Using AI without losing authorship

The argument about AI and fiction is stuck at the wrong altitude. One side says using it makes you a fraud, the other says it is just a tool like a spellchecker, and neither position helps the person who has an actual manuscript open and a genuine question about whether the thing they just did was fine.

Because the honest answer is that "using AI" describes at least four different activities that cost you completely different amounts. Asking a machine what a character said about her father in chapter nine is not the same act as asking it to write chapter ten, and treating them as one category is why the conversation goes nowhere. The useful question is narrower and answerable: for this particular request, what judgement am I handing over, and can I check the result?

This article is published by a company that ships an AI feature, which you should factor in. I have tried to write the version I would want to read before deciding, including the parts that argue against reaching for these tools at all.

The law has drawn a line, and it is not where writers assume

In January 2025 the US Copyright Office published the second part of its report on copyright and artificial intelligence, dealing specifically with copyrightability. Two of its conclusions matter enormously to novelists and are widely misunderstood.

The first: prompts are not authorship. Not even long ones. Not even very good ones. The Office's position is that however detailed and effortful a prompt is, it functions as an instruction rather than as expression, and does not by itself give you copyright in what comes back. Writers who assumed that a sufficiently elaborate prompt made the output theirs were wrong, and it is better to learn that now than at the point of registration.

The second, and this is the part the alarmed coverage tends to skip: using AI as an assistive tool does not poison the work. Where a piece contains both human and machine-generated material, the human contribution remains protectable, and the Office is explicit that using these systems for ideation or to help edit does not render the whole thing uncopyrightable. It is assessed case by case, on what you actually contributed.

Prompting is instruction. Writing is expression. The law can already tell them apart.

Put those together and you get a rule of thumb that happens to be a craft rule as well as a legal one. The protectable part of your novel is the part you expressed. Anything you merely requested belongs to nobody. If a chapter exists because you asked for it, you have not written a chapter; you have commissioned one, and badly, from something with no stake in it.

Four jobs, four very different prices

Sort any request you might make into one of these before you make it. The sorting takes two seconds and settles most of the anxiety.

1. Retrieval. "What did Marianne say about her father?" You are asking a question with a right answer that exists in your own text. No judgement transferred, and the answer is checkable in seconds.
2. Analysis. "Where does this chapter's tension drop?" A machine is applying a pattern to your prose. Some judgement transferred, but the output is a claim you can go and test against the page.
3. Ideation. "Give me five reasons she might refuse." Judgement stays with you, because you still have to recognise which of the five is any good, and that recognition is the actual skill.
4. Generation. "Write the scene." This is the one that costs something, and pretending otherwise is where writers get into trouble with themselves.

The first three are, to my mind, uncontroversial. They are versions of things writers have always done with concordances, readers, index cards and long walks. The fourth is a different act, and it deserves its own section rather than a shrug.

Retrieval is the safe end because it can be checked

The reason retrieval sits at the safe end is not that it is unambitious. It is that it produces claims about a text you own, which means you can verify every one of them without leaving your desk.

This is also the test that separates a useful tool from a confident liar. A system that answers "Marianne told Edward her father drank" is useless unless it can also tell you it read page 214, so you can go to page 214 and find out whether that is what the passage says or a plausible-sounding invention. Grounded answers are not a nicety here; without them you have added a research assistant who never admits to guessing.

A citation is not proof the conclusion is right. It is somewhere concrete to check, which is a much lower bar and a much more useful one. When Muse answers from your own pages it shows the passage it used for exactly this reason: not so you will trust it, but so you can catch it. Any tool making claims about your manuscript should be held to that standard, ours included, and if it cannot show its working the correct response is to stop using it for anything that matters.

Generated prose costs you the thing you were building

Here is the argument against generation that I find hardest to answer, and it is not the legal one or the ethical one. It is that writing prose is how you find out what the book is.

Anyone who has drafted knows this. You plan a scene where she refuses him, you write it, and somewhere in the middle of the third paragraph you discover she does not refuse, she equivocates, and that equivocation is the most truthful thing in the chapter. That did not come from the plan. It came from the friction of putting one sentence after another with a specific person in your head. A generated version of that scene is smooth precisely because nothing was discovered in the making of it, and smooth is what you do not want.

Voice compounds the problem. A voice is a thousand small consistent choices, most of which you cannot articulate. It is the reason a reader can identify a paragraph as yours. A model has read a great deal of published fiction and will hand you the average of it, which reads perfectly well and belongs to no one. The failure is not that generated prose is bad. It is that it is generic, and generic is invisible in a paragraph and fatal across a novel.

None of which is a prohibition. Plenty of working writers use generation for things they have no ambition about, and a synopsis, a blurb, or the dull connective sentence between two scenes is not where anyone's voice lives. The discipline is simply to know which you are doing, and never to let the generated thing be the part the book is actually about.

Ask what leaves your machine, one feature at a time

The most common mistake in this whole area is treating an application as though it has a single privacy boundary. It almost never does. A writing app can run its spellchecker on your device, its analysis on your device, and send a chapter to a third-party model the moment you use one particular button, and nothing about the interface will tell you which is which.

So the question is never "is this app private". It is per feature, and there are only four things worth asking:

• Which features send anything at all, and which run locally? Vague marketing language about privacy is not an answer to this.
• How much goes when they do: the sentence, the chapter, the whole book, the story world?
• Does it happen on purpose? A request you deliberately made is a different thing from a background process shipping your draft somewhere while you type.
• What happens to it at the other end: retained, logged, used for training?

It is worth saying plainly that a local-first app is not automatically clean here. Ours works offline for writing, proofreading and export, and the AI features are the deliberate exception: they are requests you make, they cost credits so you can see them happening, and what each one sends is documented. That is the arrangement I would want, but do not take my word for the boundary of a tool you are trusting with a novel. Watch the network traffic, or ask a direct question and accept nothing but a direct answer. This is the same ownership question as backups, wearing different clothes.

Nothing should enter the book without you putting it there

The single most important property of an AI feature in a writing tool has nothing to do with model quality. It is whether the thing can change your manuscript without an explicit act from you.

A system that proposes forty character records and waits is fundamentally different from one that creates forty character records and tells you afterwards, even if the proposals are identical. The first leaves you the author of your own story world. The second makes you the reviewer of somebody else's, and reviewing is a much weaker form of attention: you will approve things you would never have written, because rejecting requires more energy than accepting and there are forty of them.

This is why the boundary should be visible and hard rather than a setting. Anything that touches the prose, the records or the structure should arrive as something you accept, one at a time, with the option to accept none of it. If you cannot tell from the interface where that line falls, that is itself the answer about how carefully it was drawn.

Disclosure, and who is actually asking

Writers worry about this in the abstract and then discover the practical version is quite specific. Publishers, agents, competitions and retailers have their own requirements, they differ, they change, and they are the ones that will actually affect you. Amazon's KDP asks about AI-generated content at upload. Many competitions now ask outright. Some publishing contracts contain warranties on it. Read the form in front of you rather than reasoning from first principles.

The Authors Guild has moved in the other direction, offering a "Human Authored" certification for writers who want to make a positive claim rather than a disclosure. Whether that matters to you is a commercial and personal judgement, not a moral one, and I would not tell anyone which way to go.

What I would say is that the standard worth holding is your own, and it is easier to hold if you decided it before you were tired and behind on a deadline. Write down now what you are and are not willing to do. The decision made calmly in August is worth more than the one you will make at midnight in February.

What to do on Monday

Take the last three AI requests you made about your novel and sort them into the four jobs. If they were all retrieval and analysis, you have nothing to worry about and can stop reading things like this. If any were generation, ask what you were stuck on at that moment, and whether the stuckness was telling you something about the scene.

Then find out, concretely, what your tools send and when: not from the marketing page but from the documentation, or by watching. And check that nothing in your workflow can alter the manuscript without you pressing something. Those two facts, plus a clear head about which of the four jobs you are asking for, cover essentially everything that actually matters here. The rest is an argument on the internet, and it will still be going on long after you have finished the book.