Can Prompts Qualify for Copyright?: Section 202.1 in 2026 — Length or Human Expression?

The U.S. Copyright Office's Part 2 report on copyright and generative AI provides the surprising answer to the current prompt question: no word-count shortcut exists. For Section 202.1, the legal focus remains human authorship, not whether a prompt is brief, expansive, polished, or heavily revised.

Two inquiries matter. A sufficiently original human-authored prompt may qualify as text in its own right, but that does not make the AI response protectable. The Office says merely entering prompts is not enough because generative output is not perfectly predictable. Claiming authorship of the prompt and claiming authorship of the output are separate questions.

Across four possible human acts—drafting the prompt, selecting AI material, arranging it, and revising it—protection follows the human-created expression. Purely AI-generated text and images receive no copyright protection, while original human edits may be protected alongside AI material. These human acts and the absence of automatic thresholds frame the assessment; length supplies context, not authorship.

The § 202.1 Gate

Prompt length is not the § 202.1 gate; human authorship tied to the claimed expression is. I start with the Copyright Office’s human-authorship standard under 37 C.F.R. § 202.1(a): a machine or generative process cannot be the author where no human has determined a sufficient human contribution to the claimed expression. That is an element-specific inquiry. I ask who determined the very words, image features, composition, or revisions the claimant seeks to cover—not who merely started the process.

I then apply 37 C.F.R. § 202.1(b) claim by claim. AI material does not make an entire mixed work ineligible merely because it is present. But protection reaches only the portion for which a human author determined a sufficient contribution. I therefore separate the deposit into human-authored text, generated expression, and human acts affecting the claimed work. Rejecting a claim to generated material need not invalidate a distinct claim to qualifying human expression; treating the whole package as human-authored does.

For a DALL·E prompt, I first treat the prompt as a written document. In “A brass astrolabe rests open on an astronomer’s ledger while candlelight catches its engraved rings,” the original human-authored sentence can be evaluated for copyrightability as text. By contrast, directions to show the astrolabe from above, use a specified canvas ratio, render in a designated style, or request revisions remain functional instructions. They may explain how the model operated, but they do not become claimed image expression merely because the output displays their results. A longer prompt is not safer: length cannot convert directions into authorship.

The 2025 rulemaking record supplies my audit inventory: selection, arrangement, coordination, and modification. For each act, I ask whether it embodies an original expressive choice in the claimed work or merely makes the generative system work.

Finally, causation is not authorship. A prompt may cause DALL·E to return a particular image, while the model’s parameters and operation supply the generated expression. Neither technical causation nor bare acceptance of a returned result substitutes for qualifying human choice. A deliberate choice among fixed candidates may support a selection claim; a human-authored arrangement or substantive rewrite may support a claim in the output. My close is concrete: circle the exact prompt words and output features the claimant authored or sufficiently determined, label the qualifying human act, and cross out everything whose claimed status rests only on having caused the model to run. That is the § 202.1 boundary.

The Filing Context

The filing context concerns demand for copyright claims involving AI, not the rate at which those claims succeed. I read the U.S. Copyright Office’s January 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability, as a reminder about claim scope: counsel must identify the human expression actually being claimed rather than treat the presence of AI as the claim’s endpoint.

The supplied filing record does not establish a volume trend or a validity rate. An application is an assertion of copyright, while protection ultimately depends on the claimed work’s human-authored expression. A crowded docket can reflect experimentation with new tools, uncertainty about registration policy, or strategic claiming without implying that courts or the Copyright Office will treat an entire AI-assisted work as human-authored.

The U.S. Copyright Office’s January 2025 Copyright and Artificial Intelligence, Part 2: Copyrightability report supplies the relevant content-sensitive distinction. It says that merely entering prompts is insufficient to claim authorship of the output, while creative selection and arrangement of AI-generated elements can qualify as human authorship. The report supplies no word-count threshold. A long prompt is therefore not automatically copyrightable, and copyright in original prompt wording does not automatically carry over to the model’s output. A prompt claim reaches only original human-authored expression actually embodied in the prompt, not functional directions. A final-work claim reaches qualifying human edits or original selection and arrangement, while unedited AI-generated expression remains outside the claim.

Thaler v. Perlmutter reinforces that claim-specific approach. The court vacated the Copyright Office’s 2023 refusal for a mixed human/AI comic and remanded for analysis of the contributions actually made by the applicant. It did not adopt a categorical bar on all AI-assisted expression. For a comic, the presence of generated panels is not dispositive; the human-authored elements and creative choices must be identified and evaluated on their own terms.

For a mixed-work filing, I would use an element-by-element claim chart: record the source of each component, identify the human modification or creative choice actually reflected in the final work, and mark model-only expression as excluded. That record turns “AI-assisted” from a label into a reviewable copyright claim. The filing context underscores the need for a precise claim—not breadth of protection.

Prompt vs. Edit

The winning claim is the narrowest unit containing identifiable human expression. I separate human-written prompt text, the final output, and any compilation before testing authorship; a right in one unit does not enlarge rights in another. Prompt length is no safe harbor: a lengthy instruction set is not automatically copyrightable, and copyright in expressive prompt wording does not automatically cover model-generated expression.

That division matches the Copyright Office evidence summarized in the sources. According to Medium’s summary of the U.S. Copyright Office report, AI is a tool used in producing human-created work, not a source of authorship itself, and model output is not perfectly predictable from the prompt. The same summary states that purely AI-generated images receive no protection. Variety, quoting the report, identifies creative selection and arrangement of AI-generated elements as qualifying human authorship.

The overall winner is an accurately scoped claim to minimally creative human expression. Original prose or dialogue in the prompt can be claimed as words; a creative post-generation change can be claimed as a delta; and creative ordering can be claimed as compilation. Functional instructions and whole-output theories lose.

The hard edge is separability. If a writer supplies original dialogue and the model supplies the surrounding scene, I claim the dialogue, not the scene. If a human later replaces an image region, rewrites passages, or recomposes elements, I claim that human-created delta. When human and AI contributions are not readily separable, substantial human editing alone does not make the whole-work claim supportable.

The supplied source record does not resolve whether prompts themselves qualify or how particular edits change the analysis, and its excerpts describe a qualitative “significant human creativity” standard without a percentage, word count, or numerical threshold. That is why I do not treat prompt volume as evidence. I select the strongest accurately described human layer, not the commercially most valuable one: output novelty and model quality cannot substitute for authorship.

For a redline-ready handoff, counsel, design, and engineering should use the same schedule. Each row states the human contribution, the claim boundary, and the AI remainder that registration language must exclude.

What the Data Doesn’t Tell You

The register is a poor proxy for authorship. 37 C.F.R. § 202.1 specifies no minimum prompt length, word count, iteration count, edit count, or percentage of human contribution that guarantees protection. The inquiry concerns the human expression actually embodied, not a score generated by counting tokens or interventions. The myth that a sufficiently long prompt is automatically copyrightable—and that any resulting right travels into the model’s output—fails at both steps.

Nor do prompt text and generated output form one evidentiary unit. Variety, quoting a U.S. Copyright Office report, says that merely entering prompts is insufficient to claim authorship of the output; it does not say that every human-written prompt is unprotectable. An adverse outcome for an output-registration claim can leave sufficiently creative, nonfunctional wording in the prompt separately protectable. The reverse is equally strict: protection in that wording would not establish rights in generated expression. Functional directions remain instructions, not protectable expression merely because a person typed them.

Nor does an edit automatically cure machine authorship. A change that merely dictates the model’s existing language, applies a template, or makes a cosmetic adjustment can leave the machine-authored element unprotectable. Calling an output “final” does not convert its generated words or pixels into a human-authored work. Protection must remain limited to sufficiently creative human edits or selection and arrangement, with the machine-authored remainder excluded.

A registration record has the same limit: it establishes what was submitted, not that AI-generated words or pixels are human-authored. Copyright Office examination is record-based, not a forensic experiment. A later declaration unsupported by dated prompts, versions, or source files therefore carries less evidentiary weight than a contemporaneous creation trail.

Variation supplies a further causal warning. The same nominal prompt can produce materially different output when the model version, random seed, account setting, or generation date changes. That makes it harder to infer which human act produced a particular passage. It does not change the authorship test: nondeterminism neither removes protection from a qualifying human edit nor makes the variable machine passage human-authored. When the evidence stops at prompt entry and a “final” label, the defensible claim covers only the narrow unit containing minimally creative human expression; the rest remains outside it.

Allen’s Generation Workflow

Allen’s workflow did not turn a large generation run into copyright by arithmetic. In Allen v. Perlmutter, Jason Allen’s Thé"tre D’opéra Spatial was a visual work assembled from generative-model output and post-processed in Adobe Photoshop. The court assessed the human expression actually embodied in the submitted work; it did not place generated pixels under Allen’s control merely because he drove the workflow.

According to Allen’s account in the district-court opinion, he generated a batch of Midjourney images, selected a subset, and altered the selected set through operations including cropping, color adjustment, and upscaling before seeking registration for the assembled work. Those acts create distinct candidate layers—prompting, selection and arrangement, and post-generation changes—rather than one generalized authorship claim.

The legal question was narrow: did those acts identify original human expression in the submitted images, or did the claim remain directed at unelaborated model output? That distinction kills the prompt myth. Copyright may cover original human-authored prompt text, but a right in that text does not automatically cover model-generated expression. Nor does naming Photoshop establish that every adjustment is creative enough; each asserted human contribution must be traced to the claimed work.

The Allen-specific point is that the ruling rejected his broad copyright claim because the qualifying human contribution was not sufficiently demonstrated. It did not categorically exclude every prompt, modification, or mixed work. Copyright remains available for original human-authored prompt text and for sufficiently creative human edits or selection and arrangement, while functional instructions alone and unedited AI-generated expression cannot be claimed merely because a person entered a prompt.

The generation, selection, and editing steps did not establish authorship by arithmetic. My docket lesson is that Allen did not lose because he used Midjourney; the human-to-AI contribution map was not persuasive. Under the current 37 C.F.R. § 202.1, the disciplined filing move is a claim chart tied to the submitted work: identify the creative choices embodied in the prompt, selection and arrangement, and each asserted edit, then remove every layer for which the claimant cannot point to original human expression. That is not a request for more labels or a longer prompt. It is an auditable boundary around what the human actually contributed.

How to Choose Well

The strongest claim is usually a layer, not the whole file. When I review an AI-assisted claim file, I ask which human expression is identifiable and traceable to a qualifying act. A generation log shows activity; it does not establish authorship.

Begin with the prompt as text, not as a proxy for the result. Claim its wording only when a named human wrote original, minimally creative prose or dialogue. If the wording merely directs subject, style, format, or iterations, classify it as an instruction rather than protected expression. Mixed prompts require element-level treatment: preserve an original expressive passage, but do not sweep expressive-sounding labels, constraints, or production directions into the claim unless they independently satisfy the rule.

Inspect the unedited output next. If no human made an expressive change and no creative selection or arrangement exists, classify the model’s output as non-copyrightable machine or process material. Arrival through a publishing tool does not change that result, and the presence of multiple generated variants does not itself create a human-authored selection.

After editing, compare the before-and-after versions. Claim the delta only when a human rewrote words, drew new elements, or made a creative compositional change. Resizing, metadata changes, format conversion, or color adjustment alone are insufficient. The file history should identify the qualifying change and isolate the newly added expression; it should not convert the entire machine-generated substrate into a protectable work.

When the asserted value lies in multiple outputs, treat the compilation as a separate claim layer. Require a documented creative arrangement reflecting the choice or order of those outputs, and claim that layer only—not rights in each unprotectable element. A folder, sequence, or bundle arranged solely for convenience or access is not enough.

Proof is the final branch. Connect every claimed element to a named human, a dated version trail, and an element-level chart showing the act that produced it. If that chain is missing, the whole-work claim is unsupported. A long prompt and repeated iterations cannot cure the gap. Before filing, reconcile the prompt record, native output history, before-and-after files, and claim chart; then remove every element that cannot survive that reconciliation.

What to do next