Firefly video watermark detector scope

Firefly's video signal is Content Credentials: a C2PA manifest, cryptographically signed and embedded in the video container, asserting which tools contributed to the asset and whether a trained model produced the content. This route parses that structure and reports what it finds.

The file is uploaded into our media workflow, where the container is walked, any manifest is extracted, its signature is checked and its assertions are reported. Uploading is structural. A manifest is inside the file, so parsing one requires the file, and no page here will borrow a browser-only claim that does not apply.

Which file you supply determines the answer more than the check does.

Step 1

Obtain the original asset. A screen recording, a messaging-app copy or a platform download has been re-encoded and will normally carry nothing.

Step 2

Upload it here. The container is parsed, any manifest extracted, the signature verified.

Step 3

Interpret against the three states below, weighting a negative result far lower than a positive one.

An MP4 is an ISO base media file: a tree of typed boxes carrying video, audio, timing and index data. The format reserves one box type for payloads it does not itself define, the uuid box, which carries a 16-byte identifier declaring what its contents are. C2PA claims the identifier D8FEC3D6-1B0B-4872-BD80-E04F827E80D9, and the manifest is stored there.

Two consequences follow directly from that placement. The record is separable from the encoded frames, which is what makes removal a container operation rather than a re-render. And a screen recording can never carry it, because a recording reconstructs the presentation from decoded output while the uuid box was never part of what got decoded.

Reporting this as a binary is the standard analytical error in provenance work. The states differ not only in what they say but in how much they are worth.

Present and valid

Signature verifies, content matches what was signed. Strong evidence about the declared history, because forging a valid signature is expensive and checking one is cheap.

Present and invalid

A record exists but fails verification, or the asset has diverged from the signed state. Informative in its own right: it indicates modification after signing, not an absence of provenance.

Absent

Consistent with an asset that never carried credentials and with one that lost them in transit. Taken alone, close to worthless as evidence.

Metadata attrition is high for images and higher for video. There are at least four independent paths to a blank result, and none of them involves anyone trying to strip anything: the generating export never embedded a manifest, a screen recording captured playback rather than the file, a platform transcoded on ingest, or a share sheet rewrapped the container.

Video sits in a pipeline that re-encodes almost by default. Platforms produce multiple renditions per upload, messaging clients recompress aggressively, and screen recording is a far more common capture method for video than screenshotting is for stills. The base rate of accidental stripping is therefore high enough that a negative scan supports essentially no inference about how the clip was produced.

The operational conclusion is that provenance on video is a chain-of-custody problem before it is a tooling problem. A file received directly from its creator supports inference. The same clip pulled from a feed does not, whatever any detector reports.

Scanning a container for the C2PA identifier and reporting a hit answers a narrow question: is a C2PA structure present. Detectors that do only this state as much in their own documentation, describing it as a fast provenance clue rather than a verification.

Verification is the harder half: does the signature validate, is the certificate behind it trusted, and does the current content still hash to what was signed. Only that pass distinguishes a valid record from a broken one, which is precisely the distinction that makes present-and-invalid a useful third answer rather than a footnote to found.

This route performs provenance parsing. It does not analyse frames, does not produce a likelihood score for AI generation, and does not infer a generator from container heuristics such as resolution, duration, aspect ratio or estimated bitrate. Screening tools that use those heuristics describe their own output as preliminary and state that they cannot name a generator without credentials.

Merging provenance parsing with visual screening behind a single verdict would make the output uninterpretable, because the two techniques fail in opposite directions. A manifest answer is precise because it is narrow, and publishing the boundary is what makes the narrow answer usable.

Editorial and broadcast desks performing provenance review on supplied footage, where a declared generation assertion changes labelling obligations or disqualifies the clip. Agencies and brand teams auditing delivered video against a brief that restricted generated material.

Compliance functions building a documented record of what was checked and when, where the audit trail is the deliverable rather than the verdict. And anyone verifying, after a removal, that no manifest remains in a file about to be published.