How verdicts are scored

Every piece of content gets one score between 0 and 1 and a label: ai at 0.7 or above, uncertain from 0.4, and human below that. The card shows the score as a percentage with a ten-segment bar.

The order of precedence

1. Provenance first. If any provenance detector ran and asserted that the content is AI-generated, the score is at least 0.9. Content Credentials declaring "trained algorithmic media" score 0.95, IPTC "trained algorithmic media" 0.95, a matching EXIF generator tag 0.9, and a composite that includes AI media 0.8 (which then rises to the 0.9 floor). These are what the file declares about itself, so they win. See Content Credentials and EXIF and IPTC metadata.
2. Then the generator classifier. With no provenance signal, a pixel classifier's probability is the score. When a model recheck also exists the two are blended: 60% classifier, 40% recheck.
3. Then the recheck alone. If only a model recheck exists, its probability is the score.
4. Then writing patterns. For text with none of the above, the on-device writing-pattern score is used on its own. Community reports are not part of the score at any stage; they are displayed beside it as a count of unverified opinions.

Text under 40 characters is scored 0.1 and noted as too short.

Which detectors are which

Detector	Kind	Tier
Writing pattern (stylometry)	Text	Free, on device
SynthID and provider provenance checks	Images	Free
Content Credentials (C2PA)	Images, video containers	Free
EXIF generator tags	Images	Free
IPTC digital source type	Images	Free
Community cache and reports	Both	Free
Generator classifier	Images	Pro triggers; cached results free
Model recheck	Both	Pro, on request

What the circle uses

The circle appears when the score is at or above your sensitivity preset (95%, 90% or 70%), which is separate from the 0.7 verdict line. See Sensitivity presets.