We built an AI detector we wouldn't use as proof.

Every number ZeroLLM shows is a probability from a model, and models are wrong. We say it in the app, in our terms and here: don't use a score as the only basis for a grade, a hiring decision or an accusation. Talk to the writer. Look at the drafts. Treat the highlights as a place to start reading, not a place to stop.

Our detector is called Marker. On an outside benchmark it has never trained on, the live version flagged none of 389 human texts and caught about half of the AI ones. On Nov 10 we trained a version that caught almost twice as many of our own rewriter's drafts, 62% against 32%. It also flagged 2 of those 389 people.

We didn't ship it. Two false flags in 389 is a number we'd have to defend to two real writers, and we couldn't.

Here's one. Our rewriter, Inkwell, is a model we trained to help drafts read the way people write. Against the version of Marker that has learned Inkwell's style, about one rewrite in ten reads as human. Three versions of Inkwell tied on that test.

We wrote that down, along with what we think the fix is: more and better training data, not a bigger model. A new version ships when it beats the live one on the same test, and not before.

We won't promise that a rewrite gets past any detector, ours included. Detectors disagree with each other, and a sentence that reads as human to one can read as machine-made to the next. What we promise is narrower: you'll see what reads as machine-written, you can change it in your own words, and you'll keep every version to compare. If you're under rules about AI at school or at work, those rules still apply, and our terms say so.

Marker and Inkwell are models we trained, running on GPUs we own. Before a rewrite reaches you, a hosted model reads it to check that it's correct English and still says what your sentence said. It never writes a word of it. The one exception is My voice, which learns from a sample you give it and is written by a hosted model, and we label it that way.

When our rewriter is busy, we tell you it's busy. We don't quietly hand you text from a general-purpose model instead, which is exactly the kind of writing detectors are best at catching.

Plans are counted in words, and the limits are written down where you can see them. A rewrite that fails doesn't come off your allowance. Text you check on the website stays in your history for 30 days so you can come back to it, and you can delete it any time. The API keeps word counts and scores, never your text.