AI writing has an accent. It picked it up from us.
Almost every AI writing tell started as a human habit. The difference is how often the machine uses it.
I started keeping notes on things AI writes over and over, across different models. Not only the famous word lists. Phrases, sentence shapes, the way paragraphs end, which day of the week it picks when it wants an ordinary day, even the numbers it makes up for fictional government forms.
The longer the list got, the clearer something became. AI doesn't write differently than people do. We invented the em dash pause, the list of three, the "it's not X, it's Y" turn, the punchy last line. Those moves spread because they work.
What AI does is use all of them, all the time, at the same volume.
A person reaches for a closing zinger now and then. A model puts one at the end of every paragraph. Human writing is lumpy: a long tangled sentence, then three words, then something a little weird. AI smooths that out. The smoothness is what people are hearing when they say something "sounds like AI."
Which is why I get annoyed when people point at a single word and call a piece AI-written. Plenty of humans write evenly. Careful writers do. People writing in a second language do. Whole regions of English speakers use words that got branded as "AI words" (the Field Guide entry on delve is a good example). One tell proves nothing. If there's a signal at all, it's density.
So this site tries to do two things. The Field Guide traces each tell back to where humans started using it, then asks how much more often AI uses it than we do. The tools help you check a draft and pull the accent out of your own AI-assisted writing, if you want to.
It's an open project. If you've noticed something AI says a lot, send it in and we'll try to measure it.
Field Guide
The tells, where they came from, and how to test them yourself.
Open the guideHeatmap
Paste a draft and see where the habits cluster. Not a detector.
Run a draftPrompt index
55 copy-paste instructions for getting your own voice back into AI-assisted drafts.
Browse promptsSeen a tell we're missing?
Submissions open soon. To make a tell measurable, it helps to send it in this shape. Hunches are welcome. They go in as hypotheses until someone tests them.
- What the tell is: a word, phrase, sentence shape, rhythm, number, or name.
- A real example from an AI output.
- Which models you've seen do it.
- A prompt that reliably brings it out.
- Your guess at where humans got it, and anything you can cite.
Where the accent came from
Each entry is a habit AI leans on, traced back to the people who made it a habit first. Think of it as etymology, but for writing moves instead of words.
Sourced means the human origin is documented. Observed means someone has counted it in real text. Hypothesis means it's reasoned but untested, and it stays marked that way until someone brings evidence for or against it.
Heatmap
Paste a draft to see two things: where the habits cluster, and how much range the writing has. A highlight isn't a verdict. Most of these are things people write too. Click any highlight to see where the habit came from or how to change it.
Think of it like an audio master. Human writing usually has dynamic range: long sentences, short ones, a sudden two-word line. Writing that has been smoothed out looks like a brickwalled track, every bar about the same height. It measures sentence length only, so it's a rough picture, not a verdict.
This is not an AI detector and it can't tell you who wrote something. It counts patterns. A few hits in a long piece is normal human writing. What's worth a second look is the same move showing up again and again in a short space.
Fix your own drafts
Instructions you can paste into ChatGPT, Claude, Gemini, or any writing tool. Copy one for a quick fix, or stack several into one set of voice rules. Use the ones that fit how you write and skip the rest.
Try fewer words, or clear the category filter.