How to Copyright AI Art: The 2026 USCO Step-by-Step Guide

As a computer scientist working in AI, I spend a lot of time analyzing how to copyright ai art and where an automated system’s output stops and human judgment begins. That question turns out to sit at the center of one of the more consequential legal debates facing anyone building on generative tools today. This piece is a technical and case-law analysis of that boundary: how the U.S. Copyright Office and federal courts have drawn the line between raw AI output and human authorship, and why the underlying architecture of tools like Midjourney, Sora, and Suno makes that line harder to locate than it first appears. It draws on the Copyright Office’s own published decisions, its Part 2 Copyrightability Report, and the federal court dockets tracking these disputes as they unfold. I want to be upfront about what this article is and is not: I am a computer science researcher and patent holder, not a licensed attorney, and nothing here is legal advice for a specific project or filing. Where a question genuinely turns on the facts of a particular case, that is exactly the kind of question a licensed copyright attorney should answer, and I say so explicitly throughout.

Why a Midjourney Subscription Does Not Buy Copyright

A paid Midjourney, Sora, or Suno subscription does not, by itself, create any federal intellectual property right in what that tool generates. This is not a theoretical curiosity. It is the single most common misunderstanding I encounter when looking at how creators and technologists structure AI-assisted workflows, and it has already been tested, repeatedly, in federal court and before the Copyright Office’s own Review Board.

The practical consequence of that narrow holding is still significant. Where a competitor copies an unedited AI-generated image, a copyright-based claim generally will not succeed, because the U.S. Copyright Office treats raw output as falling outside the scope of protectable authorship in the first place. This applies identically whether the medium is a static image, a Sora-generated video, or a Suno-composed song. The output of a generative model, standing alone, does not qualify, regardless of how the underlying model was trained or how sophisticated its architecture is.

The Human Authorship Standard and the Technical Reason It Exists

Understanding how AI-assisted work is treated starts with understanding how the federal government defines authorship in the first place. U.S. copyright law exists to protect “original works of authorship,” and the U.S. Copyright Office and federal courts have consistently interpreted authorship as requiring a human origin. A camera can capture an image, but the human photographer holds the copyright, because they arranged the lighting, composed the frame, and chose the precise moment of capture. The creative judgment, not the mechanical device, is what the law protects. That same logic runs through more than a century of U.S. copyright jurisprudence, from early debates over whether photography itself could be authored, to the Copyright Office’s current framework for evaluating generative AI.

There is a genuinely technical reason this analogy breaks down for most generative AI tools in a way it does not for a camera, and it is worth spelling out plainly. A camera is a deterministic instrument: the same settings, pointed at the same scene, produce the same image every time, and every pixel traces back to a choice the photographer made about light, framing, or timing. A diffusion or transformer-based generative model is not deterministic in that sense. It samples from a learned probability distribution over a vast latent space, and running the same prompt twice, or perturbing it slightly, can produce meaningfully different results. The user supplies an instruction; the system, not the user, resolves that instruction into the specific arrangement of pixels, words, or notes that appear in the final output. That gap between instruction and execution is the analytical thread connecting the Thaler decision, the Copyright Office’s guidance documents, and the individual registration decisions examined throughout this piece.

Why a Prompt Alone Does Not Establish Authorship

This is a difficult distinction for many creators to accept, but it is worth being direct about it: writing a prompt, however detailed or iteratively refined, is not itself treated as authorship under current U.S. Copyright Office guidance.

A genuinely complex, carefully engineered prompt can represent real skill and judgment about what to ask for. The Office’s position, though, is that a prompt functions more like detailed instructions given to another party than like the act of authorship itself. If a person tells a human illustrator, “draw me a futuristic city at sunset with flying cars, in a cyberpunk color palette, with dramatic lighting,” they have given a detailed creative brief. The illustrator, not the person who wrote the brief, holds the copyright in the resulting drawing, because the illustrator is the one who executed the actual expression.

The Copyright Office’s 2023 Registration Guidance addressed this directly, explaining that because current text-to-image systems do not treat prompts as instructions for a specific, predictable result but rather as a starting point the system then interprets unpredictably, the user does not exercise the kind of control over the final image that authorship requires. This is precisely the stochastic-sampling behavior described above, expressed in legal rather than technical terms: a person entering a prompt occupies roughly the position of the client giving instructions, not the artist executing them, and current guidance treats that distinction as decisive.

This same reasoning surfaced again in one of the Office’s most closely watched and most misreported decisions, involving digital artist Jason M. Allen. Allen used Midjourney to generate an image he titled Théâtre D’Opéra Spatial — French for “Space Opera Theater” — which won a digital art award at the 2022 Colorado State Fair and, once its AI origins became public, ignited a national debate about machine-assisted creativity. Allen entered at least 624 iterative text prompts, then further refined the resulting image using Adobe Photoshop for cosmetic cleanup and Gigapixel AI for resolution upscaling. On September 5, 2023, the Copyright Office’s Review Board issued its final decision letter, affirming a refusal to register the work as submitted. The Board did not dispute that Allen’s Photoshop edits could reflect genuine human authorship, and it accepted that those specific edits could, in principle, be registered on their own. What sank the application was that Allen refused to disclaim the Midjourney-generated and Gigapixel-generated portions from his claim, insisting on protection for the piece as a whole. Because the Board found the underlying AI-generated content was more than de minimis and Allen would not exclude it, the entire application was denied — not because no human contribution existed, but because the application did not isolate it correctly.

Commercial Licenses Are Not the Same as Copyright

A specific misconception worth addressing directly involves how commercial-use terms from paid AI subscriptions, such as Midjourney, DALL-E, or a Stable Diffusion API plan, actually function legally.

Many subscribers read their platform’s Terms of Service and see language granting paid users broad rights to use, sell, and distribute what they generate, and reasonably conclude this must mean they hold federal copyright in the output. That conclusion does not follow, and understanding why matters for anyone building on generated content at any scale.

A platform’s Terms of Service is a private contract between the platform and its user. In substance, it typically says the platform will not assert a claim against the user for selling what the tool generated, and grants a commercial license to monetize that output without paying royalties back to the platform. What it cannot do is create federal copyright where none otherwise exists. Because a raw, unedited AI image generally lacks the human authorship the Copyright Act requires, no amount of contractual language from the platform changes that underlying legal status. The license and the copyright are two entirely separate legal instruments — one contractual, one statutory — and the gap between them is structural, not a matter of which platform’s fine print a person reads.

What the Zarya of the Dawn Decision Actually Established

The clearest illustration of how the Copyright Office treats hybrid human-AI works comes from a specific, well-documented case that remains the reference point for this area of law, and one worth distinguishing carefully from the Allen decision discussed above, since the two are frequently conflated in casual coverage despite reaching the same underlying conclusion through different facts.

In September 2022, artist Kris Kashtanova registered a graphic novel titled Zarya of the Dawn. The initial application did not flag the Midjourney involvement in the images, and after the Office learned of the AI use through Kashtanova’s own public statements, it reopened the registration for review. On February 21, 2023, the Office issued its final decision letter, limiting the original certificate to the work’s human-authored elements rather than canceling it outright.

The Office’s reasoning was specific and remains instructive. It held that the individual Midjourney-generated images themselves were not copyrightable, even accounting for the hundreds of prompt iterations Kashtanova had used to refine them, because the artist did not exercise the kind of direct, predictable creative control over the final image that authorship requires — the same stochastic-system distinction discussed above. The letter also noted that even where an author edits an AI-generated image afterward, the Office cannot always conclude those editing alterations are sufficiently creative to qualify for protection on their own, meaning editing alone is not automatically enough; the depth and documentation of that editing is what the analysis turns on, a point this piece returns to in the next section. At the same time, the Office confirmed that Kashtanova did hold valid copyright in two specific elements:

1. The original, human-written text and narrative dialogue throughout the graphic novel.
2. The human-authored selection, coordination, and arrangement of the images and text as they appear on each specific page.

This decision directly informed the Office’s broader March 2023 Copyright Registration Guidance, and the same underlying framework was applied again just months later in the Allen decision covered above. It has since been reinforced by the Office’s more detailed Part 2 Copyrightability Report, published in January 2025, which remains the most current and authoritative statement of the Office’s analytical framework as of this writing. Reading Zarya and Allen together is genuinely useful: Zarya shows a creator who registered successfully by clearly disclaiming the AI-generated images and claiming only the compilation and text. Allen shows what happens when a creator with an arguably stronger, more heavily edited fact pattern refuses to make that same disclaimer. The lesson is less about how much editing occurred and more about how the application itself is structured.

So, can a comic book made with AI art be copyrighted? Yes, but what gets protected is the human-authored compilation, the story, and the layout, not the individual AI-generated panels standing alone. A pirate reproducing an entire page, including the original layout and story sequence, faces a real infringement claim. Someone who crops out a single unedited AI-generated background element and reuses it in isolation generally does not.

A Pattern Across Published Decisions: Human-Modified AI Content

Reading across the Copyright Office’s published decisions, the Part 2 Copyrightability Report, and the currently pending Allen litigation, a consistent pattern emerges in how the depth of human modification maps to registration outcomes. The Office treats raw AI output as unprotected raw material, comparable to a block of digital clay or licensed stock footage, and looks for genuine human creative work built on top of it. What an examiner is trying to identify is a specific human contribution that is creative and transformational, not mechanical or trivial — and, separately from the depth of editing, whether the application itself properly discloses what the AI generated, which is the lesson the Allen case adds to the Zarya framework.

Reaching the stronger end of that spectrum means meaningfully changing the aesthetic or structural expression of the AI base output, rather than making surface-level adjustments. Across the published decisions, the human contribution that succeeds at registration functions as the dominant creative force in the final result, not as a finishing touch applied to an otherwise unedited machine output. The difference between a five-minute color correction and a multi-hour compositing session is not just a matter of effort; it is the fact pattern an examiner, or later a court, actually looks at — and, as the Allen case illustrates, it sits alongside a separate question of whether the application accurately separates the disclaimed AI material from the claimed human material in the first place.

A Step-by-Step Guide on How to Copyright AI Art

Once a workflow involves genuinely substantial human modification, the Office’s registration framework asks applicants to do something specific: separate the preexisting, uncopyrightable AI-generated material from the new, human-authored material, and describe each category in its own terms. Current guidance requires disclosing AI-generated elements and identifying the human contribution; omitting a material fact about AI involvement can put the resulting registration at risk. That structural requirement, not the specific mechanics of any one federal form, is what is worth understanding here — completing an actual application, and deciding how to characterize a specific work’s contributions, is exactly the kind of task that calls for the Office’s own instructions or a licensed attorney rather than a general analysis like this one.

Conceptually, the framework divides a hybrid work into two categories. Preexisting, AI-generated material is excluded from the claim; it is treated the way a copyright application would treat any other uncopyrightable base material the applicant did not create, such as public-domain source text or licensed stock elements. The new, human-authored material — the specific structural edits, overpainting, compositing, or original arrangement a person actually executed — is what the applicant claims and what the examiner evaluates for the kind of creative, non-mechanical contribution described in the previous section. This is precisely the distinction Jason Allen declined to draw for the AI-generated portions of his work, which the Review Board identified as the specific reason his application could not proceed as submitted.

Separately, it is worth noting that the Copyright Office periodically adjusts its fee schedule to reflect its operating costs, most recently through a proposed rulemaking process working its way through the required period of congressional review as of this writing. That process is a matter of Office administration and is unrelated to the disclosure standard itself; anyone with a live filing decision should confirm current fees directly on the Copyright Office’s official fee schedule rather than relying on a figure quoted in an article like this one.

What the Published Record Shows About Undisclosed AI Use

Some creators consider simply declining to disclose AI involvement, on the theory that an examiner is unlikely to notice. The published decisions discussed throughout this piece suggest that theory is weaker than it looks, and the statutory structure surrounding registration is worth understanding on its own terms, independent of any specific case.

Where This Analysis Ends and Legal Counsel Begins

For a solo hobbyist thinking through a few personal projects, the analysis above generally covers the conceptual landscape. For a funded startup, agency, or studio building commercial value on AI-assisted workflows, the stakes and complexity increase considerably, and specific scenarios call for a qualified intellectual property attorney directly rather than a general-audience piece like this one. I want to be direct about the boundary of what a computer scientist and patent holder, writing about copyright rather than practicing it, can responsibly cover here: explaining published registration decisions and public court dockets is one thing, but the moment an actual dispute or a high-value transaction is on the table, the analysis becomes genuinely fact-specific in a way that only a licensed copyright attorney reviewing the real documents can properly navigate. This article was not reviewed by an attorney before publication, and every specific legal proposition above is tied to a primary source rather than to a legal opinion of my own.

• Commercializing AI-heavy code: If a startup’s core product relies heavily on AI-generated code from tools like GitHub Copilot or Claude, an attorney conducting an IP audit can help separate the uncopyrightable AI-generated portions from proprietary human logic and structure the resulting “Literary Work” registrations appropriately.
• Pursuing an infringement claim: If a properly registered, substantially human-modified work is copied by a larger entity, pursuing statutory damages under 17 U.S.C. §504 through federal litigation requires the kind of case preparation an experienced litigator provides.
• Enterprise licensing agreements: A creator selling exclusive rights to AI-assisted character designs or other assets to a larger commercial partner benefits from an attorney-drafted licensing agreement that clearly establishes chain of title and avoids future ownership disputes.
• Filing during an unsettled area of law: Given that the Allen litigation described above remains pending and could shift how the Office evaluates similar fact patterns, a studio with significant portfolio value tied to heavily AI-assisted work benefits from counsel who is actively monitoring that docket, rather than relying on a snapshot of the law as it stood at any single point in time.

How UK Law Treats Computer-Generated Works Differently

The strict U.S. human-authorship standard is not universal, and comparing it to another major jurisdiction’s approach is a useful way to see how much of the U.S. position reflects a deliberate policy choice rather than an inevitability of the technology. This divergence is not a minor technicality; it is one of the more consequential structural differences in global IP law for anyone building a content operation across borders, and it was directly raised — and rejected as a controlling argument — during Thaler’s unsuccessful Supreme Court petition, where he specifically pointed to the UK and Chinese approaches as evidence the U.S. rule was out of step internationally.

Technical Safeguards Beyond Copyright Registration

Given how narrowly U.S. copyright currently protects raw AI output, creators and technologists working at meaningful scale are increasingly layering technical provenance and tracking measures on top of legal registration, rather than relying on federal registration alone. This is where the analysis moves from case law back onto more familiar ground for a computer scientist: legal protection and technical protection are complementary tools, not substitutes for one another, and the strongest asset-protection strategies I have seen in patent and IP work generally combine both rather than leaning entirely on one.

Where the underlying copyright claim to a raw image may itself be legally uncertain, making the file technically harder to strip of provenance information, or to reuse undetected, offers a genuinely independent layer of protection that does not depend on how a court eventually resolves the Allen case or any similar dispute.

The Analytical Bottom Line

Across every decision examined in this piece, obscuring AI involvement when filing a federal copyright application has consistently been treated as a misrepresentation with real downstream consequences, not a harmless omission.

Examiners have become increasingly attentive to signs of AI generation, including characteristic visual artifacts and structural inconsistencies, and the D.C. Circuit’s now-final ruling in Thaler v. Perlmutter has only reinforced how seriously the human-authorship standard is being applied across the Office’s decisions. The pending Allen litigation, meanwhile, is the clearest sign that the precise boundary of that standard — not whether it exists, but exactly how much documented human work is enough to cross it — remains an open, actively contested question rather than a settled one.

This regulatory picture, though strict, is also genuinely predictable once the underlying technology is understood. Generative AI can be used to brainstorm concepts quickly and produce raw visual, textual, or audio material at scale, because that is what a stochastic, sampling-based system is built to do. The legally protectable value lives in the documented human creative work built on top of that raw material — which is really just the human-authorship requirement expressed in the vocabulary of how these systems actually function. For anyone building a content or technology business on generative AI at real scale, understanding that boundary, and where the open questions in the case law still sit, is worth treating as a standing part of the picture rather than a one-time read.

Podcast

FAQs

Sources and Legal References