Media Verification Signals Research Note

Media Verification Signals Research Note

NIST AI 100-4, updated April 8, 2026, provides the shared technical frame. It treats provenance metadata, watermarking, labeling, and content-based detection as distinct transparency approaches. Evaluation should include unfamiliar generators, representative subject matter, cultural context, compression, resizing, background noise, adversarial changes, and realistic sample sizes.

Detector results are probabilistic. False positives can create serious reputational harm, so a score should not become a public accusation without additional evidence.

Google SynthID is a vendor-specific system for supported Google-generated images, audio, text, and video. Google says the imperceptible watermark is designed to survive common edits. A positive supported result can help attribute compatible media. A negative result does not distinguish among human creation, another generator, unsupported media, or damaged signal.

Ordinary IPTC or EXIF metadata can provide creation, rights, tool, and edit context. It is not cryptographically signed by default and can be edited or stripped.

C2PA Content Credentials add signed provenance, but coverage and preservation remain incomplete. A credential records assertions and transformations rather than deciding truth.

Visual inspection is a hypothesis-generating step. Human studies show performance above chance in some settings but not reliable individual proof. Internet transformations also reduce the value of idealized detector benchmarks.

Recommended evidence order

Use the original source and exact claim first. Add independent corroboration for the event. Inspect signed provenance and tool-specific signals. Review ordinary metadata and reverse-search results. Use detectors as supporting evidence. End with the narrowest conclusion supported by the record.