The argument about AI and writing is usually conducted between two people who are not discussing the same thing. One is talking about a machine that produces prose on request. The other is talking about a machine that reads a manuscript and reports where it drags. Those are different products with different ethics, and collapsing them makes the whole conversation useless.
Here is what the objections actually are, which of them hold, and where the line sits.
The training data objection is the strong one
Books were used to train these systems without the consent of the people who wrote them and without payment. That is the objection, and it is not answered by pointing out that human writers also read widely.
The comparison fails on scale and on directness. A novelist who reads three hundred books absorbs them slowly, through a life, and produces work no one can trace back to a specific source. A training run ingests those books mechanically, at a scale no reader could reach, to build a commercial product that competes in the same market. “Humans learn from other humans” is true and does not carry the weight put on it.
Some of this may eventually be settled through licensing, and some through courts. Neither has happened yet. In the meantime a writer using any of these tools is using something built on an unresolved question, and the honest position is to say so rather than to explain the discomfort away.
The displacement worry is real, and aimed at the wrong target
“Will AI replace novelists” is the version that gets argued and the least interesting one. Novels are bought for a voice, and the market for a specific voice is not easily substituted.
The work that goes first is the work that paid for the novels. Product descriptions. Corporate blog posts. Low-rate content writing, ghostwriting at the cheap end, the SEO articles that a lot of working writers used to treat as rent money while the real book got written. That layer is being hollowed out now, and it was load-bearing for a lot of careers.
That is a genuine harm and it does not become less genuine because the tools are also useful. Anyone who claims otherwise is not paying attention to how writers actually pay rent.
Illustrators got there first, and that is why the feeling runs so hot
The strongest anti-AI sentiment in creative communities is not abstract. It came from watching something specific happen to a specific profession, faster than anyone expected, with the visual evidence right there.
Writers who dismiss that reaction as hysteria are missing that it is evidence. It is the closest available data on what happens when a generative tool arrives in a creative market. Treating it as a data point rather than an overreaction is the minimum intellectual honesty the topic requires.
The homogenisation worry, and its decent counter-argument
A model trained on everything suggests the centre of everything. Ask for a sentence and you get the most probable sentence, which is the one nearest to what has been written many times already. Scale that across enough writers and the centre gets heavier, and the next model trains on the result.
The counter-argument is better than it usually gets credit for. Genre fiction has always run on convention. Readers frequently want the familiar beat delivered well, not a subversion of it. Whole traditions are built on repetition with variation, and “homogenised” is occasionally just a snobbish word for “popular”.
Both are true. The distinction that survives: convention chosen is a craft decision, convention defaulted into is drift. A writer who knows the beat and hits it deliberately is doing something a model suggesting the same beat is not doing.
The arguments in favour that are actually good
Two of them, and they are strong enough that dismissing them is its own kind of dishonesty.
Accessibility. Writers with dyslexia, with chronic fatigue, with motor impairments, with aphasia after a stroke, use these tools to do work that was previously slow to the point of impossible. Any position that treats all AI assistance as cheating writes those people out of the profession, and usually without noticing.
Language. A writer working in their second or third language can now reach readers who were previously unreachable. That is not a small thing in a publishing world where fluency in English has functioned as an entry requirement for decades.
Where the line is, and where it genuinely blurs
The distinction worth defending: a tool that tells you where your pacing collapses is doing something different from a tool that writes your chapter. The first gives you information about work you made. The second makes the work.
Slima is built on that distinction, so treat this as an interested party arguing its own position rather than a neutral finding.
The line blurs, and pretending otherwise would be dishonest. Line editing sits right on it. If a tool rewrites your sentence for rhythm, is that criticism or composition? Reasonable people put that in different places. So does the case where a critique is so specific it amounts to dictation: “this paragraph needs a beat where she notices the door” is feedback, and it is also, nearly, the sentence.
What does not blur: generating prose you did not write and presenting it as prose you wrote is a different act from getting told your middle sags. Everything else is a boundary dispute; that one is a category difference.
The practical part: disclosure
Platforms have begun requiring writers to declare AI use, usually with categories along the lines of AI-generated, AI-assisted, and AI-not-used. Several serialisation platforms now gate publishing on the declaration.
The useful way to think about it: the categories are asking who composed the sentences a reader will read. If the sentences are yours and a tool told you which ones were slack, that is a different answer from a tool having produced them. If you genuinely cannot tell which applies to your manuscript, that uncertainty is itself informative about how you have been working.
Read the specific platform’s definitions rather than assuming they match another platform’s. They do not. And declare accurately, which is a separate matter from declaring whatever seems safest.
What this leaves you with
If you use these tools, know what you are using and be able to say which part of the book is yours. If you refuse them, refuse for a reason you can state, which is a stronger position than a general unease.
The question that survives all of this is not whether a tool touched the manuscript. It is whether the decisions that make the book a book — what the character wants, what it is about, the specific strangeness of how you put sentences together — were made by a person.