USCO Tightens AI Disclosure in Copyright Registration
The U.S. Copyright Office is pushing AI use in creative work out of the realm of vague disclosure and into something far more structured. Under its updated registration approach, applicants are expected to draw a clearer line between the parts of a work authored by a human and the parts generated by AI, whether the material is text, images, audio, or a mixed-format output.
The practical importance goes well beyond one more compliance step. Since 2023, the Office has required applicants to disclose more than a de minimis amount of AI-generated material and to describe the human author’s contribution. Where essential facts are omitted or misstated, a registration can be challenged, corrected, cancelled, or stripped of evidentiary value later in litigation. Once disclosure becomes more structured, filing strategy, recordkeeping, and claim drafting all become tighter.
The real shift is from identifying AI use to isolating protectable authorship
The most consequential part of this development is not the Office repeating that fully AI-generated output is not copyrightable. That principle has already been stated. What is changing is the level of granularity expected from applicants. A generic sentence saying that a work was “created with AI assistance” is becoming much less useful. The Office wants to know what exactly was written, arranged, edited, or visually reworked by a human, and what came from a generative system.
That matters because mixed works are where most filings now sit. A children’s book may contain human-authored text and AI-generated illustrations. A marketing video may combine AI imagery with original scripting, editing, sequencing, and sound design. A podcast package may include AI-assisted audio elements but human-authored narrative structure and branding choices. In its 2025 report, the Office reaffirmed that prompts alone generally do not provide enough human control to support authorship. What may still be protected are the human-authored expressive parts of the work, the creative selection, coordination, or arrangement of human and AI material, and sufficiently original modifications made to AI outputs.
A separate AI disclosure step changes how applications have to be written
Once AI disclosure is pulled out and made more explicit, many teams will realise their internal project records are too thin. It is no longer enough to know that a designer used Midjourney, Photoshop, and a video model somewhere in the workflow. The harder question is where the copyright claim begins and ends. Which layer was machine-generated? Which text was rewritten by a human? Which edits were merely corrective, and which ones added original expression? That is the level at which a serious application now has to be built.
This puts pressure on both in-house content teams and outside counsel. Drafting becomes an exercise in claim discipline. Overclaim, and the application risks sweeping unprotectable AI-generated material into the asserted authorship. Underclaim, and the applicant may give away protectable human contributions unnecessarily. The filing process is starting to resemble a rights audit: break the work into text, imagery, audio, arrangement, and editing choices, then decide with precision what belongs inside the copyright claim and what must be excluded.
The risk does not end at filing; it comes back in enforcement
Many applicants still treat AI disclosure as a front-end nuisance. That is a mistake. The larger risk appears later, when a registration is tested in a dispute. The Office has already warned that if information essential to registrability was omitted, it may cancel a registration. In litigation, a court may also disregard a registration if it concludes that the applicant knowingly supplied inaccurate information and that the truth would have led to refusal. In other words, a weak disclosure record can become a structural weakness in enforcement.
This is especially important for companies that rely on copyright registrations as part of platform takedowns, licensing negotiations, M&A diligence, investment materials, or infringement claims. The immediate target may be fraudulent attempts to register wholly AI-generated output as human work, but the more common danger lies in hybrid works with messy provenance. A registration that survives initial review may still prove fragile later if the applicant cannot explain the creative chain with enough specificity.
The best response is not a disclaimer template but a stronger authorship record
For most businesses, the useful response is operational rather than rhetorical. They should preserve version history, script drafts, edit logs, source tables, layered design files, and internal notes showing who made the decisive expressive choices. Prompt records may still help, but they are rarely the strongest evidence. The more important question is how a human shaped the final expression, not simply what the human asked the model to do.
Content-heavy organisations should consider maintaining a simple authorship matrix for each project. Who wrote the first draft? Who made the final expressive edits? Which visuals remained machine output, and which were transformed through original design work? Which sound elements were generated and then creatively restructured into a larger human-authored whole? As USCO makes disclosure more granular, the registration system is doing something bigger than adding paperwork. It is raising the evidentiary threshold for copyright claims involving AI. The future question will not just be whether AI was used. It will be whether the applicant can prove where human authorship actually resides.
