No single detector is reliable across modalities. MediaMeter combines independent providers, provenance signals, and cross-modal consistency — then asks a person to confirm before an AI_GENERATED finding is published.
Four modalities
Text
Stylometric and model-based signals, with generator-family attribution where confidence allows.
Image
Frequency, lighting, and physical-consistency checks, plus C2PA manifests.
Audio
Vocoder artifacts, splicing, and speaker-consistency signals.
Video
Frame-level and temporal consistency, lip-sync alignment, and provenance.
Ensemble, not oracle
Eight detection providers are specified, each with known strengths and failure modes. MediaMeter reports provider agreement rather than a single opaque probability. When providers disagree, the result is “inconclusive” and the disagreement is visible.
Provenance beats guesswork
C2PA manifests and SynthID watermarks are checked first. A valid manifest is stronger evidence than any statistical detector. Absence of a manifest proves nothing — most cameras and editors still strip provenance — so it lowers confidence rather than raising suspicion.
Human confirmation
AI_GENERATED findings are never published automatically. A reviewer sees the provider breakdown, provenance status, and any cross-modal inconsistencies, then confirms or rejects the finding. Confirmed findings carry the reviewer’s name.
A signal that content is synthetic does not identify who made it, why, or whether it was intended to deceive. MediaMeter reports what it can support and declines to infer intent.
Technical specifications
Cross-modal consistency
Text, audio, and video are checked against each other where a piece contains multiple modalities — for example, transcript-to-audio alignment and speaker-to-face consistency.
Evaluation and drift
Detectors drift as generators change. Each provider is re-evaluated monthly against a rolling benchmark, and a provider that falls below threshold is disabled rather than silently down-weighted.