Detection signals and verdict bands
How synthetic-media detection is specified to work: modalities, the five verdict bands, provenance verification, and why detection advises while humans decide.
This describes the synthetic-media detection specification — the verdict bands, review routing, and provenance checks. Detection is a signal, never proof; design, not deployed capability.
Synthetic-media detection is probabilistic: generators improve faster than detectors, models drift, and a confident output can still be wrong. Our specification treats detection as one input into an editorial decision — never the decision itself.
The four modalities
Signals are collected separately for text, image, audio, and video, plus a cross-modal flag when a package’s components disagree (e.g., authentic audio with synthetic visuals). Different detector families cover different modalities — no single provider covers all four reliably, which is why the specification calls for a provider ensemble rather than a single detector.
The five verdict bands
Each piece of content resolves into one of five bands:
| Band | Meaning | Routing |
|---|---|---|
| AUTHENTIC | Signals consistent with human/original production | No escalation |
| likely_authentic | Mostly consistent; minor anomalies | No escalation |
| uncertain | Signals mixed or weak | Medium-priority human review |
| likely_ai | Signals lean synthetic | High-priority human review |
| AI_GENERATED | Signals strongly synthetic | Critical — mandatory human review before any verdict |
A suspected deepfake is disqualifying by design: it caps assessment at the lowest rating and always routes to a human — because a deepfake is not a quality defect to grade but an integrity event to escalate.
Provenance verification
Detection signals are statistical; provenance is a document trail. The specification checks:
- C2PA manifests — cryptographically signed content credentials recording capture device, edits, and software. A valid manifest from a trusted signatory is stronger evidence than any classifier.
- SynthID — Google’s watermarking for AI-generated media (specified, not yet enabled).
- EXIF/XMP metadata — camera and editing-tool data; weak alone (trivially stripped) but corroborating.
Detection and provenance answer different questions — “does this look synthetic” vs. “where did this come from” — and the specification treats their disagreement as its own flag.
The editorial rules
- Detection advises, editors decide — no automated publish/reject is triggered by a detection score.
- Disclosure, not just penalty — AI involvement that is disclosed (as the EU AI Act requires) is assessed differently from the same content undisclosed.
- Re-evaluation on drift — detector providers are re-evaluated on a rolling benchmark; a provider that falls below threshold is disabled rather than silently down-weighted.
- Review SLAs by band — deepfake/AI-GENERATED flags route critical; likely_ai and cross-modal flags route high; uncertain routes routine review.
Last reviewed: October 20, 2026. Part of the MediaMeter knowledge base — editorial reference material, not product documentation.