Four signals, scored independently
A check never returns one blended "deepfake score" - it scores each of these separately, so a spike in one doesn't get diluted by the others staying quiet.
| Signal | What it measures | Plan |
|---|---|---|
| Deepfake | Face-swap and reenactment artifacts - the classic deepfake signature of someone else's face mapped onto the person on screen. | Pro & Business |
| Inconclusive Video | Video quality is too degraded - compression, low light, resolution - to reach a confident verdict either way. | Pro & Business |
| AI-generated | Fully synthetic video, not a face-swap over a real feed. When this fires, we also tag a likely source model where we can (e.g. "Likely Veo 3"). | Business |
| Synthetic Audio | Voice cloning or AI-generated speech on the call's audio track, checked independently of the video signals. | Business |
Reading the number
Each signal's score sorts into one of three bands:
- 0-33 - safe. Nothing notable found for this signal.
- 34-66 - warning. Worth a look, not conclusive on its own.
- 67-100 - danger. Strong indication for this signal.
A single check can land in different bands per signal - e.g. a low Deepfake score alongside a high Synthetic Audio score, if the video feed is genuine but the voice isn't.
What can raise a score without a deepfake
Signals respond to media quality as well as manipulation. Low light, heavy video compression, and virtual backgrounds can all push scores up - that's exactly what the Inconclusive Video signal is for: telling you the input was too degraded to trust the other three, rather than letting a poor connection masquerade as a deepfake finding.
Where scores show up
Live scores appear in the Meet add-on as each check runs. Every check is also saved to your session history in the web app, alongside the evidence clip behind it if you've enabled proof retention - see proofs & recap.