open source, powers this blog
ai-humanizer: a detector for AI writing tells
46 deterministic rules + stylometry that catch AI slop: em-dash floods, RLHF artifacts, hedging, uniform rhythm. Packaged as a Claude skill. Every entry here passes through it before publishing.
The problem
AI-assisted writing has tells: em-dashes everywhere, RLHF instruction-tuning voice, the not-X-but-Y cadence, rule-of-three, sentences that all weigh the same. Readers feel it even when they can't name it. I write with AI assistance daily, so I needed a referee I couldn't argue with.
What it does
46 deterministic rules across vocabulary, punctuation, rhythm, RLHF artifacts, and structure. No ML, no API calls, same score on each run. Stylometry checks for lexical diversity and sentence-length variance. A 0-100 score where lower is more human; this site's CI fails any entry that scores above 10.
How this site uses it
Every logbook entry runs through the detector in CI before it can deploy. No badge on the post. Just a bar the writing has to clear. If an entry reads like a press release, the build fails and I rewrite it like a person.
Get it
Ships as a Claude skill and a standalone script. Source on GitHub. If a rule flags your favorite construction, the false-positive debate is open.