Can You Copyright AI-Generated Work When Human Authorship Is Still Required?

Can You Copyright AI-Generated Work When Human Authorship Is Still Required?

Yes, you may be able to copyright a work created with AI, but protection generally depends on whether a human contributed original expression that can be identified and explained. In the United States, the Copyright Office does not offer copyright to a work that a human never authored merely because it was produced through an AI system. That does not automatically disqualify every AI-assisted book, illustration, song, website, or design. A writer who selects, arranges, edits, and expresses the final content can qualify, even when some intermediate images or drafts came from a model. The legal question is not simply whether a tool was used; it is whether the submitted work contains sufficient human authorship. If the human contribution consists only of entering a prompt, copying the output, and making trivial changes, the risk of a registration refusal is substantial. As of September 2026, the prudent approach is to treat AI as an aid or a source material while identifying and documenting the human decisions that actually shape the protected expression.

Also worth reading: How Can Authors Register AI-Generated Content Under Copyright Law in 2026? · What Does the Future of Human Authorship Look Like in 2026 and Beyond? · How to prove human authorship for AI in 2026?

The same basic rule applies broadly outside the United States, although the statutes, exceptions, and treatment of computer-generated works differ. Indonesia’s Copyright Law No. 28 of 2014 protects works of authorship by human authors, and the country’s reported rules on AI-generated material follow that basic separation. International treaties generally connect copyright to natural or legal persons rather than granting rights to a machine by itself. Courts and registries are still developing answers where humans and autonomous systems contribute different elements. An author therefore should not assume that booking a publishing agreement, paying a designer, or using a human editor eliminates a missing-authorship problem. Protection attaches to the particular expression a person created, not to a commercial process merely because people were involved in that process.

What Counts as Human Authorship Under U.S. Copyright Law?

U.S. copyright requires original authorship, and the Copyright Office’s 2023 guidance, “Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” says human contributions must be identified. That guidance distinguishes assistance from control. A person may organize, evaluate, revise, and combine AI-generated material in a way that reflects creative choices, but the person cannot claim authorship of every expressive feature the model supplied. The Office has said that merely describing an intended result, receiving several options, accepting one unchanged, or making routine production adjustments is generally not enough. The application of copyright law also depends on the “modicum of creativity” standard used in registration practice, which is more demanding than the absolute absence of any human effort.

The Supreme Court’s March 2025 denial of review in the Stephen Thaler litigation reinforced the absence of a general rule that an AI can be an author. That decision did not decide every case involving AI-assisted creation, and it should not be presented as holding that no human-created arrangement of AI material is protectable. The Copyright Office’s broader work is more useful for a writer facing mixed human and AI output because it focuses on whether the human contribution is disclosed, original, and actually reflected in the work. The U.S. Copyright Office also warned that purely AI-generated material may be excluded from registration even if similar human-made output receives protection. The practical burden falls on the applicant to separate the human-created material from the machine-created material.

To register a mixed work, an applicant must explain what the tools produced and what the person produced. Saying that the author “used AI for illustrations” does not disclose whether the person redrew the line work, chose the composition, adjusted colors, corrected anatomy, and supervised the final layout. Copyright offices can request source material, drafts, instructions, and a description of the creative process. A creator who cannot reconstruct those decisions has a weaker claim than one who designed the work deliberately and retained evidence. The standard is therefore factual rather than a guarantee based on professional polish.

Do Prompts and Creative Direction Qualify as Human Authorship?

Some prompts can contribute expressive material, but there is no judicial rule saying that every sufficiently detailed prompt is equivalent to an authored work. A prompt that selects a rhyme, arranges words in a distinctive meter, or supplies the specific language later used in a poem may supply evidence of human expression. A generic request such as “write a 1,200-word chapter about a detective in Victorian London” ordinarily does not explain which language the writer actually authored. A prompt that is extremely lengthy is not automatically stronger; an examiner will examine what the words protect and how they relate to the work being registered.

Creative direction can matter in visual projects, although output selection alone may not meet the Office’s ordinary expectations. A client who supplies a synchronized shot list, dictates a particular visual hierarchy, and specifies how the final design must function may have authored the relevant arrangement. A customer who chooses among finished AI images, without materially altering or arranging them, may have commissioned a result without creating its expressive elements. A publisher should therefore not describe a model as a coauthor merely because the model generated the first draft, even if the human team shaped the broader manuscript. Keeping the model named as a tool, while identifying the human author as responsible for selection, revision, and expression, produces a clearer record.

Writers should preserve both the prompt history and the editorial chronology. Prompt logs show what the human requested, but they do not alone prove authorship of the output. Revision files are usually more important because they show what the human changed and rejected. File histories, comments, reference notes, version comparisons, and dated outlines can establish that the final work emerged through directed human decisions. A creator who uses AI to overcome blank-page paralysis or generate nonprotectable research notes can still write the final book, provided the final text reflects that person’s authorship. The safest strategy is not to avoid prompts, but to make prompt language serve the writer’s pre-existing plan and to independently create or control the expressive elements that are claimed.

Human Authorship, No Human Authorship, and Contractual Protection Compared

Authors often confuse three different questions: whether copyright protection exists, whether a particular publisher will use the work, and whether a contractual promise can stop someone else from distributing it. These are not equivalent. A contract can allocate rights and prohibit unauthorized use even if a court later finds that particular AI-generated passages lack copyright, but a contract generally cannot create statutory authorship where the law recognizes none. A human-created compilation or edited version may be protected while an embedded AI image remains unprotected. That mixed result can be workable commercially, but it should be understood before publication rather than discovered during a dispute.

The table illustrates why “I own it because I paid for the subscription” is not a sound conclusion. Payment proves a transaction, not authorship. Similarly, saying that a work is worthless if an image is unprotected is too absolute. A publisher might decline it, but some uses of an unprotectable image may be lawful under circumstances where a protected photograph or drawing would require permission. The appropriate remedy can be replacement, contractual restriction, human redesign, or refusal to publish. Deciding among those options early is usually cheaper than litigation later.

How Should You Document the Human Creation Process?

Maintain a creation file that identifies the human decisions behind the expressive elements. Begin with dated outlines, character or story decisions, reference collections, and sketches that predate the model-generated material. Preserve the prompts used, but mark which outputs were rejected and which merely supplied factual or technical assistance. Save the first machine-generated draft, the author’s annotated copy, and successive revisions. These records help distinguish a genuine creative process from a convenient account created after a registration question arises. They also help another editor, agent, publisher, or insurer understand the workflow.

The level of documentation should reflect the extent of AI involvement. A work that uses a spell-checking program needs a different record from an edition whose illustrations were generated entirely by a model. For heavily assisted work, retain the original photograph used as a reference, the author’s cropping and sequencing plan, and the final image layer created or edited by the person. For a book that combines several copyrighted and AI-generated illustrations, number each asset and label its source. Avoid vague statements such as “all artwork produced with the creative team”; the statement conflates responsibility without showing authorship. More specific records also make it easier to remove unprotected material while keeping the rest of the book eligible for registration.

Documentation does not retroactively transform AI output into human authorship. It can only demonstrate the human contribution that existed, and the applicant must still provide the minimum required identifying information. If a disputed image was only lightly adjusted with an automated filter, calling the filter a “manual edit” is not enough. A professional should compare the before-and-after files and describe the human choices involved. When the contribution is difficult to explain, the more credible option may be replacement with a human-created asset. That decision protects the stronger part of the publication rather than making an overstated ownership claim.

What Should Publishers, Authors, and Platform Writers Do?

Before accepting a manuscript or visual package, ask how much of the final expression came from each source. Require the contributor to identify AI-generated photographs, illustrations, translations, transcripts, and prose. A disclosure form should ask for the model or service, the intended role of the tool, the type of human revisions, and whether trained on the contributor’s private material. Contracts should allocate responsibility for the disclosure and require corrective work when a protected submission is not actually human-authored. Merely including AI in a general prohibition is less useful than defining a process for exceptions, such as accessibility transcription or internal research.

The right contractual wording depends on the legal position. If a publisher wants a strong warranty of copyright, it can require a defined minimum level of human authorship and delivery of the creation file. If the parties know that some material is unprotected, they can expressly limit the warranty to the declared human-created components, address private labels and confidentiality, and set a replacement procedure. A platform may be able to contractually restrict scraping, style imitation, model training on submitted files, or removal of attribution, subject to applicable law. No clause should promise that the provider will never receive similar material, because enforcement and cross-platform tracing can be difficult. These promises often sound stronger than they are.

An “AI” clause is not automatically invalid or automatically suitable. Courts have interpreted contracts by their text, context, and governing law, and a court may examine whether the actual workflow violated an agreed authoring obligation. Terms that say the author may not use any external writing tool may unintentionally interfere with editing, research, or accessibility assistance. A better clause identifies prohibited uses while permitting ordinary tools under the publisher’s disclosure rules. The agreement should also state who controls edits after delivery and whether the human editor’s work creates a separate copyrightable contribution. Early negotiation reduces the chance of a dispute over who was entitled to change an AI-generated passage into protected prose.

Common Copyright Mistakes With AI-Assisted Publishing

A frequent mistake is registering an entire package while ignoring machine-generated components. The applicant may assume the registration certificate covers every image or paragraph, although the examiner can question each expressive element. Another mistake is using multiple authors as a way to dilute responsibility; adding a name to a prompt does not make that person an author. Some creators add company names, freelancers, editors, or platforms to registration records even though those parties did not make the claimed creative choices. Parties can license and own rights they hold through assignment, but the application itself must accurately identify authorship of the work.

Do not treat polished wording, payment, publication, or extensive revision after generation as automatic protection. A person who spends hours correcting every word, image, and transition has a stronger human contribution than someone who merely reads the result, yet a reviewer may still find some elements the direct product of the model. The U.S. Copyright Office has said that naming the predominant tool or including a few human-authored elements generally does not permit a “thin” claim over an otherwise AI-generated work. The proper response is to identify the human contribution accurately rather than conceal its origin.

Several non-U.S. systems deserve separate advice. Indonesia’s approach appears stricter, and other countries recognize categories of computer-generated works under conditions that do not fit the U.S. authorship rule. A publisher distributing internationally may need separate information, particularly for a country with mandatory human-author rules. A U.S. registration certificate also does not become a worldwide ownership ruling. A small indie publication can reduce risk by checking markets individually, adopting a conservative cross-border policy, and seeking jurisdiction-specific advice when a deal is large enough to justify it. Treating international law as one uniform global copyright law is both inaccurate and commercially dangerous.

When Should You Act, and What Will It Cost?

Act before a pitch, contract, launch, or registration filing—not after a platform sends a takedown notice. Publishers should introduce the requirement when accepting submissions, and authors should begin records at the first AI-assisted stage of a project. A short creation log is usually enough to start, while a mixed visual project benefits from asset-level records. If a book uses dozens of AI images, reviewing them before layout is easier than rebuilding the book after a claim or an examiner’s question. Small authors can take conservative steps without an attorney: use AI for brainstorming, disclose the workflow, retain drafts, and replace material they cannot explain. Larger organizations may need an editorial policy, vendor review, staff training, and contract updates.

Professional fees depend heavily on the market and the complexity of the work. U.S. lawyers may charge roughly $250 to $600 per hour for copyright and publishing advice, while an experienced specialist may charge more; a complex cross-border transaction can be much costlier. A targeted authorship review of a short article may take less work than auditing an illustrated book and its platform agreements. The U.S. Copyright Office’s single-author registration filing fee is generally $45 when filed online, with other categories and group works carrying different fees, so verify the current schedule at filing time. Registration is affordable compared with litigation, but the fee cannot create rights that the underlying work does not have.

Damages introduce a strong reason to document authorship. Under U.S. law, an eligible work may receive statutory damages of up to $30,000 per work, or an amount the court finds reasonable relative to the actual damages, plus costs and potentially attorney’s fees, depending on the case. An unprotected component may therefore be used without the automatic consequence of a copyright damages award, while the surrounding human work may remain fully enforceable. Do not wait for a dispute to determine which components are valuable. The best result is usually a process that combines visible human control with reliable records, rather than a marketing claim that the technology was “human-like.”

A legal disclaimer should be clear. The information here is general and cannot replace jurisdiction-specific advice from a qualified attorney or a detailed review by an experienced copyright examiner. It is especially important to obtain advice before distributing works in multiple countries or entering a contract that assigns exclusivity based on uncertain AI-generated elements. The durable principle is straightforward: a publisher can use AI, but the human author’s original choices must remain visible, documented, and legally distinct from the tool’s output.