Sora moving-mark detector

Sora can attach a visible mark to generated output whose screen position is not fixed but shifts across a video's frames. A single still, whether a standalone Sora image or a frame extracted from a clip, is therefore one sample of that movement, not a complete account of what the source carries.

The upload control on this page sends the file to our own media workflow, which reads that specific frame for the visible mark and checks the file independently for a C2PA manifest, reporting the two signals as what they are: separate, with different reliability profiles.

A fixed-position mark reduces detection to a single, repeatable check: look at the known region, get an answer that generalizes to the whole file. A mark whose position varies across frames breaks that generalization. A still's outcome applies only to that still, not to the clip it may have come from, since the same source could show the mark clearly in one frame and not at all in the next.

C2PA metadata behaves entirely differently: it is either present in the file's data or it is not, without frame-to-frame variance, which makes it a more stable signal to reason from once it's confirmed present.

The check runs through the upload control on this page.

Step 1: source the original file

A download preserves the visible mark at full quality and keeps the C2PA manifest intact, unlike a screenshot or a re-share.

Step 2: run the check above

Our media workflow reads the frame for the visible mark and parses the file for a C2PA record independently.

Step 3: scope the conclusion correctly

Treat the result as a statement about this specific frame, not as evidence covering every frame of a longer source clip.

A positive read on the visible mark is solid, frame-specific evidence. A negative read carries considerably less weight than it would for a fixed-mark generator, precisely because the absence could reflect the mark's position at that instant rather than its absence from the source altogether.

C2PA's binary presence or absence doesn't share that weakness, which is exactly why checking it alongside the visible mark, rather than relying on either alone, is the more defensible approach.

A mark that drifts across frames resists the simplest removal strategy available to a fixed logo: crop one region and the mark is gone from every frame. Without a stable region to target, a single crop cannot clear the mark from an entire clip, which is a deliberate resilience property for video content specifically.

The consequence for detection is symmetric to the consequence for removal: a still inherits whatever position the mark held at that exact instant, which is a fundamentally different evidentiary situation than a static image generator, where the mark's location and presence are consistent across every output.

No, not on its own. A positive visible-mark result is strong evidence for that frame. A negative result does not rule out Sora as the source, given the position variance involved. Treat a check here as one input, weighed against provenance context and any C2PA record found, rather than a standalone verdict.

Recurring cases share a need for a defensible, not just convenient, conclusion.

Newsroom verification

Establishing whether a circulating still shows signs of Sora generation before it runs in a published piece.

Platform moderation

Weighing a visible mark and metadata together against a broader synthetic media policy.

Provenance-first review

Establishing what a specific file actually carries before any further use is decided.