Paused · reopens after a pre-test · verdict by 31 December 2026

Real authors. Your own pages. One honest answer.

We think we can make an AI keep a writer's voice. We are not going to take our own word for it, and we're not going to ask strangers on the internet either — they'd be judging a voice that isn't theirs, which is how you get a confident answer to the wrong question.

So we need you: someone with a manuscript, judging edits to it.

What you'd do, once it reopens

1. I send you the privacy commitment in writing, before you share anything. If you're happy with it, you send about 20,000 words of something you wrote, in one continuous stretch: roughly the first five or six chapters.
2. About 10 minutes, by email. I propose six passages from your book and you tell me which ones sound like you. You know your book; I don't. Rather talk it through? We can do a short call, but you don't have to.
3. We generate passages from your chapters — some with our technique, some with plain Claude given the same material. You won't be told which is which, and neither will we while you're judging.
4. About 45 minutes, on your own, whenever suits you. You get a single file — six pairs, nothing to install, no account, nothing sent anywhere. For each pair, two questions: which sounds more like you, and would you publish this as yours?

Nobody is watching while you judge. That's deliberate. You shouldn't be weighing up my work with me on a video call, and re-reading something twice over a coffee is how you'd use a tool like this for real. Take breaks, go back over a passage, change your mind — that's the point.

There are no dates yet. Leave your email below and you'll get one message when the study reopens, with the dates and what's involved. There's nothing to do in the meantime, and no obligation once you've seen it.

What you get

The full results before anyone else, including the raw numbers — and if the technique fails, we'll tell you that too, in the same detail. That is the likelier outcome and we'd rather you hear it from us than notice we went quiet. If it does ship, participants get it free for a year.

You also get the whole design explained afterwards — exactly how the pairs were built and what each one was testing. Some of that has to stay unsaid until you've judged, or the judging measures the wrong thing. Ask afterwards and you get all of it.

The deadline slipped once. It won't again.

This page used to say that if ten authors hadn't judged the test by 31 August 2026, we'd write that up rather than slide the date to the right. We missed that date, and we did move it, once, to 31 December 2026. There was no answer to dodge. The test simply cost more author time than we could recruit, so we made it cheaper instead of pretending it had run. If there is no verdict by 31 December, the project stops and we write up why. There won't be a second move.

Questions first? hello@writerswithai.com reaches a person. Or read what we're actually building.