AI Tools for Writers: What to Trust

Every AI writing tool describes itself in the same warm language. It will be your collaborator, your co-pilot, your tireless first reader. The copy tells you almost nothing, because the tool that finishes your sentences and the tool that quietly logs your manuscript both use those exact words. If you want to know which tools to trust with an unpublished novel, you have to ignore the homepage and ask questions the homepage was written to avoid.

There are four that do most of the work. Does the tool generate prose or analyse it. Where does your text go when you submit it. Can you see the instruction it runs on your behalf. And who owns whatever comes back. None of these is hard to ask. The reason they feel awkward is that a clear answer often costs a sale, which is exactly why a tool's willingness to answer them tells you as much as the answer itself.

This is the deepest fork, because it decides what the tool does to your craft, not just to your draft. A generating tool produces sentences you did not write and invites you to keep them. An analysing tool reads sentences you did write and tells you something true about their shape, then leaves the writing to you. These are not two settings on one dial. They are opposite jobs that happen to share a screen.

The distinction matters because the costs run in opposite directions. When a tool drafts for you, the page improves and your ear does not, because the labor of choosing each word is the labor that trains you. When a tool only diagnoses, you still write the fix in your own language, so you keep getting better at the one thing you are trying to learn. A useful habit is to ask, after the tool helps, whether you still have to write the words yourself. If you do, it helped you think. If they are already on the page, it wrote for you, whatever the label on the box said.

An AI tool has to send your words somewhere to work on them, usually to a server you do not control, often to a third-party model the tool is built on top of. The questions that follow are not paranoid, they are ordinary diligence: is your text used to train future models, how long is it retained, and can you delete it. Trustworthy tools answer these in plain sentences you can find without a lawyer. Tools that bury the answer in a thousand-word policy, or change the default to training-on and hope you never look, are telling you where their priorities sit.

For a casual draft, none of this may weigh much, and it is worth saying that plainly rather than catastrophising. The calculus shifts the moment you are writing toward submission, a query, or a contract, because agents and publishers care about first-publication rights and about where a manuscript has been. Before you paste in anything you would mind seeing retained, find the data controls for the specific tool and read what they actually say, not what the friendly homepage implies. A tool that makes that easy has already passed a test most fail.

Every AI feature runs on a hidden instruction, the prompt the tool sends to the model along with your text. That instruction is the tool's real opinion about your writing. A pacing feature might be told to flag any scene under four hundred words, or to assume every chapter should end on tension, or to prefer short sentences. You cannot judge whether the advice fits your book until you know what the advice was told to look for, and most tools never show you.

You do not need the literal text of the prompt to apply this test, though a tool that publishes it earns trust. What you need is a tool that is specific about what it checks and why, so its notes are legible rather than oracular. "Your second act has four consecutive scenes where nothing changes" is a claim you can verify against your own pages and overrule if the stillness is deliberate. "This needs more energy" is a verdict with no visible reasoning, and a verdict you cannot inspect is one you cannot trust or learn from. Prefer tools whose feedback shows its work.

If a tool only analyses, ownership is simple, because the words on the page are still entirely yours and the report is just information about them. The question gets sharp only when a tool generates text you might keep. Then you need to know two things the marketing rarely volunteers: does the provider claim any rights over the output, and is text a model generated something you can fully own and copyright at all. The honest answer to the second is genuinely unsettled in many places, which is its own reason to be cautious about pasting machine-written sentences into a manuscript you intend to sell.

This is the quiet logic behind preferring analysis. A tool that reads your whole draft and reports on pacing, plot, and whether each character's voice holds, without ever writing a line for you, sidesteps the ownership question entirely, because it never adds anything for you to own or not own. DraftProse's Reader is built on exactly that line: it diagnoses, it does not generate, so the writing stays yours by design. Whether you use that or a notebook and a sharp friend, the principle is the same. Keep the diagnosis, keep the words, and let the book leave your hands sounding like you wrote it.

Treat the four as a short checklist you actually run before you trust a tool with real pages. Generate or analyse tells you what it does to your craft. Where your text goes tells you what you are risking. Whether you can see the prompt tells you whether the advice is inspectable. Who owns the output tells you what you are left holding. A tool can be useful and still fail one of these, but you should fail it on purpose, knowing the trade, not by accident because the homepage was reassuring.

The meta-test sits underneath all four. A tool that answers these questions in plain language, without making you dig, is run by people who expect to be trusted by writers who think. A tool that dodges them is telling you something too, just not in the part of the page they wanted you to read. You do not have to become a privacy expert or a lawyer to choose well. You have to ask four plain questions and watch how readily they get answered.