Copyright and AI-generated text

You draft your own chapters, but you use a model for brainstorming, for edits, maybe for the occasional generated passage. At some point the craft question becomes a legal one: what exactly do you own? The short answer in the United States is that copyright protects human authorship, your own writing stays protectable even when a tool assisted you, and purely machine-generated text is not protectable at all. This guide walks through what the US Copyright Office has actually published on the question. It is not legal advice; for a real registration decision, read the Office's guidance or ask a lawyer.

Human authorship is the requirement

The foundation predates AI. The Copyright Office's registration guidance for works containing AI-generated material states it flatly: "it is well-established that copyright can protect only material that is the product of human creativity." Works put forward as autonomously machine-generated, with no human author, have accordingly been refused registration, and US courts have upheld those refusals.

Prompts alone are not authorship

What about prompting? The Office's January 2025 report on copyrightability addresses this directly: "Based on the functioning of current generally available technology, prompts do not alone provide sufficient control." Prompts influence the output; they do not determine its expressive details. So text generated wholly from your prompts and posted unedited is not protected, however elaborate the prompting was. The report is explicit that all of this is assessed case by case and tied to how current technology functions, which is a standing warning against treating any of it as a bright-line rule.

Assistance does not defeat your copyright

The load-bearing sentence for working writers, from the same report: the use of AI tools "to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output." And, in the Office's words, "Copyright protects the original expression in a work created by a human author, even if the work also includes AI-generated material."

In practice: a novel you wrote, edited with a model's suggestions, brainstormed with its questions, checked by it for contradictions, is still your novel for copyright purposes. The assistance is not held against you. The line is what the machine generated versus what you authored, not whether a model was in the room.

Mixed works: what is protected and what is not

A book can contain both. In a mixed work, protection extends to:

• The expression you authored that is perceptible in the final work
• Your creative selection and arrangement of generated material
• Your creative modifications of generated material, where the modifications are substantial enough to meet the originality standard on their own

The clean illustration is the Office's Zarya of the Dawn decision from February 2023: a graphic novel whose author wrote the text and selected and arranged AI-generated images. The registration covered the human-authored text and the selection and arrangement of the images; the images themselves, generated by the model, were refused.

If you register: the disclosure duty

When you submit a work for US registration that contains more than a de minimis amount of AI-generated material, the Office expects you to say so. The 2023 registration guidance puts "a duty to disclose the inclusion of AI-generated content in a work submitted for registration" on applicants, along with "a brief explanation of the human author's contributions," and the AI-generated portions themselves are excluded from the claim. Registration is a legal process with a paper trail; misstating the human contribution there is a far worse position than disclosure ever is.

Habits that keep the question small

• Keep the prose yours where it counts. Treat any generated passage as raw material to edit, not text to publish. The edit is where authorship lives, and the Office's own standard (perceptible human expression, human modification) rewards exactly that work.
• Keep your drafts. A trail of versions showing your choices documents the human contribution, and doubles as ordinary good practice. The same working records that feed a story bible are evidence of decisions you made.
• Decide your disclosure posture early. Knowing roughly what share of the book is generated makes the registration question a paragraph, not an archaeology project.
• Separate the software question from the legal one. What a tool's terms say about output, and whether you trust it with your manuscript, is the software decision; what copyright protects is decided by law, not by either set of terms.

What is still unsettled

Training data and fair use remain under active litigation, and the Office's report on that part was still pre-publication when this guide was written; authorship questions are also being litigated in cases the Office does not control. Treat the rules above as the Office's current guidance, not as a settled area of law.