Human editing carries the only registrable copyright in an AI-assisted asset

The distinction doing the work is between describing an output and fixing one. When you prompt a model you supply a description, and the model chooses the composition, the colours, the phrasing or the notes; copyright protects that specific expression, not the instruction behind it ^[8][9]. That is the whole mechanism. The question at registration is not how hard you worked but which parts of the fixed expression a human fixed ^[9].

Jason Allen's case is where the effort argument was tested. "Theatre D'opera Spatial" won a fine-art prize at the 2022 Colorado State Fair, and Allen reportedly reached it through more than 600 prompt revisions in Midjourney ^[10][11]. The Copyright Office refused registration because the image contained more than a trivial amount of AI-generated material that he would have to disclaim, and he declined to disclaim it, so it could not be registered as submitted ^[12]. The ground of refusal is the disclaimer, not the presence of machine output. Because the Office also accepts that a person's own words, editing, selection and arrangement remain protectable ^[3], the route left open is a registration that carves the generated material out and claims what the human did around it ^[18]. The 600 figure is reported rather than something the refusal itself establishes, and the revisions themselves bought no registrable interest in the pixels ^[19][11].

The extreme case is settled at both ends of the process. Part two of the Office's Copyright and Artificial Intelligence report, published in January 2025, concluded that existing law already answers the question and no new legislation is needed ^[4]. Stephen Thaler had sought to register "A Recent Entrance to Paradise" with his "Creativity Machine" named as sole author and been refused ^[5]; the D.C. Circuit affirmed in 2025 that a copyrightable work must be authored by a human being ^[6], and the Supreme Court declined the appeal in 2026 ^[7]. The doctrine is older than the models: the same human-authorship rule decided the monkey selfie case ^[14].

Thaler also ran the argument in patents, naming his AI as inventor, and lost in both the US and the UK, whose patent systems are likewise built around a human creator ^[13]. Patent law asked the same question in two jurisdictions and returned the same answer, closing off the adjacent-regime workaround ^[20].

What changes in a working pipeline is bookkeeping. If your archive is a prompt log, you have carefully preserved a record of the part that carries no rights ^[2][19]. The artefacts worth keeping are the ones that evidence the human contribution: the raw output as the model returned it, the edits applied on top, the choice made among candidates, and the arrangement in which the pieces ship ^[3].

The account is explicit that it is not legal advice and that the edges are genuinely unsettled, while the core rule rests on a detailed government report and an appeals-court ruling ^[16]. It also describes the UK as almost alone in doing something different and about to stop, without naming the provision or the date ^[15], so nothing here transfers to a UK filing.

The defensible asset is the difference between what the model returned and what you shipped ^[3][18]. That difference has to exist as a record before anyone asks to see it.