humanizelisted

# Humanize Rewrite prose to remove AI tells. Preserve meaning, strip the cadence. ## The two layers AI-ness lives at two levels, and they are not equally durable. - **Surface layer** (vocabulary, punctuation, sentence rhythm, paragraph scaffolding): the tells in [references/tells.md](references/tells.md). Cheap to detect and cheap to remove, but transient. Newer models shed them on their own, and a light edit strips the rest. - **Discourse layer** (whether the piece took a real, contestable position; whether it holds genuine tension; whether its specifics are named and concrete): the moves in [references/rewrite-toward-human.md](references/rewrite-toward-human.md). Stripping the surface layer is necessary but not sufficient. In a controlled test, running AI text through exactly this kind of span-level artifact removal (clichés, purple prose, redundant exposition) moved a discourse-level detector by only 1.6 points: the structural fingerprint survived the scrub (Russell et al., 2026; see [references/sources.md](references/sources.md)). So for anything longer than a short message, treat the positive-direction pass (step 5) as load-bearing work, not a finishing gloss. If a passage has no point of view, no amount of tell-stripping will make it read human. ## When to invoke User pastes text and asks to clean it up, flags something as sounding like AI or ChatGPT, asks for "tells" to be removed, or asks for general humanizing of a paragraph, email, blog post, doc, or message.