Who Owns an AI-Assisted Manuscript in the EU?

Every conversation about AI and manuscripts reaches the question of who owns the book. It is the wrong place to start. Under EU copyright, a text that a model produced on its own is not a protected work, so there is no owner to argue about. The question that decides your position is narrower: if a publisher asks, or a dispute arrives, what can you show that proves the book is yours? This article is an explanation, not legal advice. A contract or a publisher's policy can require more than the law does.

What has to be true before ownership is even possible?

EU copyright protects a work that is original, and originality has a specific meaning in EU law. In Infopaq, in 2009, the Court of Justice generalised a test that had been written for photographs, databases and software: the subject matter must be the author's own intellectual creation. Later judgments sharpened it. In Cofemel the Court held that a subject matter is original when it reflects the personality of its author, as an expression of his free and creative choices. A 2020 study commissioned by the European Commission on AI and intellectual property quotes that sentence and builds its analysis on it.

Originality is not effort, and it is not quality. A manuscript typed over ten years without a choice of its own is no more protected than one typed in a weekend. What the test looks for is a chain of decisions that could have gone otherwise, and that can be seen in the text.

The second requirement is human authorship. The same study puts it directly: a production wholly generated by an AI system, without any human intellectual effort, is excluded from copyright protection. It also says the requirement does not rule out works made with the aid of a machine, as long as the human contribution meets the originality standard. The comparison used in the case law is a camera. A photograph is protected even though the camera did most of the optics, because the person chose the framing, the moment and the light.

There is no EU rule written specifically for AI output. A European Parliament research briefing from December 2025 says so plainly, and adds that the case law of the Court of Justice and the practice of member states show a strong need for human creativity. So this is not a new statute but the old test applied to new facts. Anyone who tells you the matter is settled is guessing.

One detail matters for attribution. Economic rights are harmonised across the EU, but moral rights are not, as the Commission's copyright page notes, so the right to be named as author is shaped by national law. Keeping evidence matters more when the rule is not the same everywhere.

What happens to a manuscript the model wrote alone?

It falls outside protection. That has two consequences, and the second is the one writers underestimate. The first is that you cannot stop someone else from using a text that is not a work. The second is that the provider's promise to give you the output transfers nothing of value, because there is nothing to transfer. A purely generated chapter is not a badly protected asset. It is not an asset.

The study is blunt about the user who only prompts. Where the user's role is reduced to an initial prompt, or to pushing buttons, the user is not the author, in the same way that a person playing a video game does not own the footage. That is the case to avoid landing in, and it is avoided by what happens after the model answers, not by the prompt's wording.

The same analysis has one hopeful sentence. Even an unexpected, quasi-random output might still become a protected work at the redaction stage. The law does not ask where the sentences came from. It asks whether the published text expresses choices that are yours. Redaction is where those choices become visible, and it is also where they are easiest to document.

Which choices count, and how do you show them?

The study uses a three-phase model: conception, execution and redaction. It is a description, not a legal rule, but it is a useful map. Conception is what the book is: the premise, the structure, the voice. Execution is the scene-by-scene work, including what you ask the model to do and what you reject. Redaction is the pass that turns a draft into the text you publish: cuts, reordering, rewrites, corrections of fact and tone.

A court or a publisher will not read your chat log for atmosphere. They will look for identifiable decisions at one of those stages, reflected in the final text. That is why the record has to show choices and not hours. A folder of drafts proves you were busy. A diff proves you decided.

What should a writer keep?

• The raw output, exactly as the model returned it, before you edit. Without it, the diff has nothing to compare against.
• Prompts with the date, the model and the version. A prompt is a decision about direction even when the output is discarded.
• Drafts at each stage, under version control. Plain text in a git repository works better than most writing software here, because every commit is a dated snapshot and every change is a diff.
• The diff between what the model returned and what you published. This is the single most useful artefact you can hold. It is the shape of your contribution, in the order you made it.
• Dated exports at milestones, as PDF, so the text has a fixed form outside your working files.
• Editorial correspondence: the editor's notes, your replies, the version that was accepted.

Two honest caveats. A file's modification time is easy to change and proves little. A hash proves the file is the same one later, not when it existed; a trusted timestamp from a timestamping authority binds a hash to a time. For most writers, a repository with dated commits and a few exports is proportionate.

Keep this material where it survives. Chat histories are retained on the provider's terms, can be deleted, and are not designed as an archive. Export what matters on the day you make it.

Does the AI Act decide ownership?

No. Article 50 of Regulation (EU) 2024/1689, the AI Act, is a transparency provision, and the Commission describes its purpose as addressing the risks of deception and manipulation. It governs disclosure. It says nothing about who holds rights.

Two of its paragraphs matter to a writer. Article 50(2) puts a duty on providers of generative systems to mark synthetic output in a machine-readable format so that it is detectable as artificially generated. That duty sits with the company that ships the model, not with you. Article 50(4) puts a duty on deployers, the people using the system, to disclose AI-generated text published with the purpose of informing the public on matters of public interest, and to disclose deepfakes. The transparency obligations apply from 2 August 2026. The Commission has published a voluntary code of practice on marking and labelling, and a set of EU icons, but the obligations themselves are law.

For a novel, the public-interest limb is unlikely to apply, but the exception is worth reading. The disclosure duty for text falls away where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. The law treats reviewed, responsibly edited text differently from text pushed straight out of a model. That is the same distinction copyright makes, arrived at from a different direction, and it is another reason the record of your editing is worth keeping.

What do the provider's terms actually give you?

Both OpenAI and Anthropic now say you own the output. OpenAI's Terms of Use state that, as between you and OpenAI and to the extent permitted by applicable law, you own the Output, and that OpenAI assigns to you all its right, title and interest, if any, in it. Anthropic's Commercial Terms say the customer owns its outputs and that Anthropic disclaims any rights it receives in the customer content; those terms are dated 17 June 2025.

Read the qualifiers. To the extent permitted by applicable law. If any. A contract can move rights that exist. It cannot create a copyright in a text that is not a work. So the terms settle whether the provider will claim something from you. They do not settle whether there is anything to claim. The same terms add that output may not be unique, a fair warning about treating the text alone as an asset.

Check which terms apply to you. Consumer, business and API terms differ, and they change. A clause you read last year may not govern the account you use today.

Which contract binds you day to day?

For most writers the operative document is not the model's terms. It is the publishing contract, the commissioning agreement, the client's statement of work, or the platform's upload policy. These can require you to disclose AI use, to warrant that no AI was used, to indemnify someone if a claim arrives, or to deliver a clean chain of title. None of that is copyright law. All of it binds you more tightly than copyright law does, because it is a promise you signed.

Read those clauses before the manuscript is finished, not after. If a contract asks for a record of how the text was produced, the routine below produces it as a by-product.

What routine makes the record almost free?

• Keep the manuscript in a git repository. Commit at the end of every working session, with a message saying what changed and why.
• Save every model response you use in a dated folder before editing it, including the ones you reject.
• Log the prompt and the model beside each saved response.
• Export a PDF at each draft milestone and keep it.
• Never edit the only copy. The raw version is the evidence; the edited version is the book.

This is also how we build writing tools. Vellum, our plugin for long-form fiction, runs the manuscript as files on disk with an outline, gates and a state card, so the drafting trail exists without anyone maintaining it by hand. The principle is older than AI: the work you can describe is the work you can defend.

What changes if you take this seriously?

You stop treating the model's output as the product and start treating it as material. Models are good at producing a plausible passage and bad at knowing which passage the book needs. The choices that make a manuscript yours are the ones you make against the model's first answer: what to keep, what to cut, what to contradict, what to rewrite until the sentence sounds like you. Those choices are also, conveniently, the ones a record can show.

If you write with a model and publish the result, copyright law will not ask you to prove that no machine was involved. It will ask what you contributed. Keep the answer where you can find it. For the wider compliance picture, our guide to the AI Act for small businesses covers the obligations that apply to a company rather than a manuscript.