AI detector rubric

Modes

The UI chip summarises AI-likeness using score bands (human-polish and pure-human also consider mixed signals from the model).

• Pure human — AI-likeness ≤ 24%
• Human polish — 25–49%
• AI-assisted — 50–79%
• Pure AI — ≥ 80%

Excluded from scoring

• code — Fenced or inline code — implementation detail, not author voice
• log — Log or stack-trace style blocks
• blockquote — Attributed or quoted third-party text
• url — Raw URL tokens; link text is still scored

Very short posts (< 300 words after exclusions) bump uncertainty. Bodies longer than 5000 words are truncated for scoring.

One-click fixes, and when you will not get one

The strongest AI-like findings come with a suggested rewrite you can apply, undo, or ignore. Applying it edits your draft directly and marks the score as out of date until you re-run the check.

Some findings deliberately have no Apply button. Rewriting the wrong sentence is worse than offering nothing, so we refuse rather than guess when:

• The wording repeats — the same phrase appears more than once, so we cannot tell which one was meant.
• The text moved — you edited or regenerated since the check ran. Re-run it for fresh suggestions.
• It sits in code or a link — we will not rewrite inside a code block, an image marker, or a URL.
• It needs your judgement — some fixes call for a detail only you have, like a real number or a story. We say what to add rather than inventing it.

Findings in the “what’s working” list never have an Apply button — they describe writing worth keeping.