the public domain, and copyrightable material that is owned by another party. Where a work contains an appreciable amount of it, the applicant must exclude it and limit the claim to what the author made.

That fourth category is your stock footage. The Pexels licence says all photos and videos on Pexels are free to use and that attribution is not required, and says nothing about copyright transfer. Getty's agreement, in the Dutch text its site serves, is blunter: all licensed content is the property of Getty Images or its content suppliers, licensed to you perpetually. A perpetual worldwide right to use somebody else's footage is a good deal. It is not authorship.

Vendor terms are the same trap from the other side. ElevenLabs, in terms last updated 31 March 2026, says you retain all rights in and to your Output, and in the next breath takes a licence back that is perpetual and irrevocable, nonexclusive, royalty-free, worldwide and sub-licensable through multiple tiers. Both halves can be true and neither is about copyright. A contract settles what the vendor will not do to you. It cannot create a right the Copyright Act does not recognise, and it is not the document the Material Excluded field asks about.

The two layers you can claim are the script and the running order

Put the guidance and Part 2 side by side and what falls out is not a yes or a no. It is a production process. The Author Created field wants selection, coordination and arrangement described in words, and Part 2 holds that human authors own exactly that: the creative selection, coordination or arrangement of material in AI-generated outputs, and any creative modifications of them. More than two or three elements of it, says the footnote.

So the written script is claimable: a person wrote it and it is perceptible in the finished cut. The order of the shots is claimable because a person chose it. Everything between those two layers is either somebody else's or nobody's. A pipeline that stops to make a founder approve a script, then assembles shots in an order a person can point at, has something to write in the field. Footnote 25 of the Register's testimony says more guidance on the Office's disclaimer practices is coming later this year, and only a team that kept a record of who chose what will have an answer for it.

A human approval step is not a slower pipeline. It is the part of the video you can put your name to on the form.

What to do about it

1. Before a video containing any generated element is published, write the two sentences the Standard Application will ask for: what a person authored, and what to name under Material Excluded.
2. File on the $65 Standard Application rather than the $45 single application, and correct an omission by supplementary registration at $100 instead of leaving the public record wrong.
3. Keep the approval step and keep its record, because selection, coordination and arrangement is the layer you can claim and a selection of two or three elements is too thin to count.