Does AI writing belong to you? Copyright for AI-assisted novels

How copyrights apply to work that has been modified or produced with AI is a topic there's a lot of misinformation about. Some people say such works can't be copyrighted at all, but that's not true. In fact, something you wrote with AI can be copyrighted in many circumstances.

Copyright attaches to what a human contributed: the sentences you wrote, the machine drafts you rewrote into your own words, the shape and order you decided on. What it does not cover is the material no person authored — a chapter generated from a prompt and kept exactly as it came back, a book produced whole from a premise and pushed onto a storefront unread.

This is a far narrower exclusion than most people realize. This includes anyone who used AI to edit their book, revise their book, suggest prose that's edited into the book, or anyone who substantively made editorial or directive choices about what was written. It doesn't include... people who generated AI slop and threw it online. Which is a good thing.

Caveat: This is a summary rather than legal advice, and it describes the United States, where the guidance has developed furthest, as of July 2026. Copyright law differs by country. If a publishing contract or real money is at stake, you want a publishing attorney rather than an article.

The rule protects human authorship

In March 2026 the Supreme Court refused to hear Thaler v. Perlmutter, in which an inventor sought to register a picture his AI system had made with no human contribution at all, neither prompted nor afterwards touched. No human authorship, no copyright. Where the line falls above that floor is the part still in motion, and it is what the rest of this is about, along with the question almost nobody asks until it is urgent: what you could prove.

The Copyright Office has been consistent about the principle throughout, and Thaler changed it less than it confirmed that nothing was going to. Copyright protects works of human authorship. Material generated entirely by a machine, with no human creative contribution whatever, is not copyrightable however good it may be.

The requirement is for some human authorship, not for hand-typed authorship. The Office's guidance and its 2025 report on AI drew the line at what a human contributed rather than at whether AI was involved at any point, which the headlines flatten into AI work cannot be copyrighted. That is not what the Office said.

Prompts alone are generally insufficient. Typing "write a chapter in which the detective confronts her partner" and keeping whatever returns does not make you the author of that chapter in the Office's view, however many times you regenerate it.

Your own writing is protected without qualification. Everything you wrote, and the portions of machine drafts you substantively rewrote in your own words, belongs to you.

Selection, arrangement and creative editing count. A novel in which a human devised the story, structured the whole, chose among outputs, arranged them and creatively revised them may be protected as a work, with purely machine-generated passages potentially carved out of the claim.

The well-known early test case concerned a comic whose text and page arrangement were registered while its AI-generated images, taken individually, were not. Registrations since have followed that pattern. The human contribution is protected and the raw machine output is not.

The floor and the threshold are different questions

Two things get blurred together here, and separating them helps.

The floor, meaning no human authorship and therefore no copyright, is the part the Supreme Court has now left firmly in place, and there is little reason to expect it to shift. The Constitution's copyright clause protects "authors," the courts have read that word to mean humans for over a century, and nobody is lobbying for machine authorship.

The threshold, meaning how much human contribution suffices, is where the law is still moving. The Copyright Office says these determinations are made case by case, and it has recognized several routes to protection: significant creative modification of AI output, creative selection and arrangement, AI confined to assistive roles, and human-authored material that remains perceptible in the finished work.

If you are attempting to guess where the threshold travels from here, the pressure runs toward generosity rather than strictness. AI assistance is becoming an ordinary part of how professional writers, filmmakers and studios work, and those are the same studios and publishers whose entire business model depends upon owning what they release. A legal regime that consigned most professionally produced work to the public domain by default would be intolerable to every powerful interest in the creative economy.

Nobody can promise you an outcome. But a seriously directed, human-revised, AI-assisted novel sits on the safe side of every version of this rule that has ever been proposed.

The rule is less alarming than it sounds

Step back from the legal vocabulary and notice how little the rule excludes, because the anxiety surrounding this subject is largely misplaced.

Nobody sits down wanting to generate a random book. Writers have a story, this character and this ending and the thing they have been circling for years, and that is the entire reason they are at the desk. Watch what happens when such a writer works with AI. The model offers a scene and she says no, she wouldn't do that. It offers another and she says closer, but the ending gives itself away. She redirects, cuts, rewrites the middle, throws out the version that was technically fine and was not hers. Ten passes later the chapter is the one she meant.

That process — the wanting it a particular way, and the hundred small refusals that arrive at it — is authorship. It is what authorship has always been. AI changed only what the raw material looks like on the way in.

It is also what the Copyright Office protects. Its recognized routes read like a description of that session, which is to say creative selection among outputs, arrangement of the whole, significant creative modification, and human-authored material perceptible in the finished work. The line the law draws separates authored work from generated work, and it falls almost exactly where a working writer would draw it herself.

What sits on the wrong side is the thing the rule was aimed at. Books mass-produced from a prompt and pushed onto a storefront, dozens a month, which nobody directed and nobody revised because the point was volume. Those are unprotected, and no novelist has ever mourned the fact.

One wrinkle, since it cuts both ways. Iteration counts because it is directed, not because it is repeated. The Office was explicit that regenerating a prompt many times does not by itself make you the author. What earns protection is that each refusal was a judgment about your story, and that your decisions and your sentences ended in the book. "Keep rolling until something is good" is a weaker position than "keep working until it is mine." In practice, real writers are doing the second thing. It is why they are writing at all.

What this means at each level of AI use

In How to write a novel with AI we described a spectrum of five positions, and it maps onto copyright closely.

With AI as editor or as sounding board there is no serious question. The words are yours, and suggestion and critique transfer authorship no more than a human editor's do.

Where AI drafts and you substantively rewrite, you stand on solid ground, provided the rewriting is real. Your sentences, your rhythm, your choices, and not a synonym pass over machine prose.

Where AI drafts and you keep it largely as it arrived, you are in the gray zone. Your novel's structure, its plot and characters and arrangement, reflects your authorship even here, though the kept passages themselves may not be yours to claim.

Where AI writes the book from your premise, the work is mostly not protectable under current U.S. guidance, which among its other consequences means anyone could lawfully republish it.

One practical note if you intend to register. The Copyright Office requires applicants to disclose AI-generated material in works submitted for registration and to exclude it from the claim, and overclaiming can void a registration. Registering machine prose as your own without disclosing it is the one unambiguously bad move available here.

The part nobody discusses: proving it

Copyright disputes turn on what you can show happened rather than on what happened. Most writing arrangements show almost nothing.

Draft in a text editor or a cloud document, paste in machine output, revise it heavily across six months, arrive at a manuscript that is overwhelmingly your own authorship, and you may have no means of demonstrating any of it. The finished file is indistinguishable from one somebody generated in an afternoon. If your registration is ever challenged, if a publisher's contract warranty is ever questioned, if you must one day sue a pirate, the question becomes what evidence exists that a human did the work.

This is where the present legal moment rewards a particular kind of workflow. One that leaves a record.

How Proselon handles this

Proselon was built around the answer to that question before it became a legal question.

Your manuscript lives as plain files on your own computer, and Proselon keeps a complete version history of it automatically as you work, timestamped and continuous, reaching back to the day you began the book. Inside the application that history appears as a list of earlier drafts. Browse it, read any page as it stood at any moment, name the versions that matter, restore a page or the whole book. Nothing to configure, nothing to remember to do.

Underneath, it is not a proprietary format. It is the standard open version-control record the software industry has used for two decades to answer this exact kind of question about who changed what and when, sitting in your own book's folder on your own disk. That distinction matters on the day it counts, because you can hand it to a lawyer and her expert can read it with ordinary tools. No export request, no vendor cooperation, no subpoena to us.

What it preserves is the sequence of the work. Every co-writing session is bracketed, the book's state recorded the moment before the co-writer begins, and ordinary writing is snapshotted continuously as you go. Across a year that accumulates into something specific: the alternation of your hands and the co-writer's over hundreds of sessions, the shape of the book changing, passages entering and being reworked, the pace and pattern and timing of every change. An examiner reading that has the raw material with which to characterize how the book was made, which is what a finished file, however excellent, can never show. A manuscript generated in an afternoon does not look like this and cannot be made to.

That is the substance of the claim. A documented, independently readable history of human direction and revision, on your own disk, from the first day. For a registration you must defend, a contract warranty somebody questions, or an infringement suit you must bring, that is a materially stronger position than most writers can put their hands on.

Two things it is not, because reassurance is worth something only when it is true.

It is a record of changes rather than a per-sentence authorship label. It shows when every change occurred and brackets each co-writing session, and it does not tag individual sentences as yours or the co-writer's. Reading authorship out of it is analysis, the sort of thing an expert does with evidence, rather than a report the application hands you. If you register and must disclose AI-generated material, that determination remains yours to make.

And a record demonstrates what you did without upgrading what you did. A well-documented history of accepting machine prose unchanged is a well-documented gray zone rather than a loophole. It is evidence and not a certificate, and courts and the Copyright Office weigh evidence case by case.

Which is rather the point. The strongest position, legally and artistically, has turned out to be the same one. Remain the author. Make the structural decisions, write or rewrite what ends on the page, keep the AI in the roles where human authorship is never in doubt, as editor and sounding board and first-drafter of material you then make your own. Then keep the proof.

That is the workflow Proselon is built around. The co-writer proposes and you decide, scene by scene, every draft passing through your hands in a real manuscript that lives in your own files, with a complete and verifiable history behind it. It is free for 30 days if you would like to see it for yourself.