de-LLM-ify · voice-fit meter

Does your writing sound like a person?

Paste a paragraph. The meter scores how close it reads to unedited human prose - the stock vocabulary, the reflexive “not X, but Y,” and the flat, even cadence that gives machine text away. It grew out of a tool I built to catch my own site drifting into that voice.

This is a voice-fit meter, not an AI detector. It measures how close your writing is to unedited human prose - the tells, and the deeper rhythm underneath them - not whether a machine wrote it. It runs entirely in your browser; nothing you paste is uploaded or stored. It is a mirror, not a verdict.
Paste or type your prose
0 words
What’s it about?

Name the subject and its own words stop counting against you - a post about a harness says “harness” because that’s the topic, not because a model reached for it.

Try one
Start writing and the voice-fit meter scores it live - the number, the band, and exactly which tells are firing.
How it reads this

This is a voice-fit meter, not an AI detector. It does not try to answer whether a machine wrote your text, which is both unanswerable on modern models and beside the point. It measures something smaller and more useful: how close your writing is to the way unedited humans actually write. It looks for the tells. Stock surge vocabulary. The reflexive "not X, but Y". Connective scaffolding like moreover and ultimately. And, underneath the words, the deeper signal the research points to: burstiness, the natural swing between long and short sentences that models tend to iron flat. Everything runs in your browser. Nothing you paste is sent anywhere or stored. The score is a mirror, not a verdict, and it is honest only inside those limits.

Burstiness - How much your sentence lengths vary. Humans swing between long and short; models flatten out. Higher is more human.
Excess vocab - Rate of the post-2022 surge words (delve, underscore, robust, seamless…) per 1,000 words. Lower is better.
Noun-heaviness - Verbs buried in -tion / -ment / -ness endings per 100 words. Lower reads more direct.
Em-dashes - Em-dashes per 1,000 words. Only meaningful against a personal baseline, so it is shown here for information, never scored.

Keep reading

The essay that built this

Why the meter calibrates on pre-model writing, and where it is honestly wrong.

All the browser tools

The rest of the small, private-by-default tools that run in your browser.

The AI benchmarks

The same show-your-work habit, pointed at how the models themselves score.

The reading lists

Another public, content-as-code corner of the site.