Generative art tools can produce a finished image, poem, or piece of music from a short text prompt in under a minute. That speed has forced a question that traditional art practice rarely needed to answer with precision: who, exactly, made the work? Courts, copyright offices, and platforms are now drawing lines between a person who types a few words into a text box and a person who spends hours adjusting parameters, selecting among hundreds of outputs, and editing the result by hand.

This article lays out the current state of that debate: the legal thresholds that have emerged, the arguments for and against treating prompting as authorship, the rulings decided so far in the United States, the United Kingdom, and China, and a practical checklist for anyone who needs to document a claim to a generative work. Where the law has not settled a question, that is stated directly rather than implied.

Why Authorship Became Contested

Copyright law in most jurisdictions was built around a simple assumption: a human being fixes an original expression in a tangible medium, and that act of fixation creates authorship automatically. Photography tested this assumption in the 19th century, and courts eventually accepted that choices of framing, lighting, and timing were sufficient human input to support a copyright claim. Generative models test the same assumption from a different angle: the model, not the person, performs the final act of fixation.

The contested part is not whether a computer is involved. Photographers use cameras, digital painters use software with auto-fill and smart brushes, and musicians use sequencers with built-in generative arpeggiators. The dispute centers on degree: at what point does a tool's contribution become large enough that the human's role shrinks from "author" to "requester"?

A second source of friction is training data. Diffusion and transformer-based image and text models are trained on large corpora of existing works, often scraped without individual licensing. Several lawsuits, including Getty Images v. Stability AI (filed in the UK High Court in 2023) and the consolidated artist lawsuits against Stability AI, Midjourney, and DeviantArt in the US District Court for the Northern District of California, argue that outputs are derivative of copyrighted training material. That question is separate from authorship of the output itself, but the two issues are frequently discussed together in public debate, which adds to the confusion.

The Human-Decision Threshold

Most copyright offices and courts that have addressed generative works ask a version of the same question: did a human make the creative decisions that determine the work's final expressive content, or did the machine make them? The US Copyright Office's guidance of March 16, 2023, states that copyright protection depends on whether the "traditional elements of authorship" were produced by a human or by a machine "mechanically in response to text prompts."

Under that standard, several categories of human input are treated differently:

• Prompt text alone: treated by the US Copyright Office as insufficient, because the model, not the person, determines how the prompt's concepts are translated into pixels or words.
• Selection and arrangement: choosing specific outputs from a larger set and combining them into a sequence or collage can receive protection for the selection and arrangement itself, even if the individual generated elements remain unprotected.
• Post-generation editing: manual retouching, compositing, color correction, or substantial hand-drawn additions on top of an AI output are treated as new human authorship covering those specific contributions.
• Parameter and seed control in some tools: fine-grained control over pose, layout, or composition through tools like ControlNet or layered inpainting masks has not yet been directly ruled on, but legal commentators generally argue it moves closer to the human-decision threshold than text prompting alone.

This threshold is not binary in practice. A single project might combine unprotectable AI-generated backgrounds with protectable hand-painted foreground elements, meaning different parts of the same image can have different copyright status.

Prompting as a Form of Authorship: Limits and Objections

Some practitioners argue that prompt engineering is itself a skilled creative act. A prompt that specifies lighting direction, lens type, color palette, artist reference, aspect ratio, and negative terms to exclude unwanted elements can run to several hundred words and take dozens of iterations to refine. Proponents of this view point to analogies with film directors, who do not personally operate the camera or edit the footage but are still credited as authors because they make the governing creative decisions.

The counter-argument, which has prevailed in US Copyright Office decisions so far, is that a prompt describes a desired result but does not control how the model arrives at it. Two users entering the identical prompt into the same model and seed will often get different outputs across sessions, and even small changes in prompt phrasing can produce results the user did not anticipate or intend. Courts and offices have treated this unpredictability as evidence that the causal link between human intent and final expression is too indirect to support authorship.

A middle position, sometimes called the "controlled generation" argument, holds that authorship should scale with the degree of deterministic control a tool offers. Tools that let a user fix a seed, lock a composition, and iterate only on texture or color are argued to sit closer to traditional digital painting than tools that generate an entire scene from an open-ended text description. No jurisdiction has yet adopted this graduated standard as formal law, though it appears in scholarly commentary and in some platform terms of service.

Case Law So Far: What Has Been Decided

Three decisions anchor the current landscape, each from a different jurisdiction and reaching a different practical outcome.

These three outcomes are not contradictory once the underlying facts are separated. Thaler explicitly disclaimed human authorship as a test case. Kashtanova claimed human authorship over text and arrangement but conceded the images themselves were machine output. Li documented a specific, iterative prompting and selection process and argued for authorship, and the Chinese court accepted that argument on its facts. The pattern across all three: outcomes track the amount and specificity of human decision-making that was claimed and documented, not the tool used.

Collective and Shared Authorship Models

Generative projects increasingly involve more than one person, which raises joint authorship questions separate from the human-versus-machine question. A common structure in commercial generative art work: one person writes and refines prompts, a second person curates outputs and performs post-production, and a third person directs overall concept and style guides. Whether this produces joint authorship depends on jurisdiction-specific tests.

Under US law, joint authorship under 17 U.S.C. §101 requires that each contributor intended their contributions to be merged into a unified whole and that each contribution be independently copyrightable. This creates a practical complication for generative workflows: if the prompting contribution itself is not independently copyrightable under current US Copyright Office guidance, a prompt writer may not qualify as a joint author even where the final piece, as edited by a second person, is registered.

Studios and agencies working with generative tools have adopted a few recurring contractual structures to manage this uncertainty:

• Work-for-hire agreements that assign any rights arising from an employee's or contractor's prompting and editing work to the commissioning company, regardless of how the authorship question is eventually resolved.
• Explicit written agreements among collaborators specifying percentage splits of revenue rather than relying on default joint-authorship rules, since those defaults are unsettled for AI-assisted work.
• Internal logs that record who entered which prompts, who selected which outputs, and who performed which edits, kept primarily as evidence in case a dispute or registration challenge arises later.

A Working Checklist for Claiming Authorship

Anyone seeking to register or defend a claim to authorship over a generative work should document the process, not just the outcome. The following steps reflect what US Copyright Office examiners and the courts above have treated as relevant evidence.

1. Record the original creative brief or concept note written before any generation began, dated and time-stamped if possible.
2. Save every prompt version used, including negative prompts, parameter values, and seed numbers, in sequence.
3. Keep the full set of generated candidates, not only the one selected, to demonstrate a documented selection process rather than single-shot acceptance.
4. Log every manual edit made after generation: layers added, retouching performed, compositing steps, and the software used for each.
5. Separate, in any registration application, the elements claimed as human-authored from the elements acknowledged as machine-generated, as the US Copyright Office now requires applicants to disclose AI-generated content and identify the human-authored portions.
6. Where multiple people contributed, draft a written agreement covering ownership and revenue splits before the project is finalized, rather than after a dispute arises.

Common Mistakes

Several recurring errors appear in both informal disputes and formal registration attempts. Treating prompt text as automatically protectable is the most frequent one, given that US guidance explicitly rejects this. A second common mistake is omitting AI disclosure on a registration application, which the US Copyright Office has stated can lead to cancellation of the registration if discovered later. A third is assuming that one jurisdiction's ruling, such as the Chinese court's decision in Li v. Liu, applies elsewhere; authorship standards differ by country and sometimes by type of work within the same country. A fourth mistake is relying on verbal agreements among collaborators instead of written splits, which becomes a problem specifically when a generative piece gains unexpected commercial value.

Next Steps

Anyone working professionally with generative tools should start a documentation habit now rather than after a dispute begins: save prompts, candidate outputs, and edit histories as a matter of routine workflow. Those planning to register a generative work for copyright should consult the current disclosure requirements published by their national copyright office before filing, since these requirements have changed more than once since 2023 and are likely to change again. For contracts involving multiple collaborators or commissioned generative work, a lawyer familiar with intellectual property in the relevant jurisdiction should review ownership terms before a project starts, since default legal rules in this area remain unsettled and vary by country.