When a publisher drops an author for using AI, what must the contract say?

What happened

In early 2025, Swedish publisher Natur & Kultur announced it had cancelled the contract for author Thélyson Orélien's novel Guds vrede (The Wrath of God) after discovering he had used AI to write it ^[1]. The publisher said the decision followed internal discussions and that it had happened in agreement with the author. Orélien had already published several novels with the same publisher.

The news triggered a wider discussion about whether authors must declare AI use, and what the legal rules actually say.

The disclosure duty in Article 50.4

Article 50.4 of the AI Act says that users who generate or manipulate text with AI and publish it to inform the public on matters of public interest must disclose that the content is AI-generated or AI-manipulated ^[2]. There is an exception: the duty does not apply if a human has reviewed the content and a natural or legal person has editorial responsibility for it.

That duty has applied since 2 August 2026. It covers news, reports and opinion pieces intended to inform the public. It does not cover fiction, because fiction is explicitly an "artistic, creative, satirical, fictional or analogous work" under the same article, and for such works the only requirement is that it be made clear "in an appropriate manner" that AI content exists, in a way that does not hamper the display or enjoyment of the work ^[2].

Think of it like allergen labelling on a menu: you have to say nuts are in the dish, but you don't have to shout it on a sign above the table. For a novel, that appropriate disclosure might be a line in the author's note or on the copyright page. The AI Act does not say what counts as appropriate for fiction, only that it must not spoil the reading experience.

So for a novel like Guds vrede, Article 50.4 does not require the author to tell readers that AI was used, as long as there is some appropriate mention that AI content is present. Whether that mention is sufficient is a judgement call.

What the contract can require

The AI Act does not stop a publisher from asking authors to declare AI use in the contract. Many publishing contracts now include a clause requiring the author to state whether AI tools were used and, if so, how. That is a matter of contract law, not AI regulation.

If the contract says "you must tell us if you use AI", and the author does not, then the author has breached the contract. The publisher can terminate the agreement and, depending on the terms, reclaim any advance already paid. The author also risks being dropped by the publisher for future projects, which in a small market like Sweden's can be damaging.

What the AI Act does is set a floor: certain uses of AI must be disclosed to the public. Contracts can add requirements on top of that.

Machine-readable marking: the provider's job

Article 50.2 of the AI Act requires providers of general-purpose AI models with systemic risk to ensure that outputs are marked in a machine-readable format as artificially generated or manipulated, and detectable as such ^[2]. That duty lies with the model provider, not with the author or publisher using the model. In practice, it means that if you generate text with a large model, the output should carry a hidden mark that tools can read.

That marking is meant to help platforms, fact-checkers and content moderators spot AI-generated material at scale. It does not replace the disclosure duty in Article 50.4, which is aimed at the human reader. One is a label for machines, the other is a notice for people.

Why it matters for publishers and authors

Publishers care about two things: legal risk and reputation. The legal risk under the AI Act is small for fiction, because fiction does not have to be disclosed in the same way as journalism. The reputational risk is larger. Readers who feel they have been sold a book written by AI when they expected a human author may not buy the next one, and they tend to say so loudly online.

For authors, the risk is contractual. If the publishing agreement says you must declare AI use and you do not, you have broken the contract. In Orélien's case, the publisher's decision to cancel the contract suggests either that such a clause existed or that the relationship had broken down for other reasons.

There is also a question of craft. Many authors use AI tools for research, editing suggestions or to try out ideas. Using AI to write entire chapters or books is seen differently, and that perception matters in an industry built on voice and authorship. Whether that distinction makes sense is another conversation, but for now it drives how contracts are written and how publishers react.

What this means for you

If you are a publisher:

• Add a clause to your standard author contract requiring the author to declare any AI use, and to say how it was used.
• Decide what your policy is: some uses allowed, others not, or declaration only.
• Make it clear to authors what the consequences of non-disclosure are: contract termination, repayment of advances, or loss of future deals.

If you are an author:

• Read your contract. If it asks about AI, answer truthfully. A lie can cost you more than an honest answer.
• If you do use AI, keep a record of how: brainstorming, editing, outlining, drafting. Courts and arbitrators like specifics.
• Remember that Article 50.4 does not require you to tell readers you used AI if you are writing fiction, but your publisher may still want to know, and your readers may care.

A record of what you generated

Here is where a digital witness fits, for authors and publishers who want a clear record.

Say you draft a book with the help of an AI model. At each session, you seal a note of what you asked the model, what it generated and which parts you kept. That note stays with you. What gets sealed is a fingerprint: a short string that changes completely if a single word in the note changes.

Months later, if a dispute arises about what the AI produced and what you wrote yourself, you can share that record. Anyone can check that it matches the sealed fingerprint, so it has not been edited after the fact. It doesn't prove your work was original or lawful, but it does show what you generated, when, and that the record has not been altered since.

The technical side is simple. You send the fingerprint to WitniumChain through its API, SDK or MCP server, and get a witness ID back. To check later, you calculate the fingerprint again and see if it matches. The content itself never leaves your systems. The first 2,000 witnesses a month are free, and the documentation is short.

For publishers, the same approach works for contracts, approval emails and manuscript versions. Seal each stage as it happens, and later you can show which version was agreed, without anyone having to trust your word.

What to watch

The AI Office has not yet issued detailed guidance on what counts as an "appropriate manner" of disclosure for creative works under Article 50.4. That guidance, when it comes, will shape how publishers and authors handle the disclosure duty for fiction.

Meanwhile, the industry is writing its own rules through contracts. If you are negotiating a publishing deal in 2025, expect a clause about AI. Read it carefully.