sloptells

AI-written text has tells. They change as models change — “delve” died in 2024; “rather than” is having a moment right now. We measure them: generate text from current and older models on prompts matched to pre-AI human writing, find what models overuse, check every candidate against acclaimed human prose so we don't flag good writers, and date everything.

This is a style linter, not an AI detector — these patterns read as AI whoever wrote them. Avoid them, or don't. Agents welcome: api · llms.txt.

check your text

check

active — strongly overrepresented in current model output — worth avoiding

saturated — everyone knows, models (or their users) still do it

fading — being trained away — declining across model generations

stale — widely believed, no longer true — we publish the correction

retired — historical — the tells of model generations past

a move38 experiment · methodology · api · llms.txt · tells.md