detect

❌ Don't own:

• Rewriting, humanizing, or "making it undetectable" — editing work, and evasion is refused outright
• Scoring content quality for a platform — that is content:audit
• Naming a specific model or tool as the author — style cannot identify a vendor
• Deciding whether someone cheated, lied, or broke policy — that is a human judgment that needs process evidence

Evidence hierarchy, strongest first: hard artifacts → fabricated or unresolvable references → divergence from supplied author samples → co-occurring stylistic cluster → single stylistic marker (never sufficient).

Don't emit percentages, probabilities, or "X% AI". Research on detectors shows results vary by language, genre, model, and sample; a number implies precision you don't have.

Don't treat any word or construction as proof. "Delve", contractions, em dashes, sentence-length variation — none of these settle anything. Explain the effect in this passage or drop the marker.

Don't name a vendor or model. Say "machine-generated patterns", never "written by ChatGPT".

False-positive risk is highest for formal registers, non-native writers, templated genres (press releases, abstracts, legal text), and text passed through grammar tools. Say this explicitly whenever the band is Mixed or higher.

Short texts (under roughly 150 words) rarely carry enough signal for any band above No notable pattern; say so rather than stretching.

For consequential use — grading, hiring, publication, moderation — the recommended next step is always process evidence: ask for drafts, version history, a live writing sample, or a conversation about the content. Never recommend acting on the band alone.

# AI Authorship Read — {filename or "Pasted text"}

Text type: {genre / purpose}
Length: {word count}
Author sample provided: {yes — N words / no}

Band: {No notable pattern | Mixed signals | Strong pattern cluster}

## Hard artifacts
{list with quoted excerpt and what it is, or "None found"}

## Markers
- {marker name} — "{excerpt}" — why it reads as machine-typical — what else explains it

## Author sample comparison
{register, rhythm, error profile, vocabulary — or "No sample; skipped"}

## What this does not establish
{two to three plain sentences: no percentage, no vendor, no proof of authorship or misconduct; the false-positive profile for this genre}

## Recommended next step
{process evidence to gather, or "No action warranted"}

If the user says "save it": write ai-authorship-{slug}-{YYYY-MM-DD}.md in CWD, slug = first 6 lowercase-kebab words of the text's title or opening line, then surface the path.

Do:

• Quote every marker from the text.
• Name the innocent explanation next to every marker.
• Put "What this does not establish" in every report.
• Recommend process evidence for any consequential decision.