Writers were among the first people to be genuinely uneasy about AI assistants, and for a reason that has nothing to do with whether the technology is any good.
An unpublished manuscript is a strange kind of property. It has no protection worth much until it exists publicly, it’s often the product of years, and its value depends partly on nobody having seen it. The instinct to keep it in one place is not paranoia — it’s how the work has always been handled.
This article sits under What People Actually Use Private AI For.
Who this actually matters for
Novelists with a manuscript years from submission, under contract or hoping to be.
Journalists, where the concern is considerably more serious than commercial. A draft can identify a source through detail alone — a location, a timeline, a turn of phrase. Source protection is a professional obligation and sometimes a matter of someone’s safety. It doesn’t survive “the provider has a good privacy policy.”
Academics with an unpublished argument, an unreviewed dataset, or a grant application in progress.
Screenwriters whose premise circulating is a commercial problem.
Ghostwriters and memoirists working with material that isn’t theirs to disclose, frequently under an NDA that makes the question straightforward.
For all of them the concern is not that a provider would misuse the work. It’s that a draft should exist where the author put it, and nowhere else, until they decide otherwise.
What a model is genuinely useful for
The honest finding, from writers who actually use these tools: it’s a diagnostic instrument, not a writing one.
Structural interrogation. “Summarize what this chapter accomplishes.” If the summary is thinner than you expected, the chapter is thinner than you thought. Brutally effective and hard to do for yourself.
Repetition detection. Models are excellent at noticing you’ve used “flickered” four times in nine pages, or that three characters share a speech rhythm.
Momentum. “Where does this scene slow down?” It’ll usually point at the right paragraph.
Continuity. Eye colour, timeline, who knows what and when.
Motivation. “Why would she agree to this?” — if you can’t answer, neither can the reader.
Dialogue tests. Reading a character’s lines back in isolation reveals whether they have a voice.
Compression. “Cut this by 30% without losing information.” Reliably good, and genuinely tedious to do by hand.
Line-level cleanup. Passive constructions, filter words, adverb pileups.
Talking it through. Explaining a plot problem out loud, to something that asks questions.
What it’s bad at
Writing prose worth keeping. Model prose is fluent, competent, and dead. It has no idiosyncrasy, and idiosyncrasy is most of what makes writing worth reading. Small local models are especially prone to a smooth, generic register.
Taste. It cannot tell you whether a risky choice works. It’ll usually say yes.
Whole-manuscript reasoning. A phone’s context window holds perhaps 6,000–24,000 words. There is no reading the whole novel at once.
Genre-specific judgement. It doesn’t know your market, your editor, or what’s been done to death this year.
The writers who get value from this treat it as an extremely patient, slightly literal-minded first reader who has read everything and understood none of it. That’s a useful thing to have. It is not a collaborator.
A working setup
Model choice. Longer excerpts and structural questions favour a larger model — Qwen 3.5 4B (2.9GB, requires 10GB of RAM) if your device can hold it, otherwise Qwen 3.5 2B (1.6GB, requires 6GB). Leave thinking mode off for prose work; it makes responses slower and more stilted without improving literary judgement.
One project per manuscript. In Cloaked, a project holds its own system prompt and conversations, on-device:
You are a first reader for a literary novel in progress.
- Never rewrite my prose. Describe problems; do not fix them.
- Be specific: name the paragraph, quote the line.
- Do not praise. If something works, say so in one sentence and move on.
- When I paste a scene, tell me what it accomplishes and where it sags.
- Match nothing to my style. I want your reading, not my echo.“Never rewrite my prose” is the important line. Left to itself a model will hand back a smoothed version of your paragraph, and reading it is corrosive — you start writing toward the average.
Work in scenes. Paste one scene or chapter. Ask one question. Start a new conversation for the next. Fits the context window and produces sharper critique.
Keep the manuscript in your real tool. The model is for interrogation, not storage.
Why local specifically
Every cloud provider now offers some form of training opt-out, and for most writing that’s probably sufficient.
But the guarantee is contractual, and contracts describe present intentions. Terms are revised. Companies are acquired. Data is retained for abuse monitoring separately from training preferences, and disclosed under legal process when required. How to Read an AI Privacy Policy covers what the standard clauses actually permit.
For a journalist protecting a source, none of that is acceptable — and the exposure isn’t hypothetical. For a novelist, the question is simpler: your unpublished manuscript now exists on infrastructure you don’t control, under terms you didn’t write, for a period you can’t verify.
On-device inference makes it a non-question. The draft is read into memory on your phone, processed, and discarded. There’s no upload, no retention, no policy to re-read next year. The manuscript stays in one place, which is where it was always supposed to be.
You can confirm this in about thirty seconds by turning on airplane mode and continuing to work — see AI Apps That Don’t Collect Your Data.
One caution worth stating
Using a model to critique your writing doesn’t make it any less yours. Using one to generate text you then publish is a different question, and the legal and professional answers are still settling — they vary by jurisdiction, by publisher, and by contract.
Most writers resolve this simply: the model doesn’t write anything that ends up in the book. It asks questions, points at problems, and counts your adverbs. The sentences stay yours.
That’s also, conveniently, where the tool is actually good.
Download Cloaked on the App Store — work on a manuscript with a model that runs entirely on your device, offline, with nothing transmitted.
Frequently asked questions
Is it safe to use AI on an unpublished manuscript?
With a cloud assistant your manuscript is transmitted and retained under terms that can change. With an on-device model it is processed locally and never leaves the device. For work you intend to publish or sell, that difference is worth taking seriously.
Will AI-assisted editing affect my copyright?
Using a tool to critique or edit your own writing does not transfer authorship of your words. The questions get more complicated when a model generates substantial text you then publish. Rules vary by jurisdiction and are still developing — check the terms of any service you use, and your publisher's policy.
What is AI actually good at for fiction?
Diagnosis rather than generation. It is good at identifying repetition, flagging where a scene loses momentum, summarizing what a chapter actually accomplishes, and asking questions about a character's motivation. It is poor at writing prose worth keeping.
Can a local model handle a whole novel?
No. On-device context windows hold roughly 6,000 to 24,000 words, so work chapter by chapter or scene by scene. In practice this suits editing better anyway, since useful critique is specific rather than global.