A federal report says prompts alone do not make an AI user an author

The process wall

On 29 January 2025 the US Copyright Office published Copyright and Artificial Intelligence, Part 2: Copyrightability, the second instalment of a report drawn from more than 10,000 public comments. The process wall it addresses sits at the exact point a maker hits when a generated output does not quite match what was typed: the report's own illustration shows a prompt for a bespectacled cat smoking a pipe in a highly detailed wood setting, and the resulting image supplies the cat and pipe but invents a robe, a specific breed and a background the prompt never specified. That gap between instruction and output is the report's central evidence.

What the documents show

The report concludes that, given current generally available technology, prompts do not alone provide sufficient human control to make a user the author of an AI system's output, because prompts function as unprotectible ideas rather than as fixed expression, and repeating or revising a prompt many times does not change that, since each attempt is closer to re-rolling dice than to authorship. It draws a different line, however, for two other kinds of human contribution: a human-authored expressive input, such as an uploaded drawing that remains visibly present in the output, can carry its own copyright into the result, as the report's Rose Enigma registration example shows; and a human who selects, arranges or creatively modifies AI-generated material can hold copyright in that arrangement, the same compilation logic the Office had already applied to Zarya of the Dawn two years earlier. The report also surveys other countries, finding that Korea, Japan, the EU member states and a Beijing court have each reached broadly similar human-authorship positions.

Authorship and limitation

The mechanism is the same one running through the Office's 2023 guidance and the Thaler opinion: whether a human, not the system, determined the work's expressive elements. A detailed prompt still leaves an unpredictable, black-box system to fill gaps the words never addressed, which the report treats as a control failure rather than a merely aesthetic one. The limitation on legal change is explicit: the report recommends against creating new copyright or sui generis protection for AI-generated material, reasoning that flooding the market with unearned protection could undermine, rather than serve, the incentive structure copyright is meant to provide.

What survives

What survives for a maker is control over everything upstream and downstream of the generation step itself: the expressive inputs fed in, and the selection, arrangement and edits made afterward. Editorially, the practical move the report supports is keeping and describing those inputs and edits, since they are what will separate a registrable compilation from an unregistrable output.

• Compare the report's cat-and-pipe prompt example against a real prompt-to-output gap in a maker's own workflow.
• Check whether an input image, not just a text prompt, was used and remains visible in a finished piece.
• Read the report's international survey section for how another country's registrar treats the same question.

The report changes no statute; it explains one, and in doing so gives makers a clearer account of which of their choices a registrar will actually credit.