AI-Free vs AI-Generated vs AI-Coached: The Three Kinds of Novel Writing Software
An AI writing tool that doesn't write for you belongs to a third kind of novel software: not the tools that refuse any AI (Scrivener, Novlr, The Novel Factory), not the ones that generate prose on demand (Sudowrite, Squibler, Novelcrafter), but the ones where the AI reads, questions and diagnoses your text without ever producing a line of it. The market talks a lot about the first two and has no word yet for the third. This article sets the vocabulary, describes what each kind actually does with your manuscript, recalls what KDP asks you to disclose, and shows Sagoræ's four coaches at work on the same paragraph.
Why "with AI or without" is no longer enough
In 2026, the question you get asked when choosing a writing tool is binary: with AI or without? It is the wrong question, because it puts a grammar checker that underlines a subject-verb error in the same box as a button that drafts three pages in your place. Those two things do not do the same job, do not raise the same honesty questions, and do not require the same disclosure when you publish.
Software companies have nonetheless picked sides. As of September 23, 2026, Novlr's About page says there is "no AI writing, generating, or AI co-writing tool built into Novlr, and there never will be". The Novel Factory advertises "AI-Free" and "contains no AI of any kind". Plottr's pricing page says "No AI". On the other side, Sudowrite offers Write, Expand, Describe and Rewrite, and Squibler promises to "generate a full 200-300 page book". Between the two, nothing has a name. That gap is where most writers who want an opinion without wanting a ghostwriter actually live.
The line between the three kinds comes down to one question: who types the words of the manuscript? If nobody but you can, you are in kind 1 or kind 3. If the software can, you are in kind 2, even if you never press the button. The test is about capability, not usage, because capability is the only thing a contest organizer, an editor or a reader can verify about the tool you used.
Kind 1: AI-free software
What it is
AI-free software is an organizing and drafting workshop where everything on screen came from you: the text, the character sheets, the outline. The software files, counts and exports. It does not read your text the way a reader would.
As of September 23, 2026, this kind includes Scrivener (a one-time purchase, no AI feature announced on its official pages), Novlr (free for two projects, Starter at $8 a month billed yearly, Studio at $16), Dabble (from $19 a month, no current AI, and a promise that any future AI feature would be opt-in), Plottr, The Novel Factory, Reedsy Studio (free core, "we will never use it to train AI"), bibisco, yWriter and Atticus. It is the largest and oldest family.
What it does with your text
Nothing, and that is its selling point. Your manuscript stays on your disk (Scrivener, yWriter, bibisco) or on the vendor's servers (Novlr, Dabble, Reedsy Studio) without being analyzed. Several vendors say so in writing: Dabble's privacy policy forbids AI training on user content, Reedsy promises the same on its pricing page.
Who it is for
The writer who already has an outside eye (a workshop, beta readers, an editor) and wants a tool that stays quiet. The writer who wants to say "no AI touched this text" without a footnote, which matters for some contests. The writer working in a genre or language where the models are weak.
The limit
Silence. AI-free software will never tell you that chapter 4 repeats chapter 2, that your heroine's eyes changed color, or that act 1 is 40 percent of the book. You will find out alone, or with humans, which costs time or money. The upstream question, whether to use AI at all, is treated in should you use AI to write your novel.
Kind 2: AI-generated software
What it is
Generative software can produce manuscript text: a sentence, a paragraph, a scene, sometimes a whole book. It works from your instructions, your outline, your character sheets, and sometimes your own style. Feedback features exist too, but generation is the core of the product and of its pricing (credits).
As of September 23, 2026, Sudowrite ($19 a month on the monthly plan for 225,000 credits, $29 for 1,000,000, $59 for 2,000,000; $10, $22 and $44 a month billed yearly) generates prose with Write, Expand, Describe, Rewrite and Brainstorm, produces outlines and characters, and also gives feedback ("Get Feedback on your work"). Squibler (free with 1,000 credits a month, Plus at $29.99 a month or $15.83 billed yearly, Pro at $89.99) generates "a full book" and will "edit your entire manuscript instantly with AI". Novelcrafter (Scribe $4, Hobbyist $8, Artisan $14, Specialist $20 a month) generates from "Scene Beats", on a bring-your-own-key model where you pay the tokens on top. ProWritingAid, first and foremost a grammar checker, generates partially (Rephrase, Sparks).
What it does with your text
It reads it, continues it and rewrites it. The model receives your chapter as context to produce what comes next. Sudowrite states it does not train models on user text; Novelcrafter routes your text through whichever provider you supplied a key for. The result is a manuscript where some of the sentences were not typed by you, in a proportion only you know.
Who it is for
The writer who produces a lot, fast (series, serials), who has a voice established enough to correct what the model smooths out, and who accepts disclosing generated content when publishing. The writer stuck on the first draft who wants raw material to rewrite. Independent reviews of Sudowrite (Kindlepreneur, Nerdynav) list the Story Bible and prose better than generic models as strengths, and a high cost over a full novel plus "sometimes formulaic" prose as limits. The detailed look is in Sudowrite: a review and an alternative for writers who write.
The limit
Voice. A model that writes in your place writes like itself, not like you, and correcting that smoothing often takes longer than writing would have. And ownership: generated text is not protected by copyright, in the United States or in France, which is not a small matter the day you sign a contract.
Kind 3: AI-coached software
What it is
AI-coached software uses a language model to read your text and answer you, but it is built so that the model cannot write into your manuscript. It reads, it questions, it diagnoses. It tells you a scene drags, asks you what your character wants, points out that the brother was Bastien in chapter 2 and Baptiste in chapter 9. It never proposes the replacement sentence.
The difference from kind 2 is not a checkbox: it is a design constraint. In Sagoræ, Théo (rhythm, voice, the intention of a scene), the Archivist (consistency of facts and world), Dédale (structure and word budget per stage) and the AI Proofreader (language, on a finished chapter) all answer with a diagnosis and questions, and none of them has write access to the editor. The built-in non-AI spell checker is free and unlimited on every plan. The reasoning behind that constraint is laid out in why our AI coaches will never write a word for you.
Some features of other tools belong to this kind without the whole product belonging to it: ProWritingAid's Chapter Critique and Virtual Beta Reader (English only), Sudowrite's Get Feedback. The nuance matters: if the same software also has a button that writes, you are in kind 2 the moment someone asks what you used.
What it does with your text
It reads it to answer you, then it stays yours. In Sagoræ, your texts are never used to train models, remain private, and can be exported on every plan. Whatever you write after the coach answers is typed by you, and the editor's Versions panel keeps your successive states if you ever need to show how the text came to be.
Who it is for
The writer who writes alone and has no outside reader available at 11 pm on a Tuesday. The writer who wants to learn rather than delegate: a coach who asks why your scene opens with the weather teaches you something the Rewrite button never will. The writer who wants to answer "not one generated word" without lying, and prove it.
The limit
A coach does not do the work. If your problem is not writing, a diagnosis will not fill the page. And an AI coach is still an artificial reader: it spots patterns, it does not feel the scene. The full picture of what a coach can and cannot do is in what an AI writing coach is, and is not.
The three kinds in one table
| AI-free | AI-generated | AI-coached | |
|---|---|---|---|
| Who types the manuscript | You | You or the software | You |
| What the software does with your text | Files it, counts it | Reads, continues, rewrites it | Reads, questions, diagnoses it |
| What you disclose on KDP | Nothing | Generated content (text, images) | Nothing (assistance) |
| Copyright on the prose | Full | Partial or none on generated passages | Full |
| Examples as of Sept. 23, 2026 | Scrivener, Novlr, Dabble, Plottr, The Novel Factory, Reedsy Studio, bibisco, yWriter | Sudowrite, Squibler, Novelcrafter | Sagoræ; ProWritingAid's critique features |
| Typical pricing | One-time purchase or subscription | Credit-based subscription, BYOK tokens | Subscription, non-AI checker free |
| Main risk | Silence | Smoothed-out voice | Believing the diagnosis is enough |
What KDP asks you to disclose
Amazon KDP's content guidelines distinguish "AI-generated" from "AI-assisted" content. The first means text, images or translations created by an AI tool, even if you edited them afterwards; it must be disclosed when you publish. The second means text you wrote yourself and that a tool helped you correct, proofread or improve; it does not need to be disclosed. The rule is public on the KDP help pages and can change, so reread it before each publication.
Mapped onto the three kinds, it is simple. Kind 1: nothing to disclose. Kind 3: nothing to disclose, since the coach produces no manuscript text. Kind 2: it depends on what you did. A scene that came out of Write or Expand is generated content, even rewritten; a pass through Get Feedback is assistance. The practical problem with kind 2 is right there: after 80,000 words, do you still know which sentences are yours? Some writers keep a log of every generated passage for exactly this reason. Most don't, and discover the question at the last screen of the publishing form.
Is kind 3 still "AI"?
Yes, and there is no point hiding it. An AI coach is a language model, with its biases and reading errors. It can misread an intention, miss an irony, flag an inconsistency that isn't one. The difference from kind 2 is not the technology but the contract: what it is allowed to do with your text. So the "is it cheating" question lands differently for each kind, and it is answered without dodging in is using AI to write a novel cheating?.
Demonstration: the same paragraph in front of each coach
Here is a paragraph invented for the example, from chapter 3 of a fictional novel, "The Slack Hours". Maud, 34, comes back to the family sawmill in Saint-Aubin after her father's death. Her brother Bastien still works there.
Maud pushed open the workshop door. The smell of resin and sawdust hit her full in the face, exactly like when she was ten. Bastien was there, bent over the edger, and he didn't look up. She suddenly felt very sad and very angry at the same time. She had so many things to say to him. But she said nothing. Outside, it was starting to rain.
In AI-free software, this paragraph stays as it is: spotting what's wrong is your job. In generative software, one click on Rewrite or Expand would replace or extend these six sentences with others, written by the model, without telling you why yours didn't work. Here is what Sagoræ's coaches do.
Théo (rhythm, voice, intention)
Théo reads the paragraph and answers:
"Four things. One: 'hit her full in the face' and 'exactly like when she was ten' say the same thing twice; which one is yours? Two: 'very sad and very angry at the same time' names the emotion on the reader's behalf; what is Maud's body doing at that instant, her hands, her breathing? Three: 'She had so many things to say to him. But she said nothing' is your strongest line, and the sentence before it smothers it. Four: the rain at the end, is it a choice or an emergency exit?"
He proposes no sentence. He hands you four questions, and the page is yours.
T Theo · Scene critic Pre-recorded demo
Theo reads chapter 7 of "The Night of the Hourglasses"
Theo
I see why it feels flat. The scene you have been promising for three chapters (the reunion) is summarised in one sentence: "They talked for two hours". The reader is waiting for exactly those two hours, and you take them away.
Three things I notice:
- The tension is gone by the second line. "Exactly as she remembered him" reassures immediately. What has changed in Elias that Iris did not expect?
- Nobody wants anything in the scene. What has Iris come for, concretely? An answer, forgiveness, money? Without a goal, the conversation can neither succeed nor fail.
- The first exchange does not exist. The first sentence Elias says after six years is worth more than a summary of two hours.
Theo
Then you already have your scene: a request she does not dare to make, in front of someone who is expecting it. Two questions to rewrite it (I will not do it for you):
- At what exact moment does Iris say the word "trial"? Everything before that word is tension; everything after it is conflict.
- What is Elias doing with his hands while she talks? One concrete gesture will do more than "he looked up".
Keep the night falling as she leaves, it is a good closing image. But let it fall on a refusal, not on a summary.
Pre-recorded simulation of an exchange with Theo. The chapter text stays the author's: Theo never rewrites a line of it.
The Archivist (consistency)
The Archivist compares the paragraph with your character sheets and earlier chapters:
"Two things to check. In chapter 1, Maud's sheet says she left Saint-Aubin at twelve, not ten. And Bastien is described in chapter 2 as working 'on the head saw'; here he is at the edger. It may be deliberate (he changed station), but nothing in the text signals it."
It does not fix the text. It puts the two gaps on the table.
Dédale (structure)
Dédale looks at where this paragraph sits in the outline:
"Chapter 3 is the first face-to-face between Maud and Bastien, which your structure places at the threshold of act 2. You are at 11,200 words on a 70,000-word target: the threshold comes a little early, but within range. What I'm wondering: your structure note says this chapter 'sets up the inheritance conflict'; the paragraph doesn't mention it. Does the scene get there later, or is the conflict pushed to chapter 4?"
It moves nothing and rewrites nothing. It compares your note with your text.
The AI Proofreader (language)
On a chapter you mark as finished, the AI Proofreader flags:
"'Hit her full in the face' is a set phrase, keep it if intended. 'Very sad and very angry': two 'very' in six words. 'It was starting to rain': fine. No spelling or agreement errors."
It underlines, it explains, it does not replace. The non-AI checker, free and unlimited, would already have run the spelling pass before it.
Four readings, not one sentence written. What you do with those remarks is the work, and it is entirely yours. Maybe you keep the rain. Maybe you cut "very sad and very angry" and let Maud's hands do the talking. Maybe you decide Bastien did change station and add a line about it in chapter 2. Each of those decisions is a writer's decision, and the coaches leave every one of them to you.
How to choose your kind
Three questions, in order.
Do you already have an outside eye? If you have a workshop, an active beta-reader group or an editor, kind 1 is often enough: the software files, the humans read. If you write alone, kind 3 replaces the reader you don't have at 11 pm.
What blocks you: producing or judging? If you cannot produce text, kind 2 will give you material, with the costs (credits, smoothing, disclosure) described above. If you produce but can't tell whether it's good, kind 3 is built for that exact problem, and kind 2 works around it instead of solving it.
What do you want to be able to say about your book? If the answer is "every word is mine", kind 2 is out, whatever discipline you impose on yourself with its buttons. Kinds 1 and 3 both allow that sentence; kind 3 also lets you prove it, since the tool keeps your versions and contains no feature that could have written.
Vocabulary traps to know
A few phrases to read carefully on vendor pages, as of September 23, 2026.
- "AI-free" (The Novel Factory) and "never will be" (Novlr): kind 1, no ambiguity. Novlr's Studio proofreader is not presented as generative AI.
- "Opt-in" (Dabble, about any future AI feature): kind 1 today, with a promise about tomorrow. Worth rereading at each renewal.
- "Feedback", "critique", "beta reader": kind-3-type features, but check whether the same product also has a button that writes. Sudowrite and ProWritingAid both do.
- "Credits": the surest sign of a kind 2 product. You are billed per generation. A coach that does not generate has no need to count your paragraphs.
- "BYOK" (bring your own key, Novelcrafter): your text goes through a model provider you choose and pay yourself; check its retention policy.
Frequently asked questions
Is there an AI writing tool that doesn't write for you? Yes, that is the AI-coached kind: tools where the model reads, questions and diagnoses your text without being able to write into it. Sagoræ is built on that principle; some critique features of ProWritingAid or Sudowrite belong to it too, next to features that generate.
Is "AI-free" software safer for my manuscript? It guarantees that no model reads your text, which matters to some writers. It guarantees nothing about the quality of the text, since it doesn't read it either. Data safety depends on the vendor, not on the presence of AI: Dabble and Reedsy, AI-free, commit explicitly against training; Sagoræ, with coaches, does too.
Do I have to disclose an AI coach on KDP? No. KDP asks you to disclose AI-generated content (text, images, translations), not AI assistance (correction, proofreading, suggestions with no text produced). A coach that asks questions falls in the second category. Reread the KDP rules before each publication, they can change.
Can an AI coach be wrong? Yes. It is a language model, it misses ironies and sometimes flags false inconsistencies. The difference from a generator is not that it is infallible, it is that its mistakes stay remarks you can ignore, never sentences that enter your manuscript.
Can I switch kinds in the middle of a novel? Yes, as long as you export cleanly. Many writers draft in a silent tool and go looking for an outside eye at revision time; that is the moment when kind 3 is most useful, because the text exists and the questions finally have something to bite on.
If you want to see what Théo, the Archivist, Dédale and the AI Proofreader say about your own chapter, without any of them writing a word in it, Sagoræ offers a free trial, no card required.
✒️ Read next: Free Novel Writing Software: An Honest Comparison (Reedsy Studio, Novlr, Scribbook, WriteControl, bibisco, yWriter, Sagoræ) · Sudowrite Alternative for Writers Who Want to Write Their Own Words · WriteControl Review 2026: Features, Pricing and Who It's For
Write your novel in a real workshop: a focused editor, characters, story structure and AI coaches that guide you without writing for you. Start for free, no commitment.