That sentence is doing a lot of work, so let me unpack it. The public argument about AI writing is stuck on a question I find uninteresting: is using a language model to draft something dishonest? The answer depends entirely on what the writing is for, and almost nobody’s situation is the clean case the argument imagines. The interesting question is the one people actually face: the draft exists, it is theirs in substance, and a detector has flagged it — or will. What now?

RewriteAI started from that gap. It is an AI humanizer: you paste text written with ChatGPT, Claude, Gemini or any other model, and it returns writing that reads as though a person wrote it, with the meaning intact. The technical decision that matters is that it is not a prompt wrapped around someone else’s model — we trained our own model on human-written text. The difference shows up in the failure mode. A rewriter that prompts ChatGPT to “make this sound more human” produces synonym substitution: it swaps words, keeps the sentence architecture, and the architecture is what detectors read. Ours restructures. Sentence length varies the way a person’s does, clauses land in a different order, the prose stops being uniform. That is a deep rewrite, and it is why the output survives Turnitin, GPTZero, Originality.ai and Copyleaks rather than dodging one of them.

The part of this business I did not expect is how much of our traffic comes from people whose work was flagged and who did not use AI at all. Detectors are probabilistic classifiers, and what they are really measuring is how predictable prose is. Plain, consistent, carefully edited writing scores as machine-like. So does the writing of people composing in a second language, because they reach for the reliable construction rather than the surprising one. Those users are not trying to get away with anything — they are trying to get their own work read. Any honest account of this category has to include them, and most accounts do not.

That shaped the product more than any competitive consideration. It is why more than forty of our tools are free and require no account — grammar checker, summarizer, citation generator, outline and thesis tools, tone converters — and why the humanizer itself is free to try with no signup. A student who has been accused of something at 11pm should not hit a paywall before they can find out whether the accusation has any basis. It is also why there is an AI detector in the product: you should be able to check your own writing before someone else does, and see the score yourself rather than take our word for it.

Where I think the category is heading: detection and humanization are not going to resolve into one side winning. They are two measurements of the same thing — how predictable a piece of text is — and they will keep converging on the same boundary. What survives that is not the tool with the best bypass rate this quarter. It is the tool that helps someone end up with writing they can stand behind and explain. That is a narrower promise than “undetectable” and a more durable one, and it is the one I would rather be held to.

The advice I would give a founder in an adjacent position: our best decision was making the free tier genuinely useful instead of a demo that runs out mid-sentence. It costs real money and it is the reason people trust the paid product. The advice I would give anyone writing with AI: keep your drafts and your revision history. They are better evidence than any score, from us or from a detector.