Contents
- What does a SynthID check on a video actually read?
- Which AI video carries SynthID, and which does not?
- What does a hit mean, and what does a miss mean?
- What about Sora and other AI video?
- Where can you actually check a video?
- How strong is the mark, and where does it fail?
- A watermark is one signal, not a verdict
- Sources
A SynthID check on a video tells you one narrow thing, whether Google’s invisible SynthID watermark, applied at generation time to Google’s own video models, is present. It is not a general AI-video detector, and it is blind to AI video that carries a different provenance mark. Read a hit as “a Google AI video tool made this,” and a miss as “no SynthID watermark was found,” never as proof the footage is real.
What does a SynthID check on a video actually read?
SynthID embeds an invisible pattern into the frames of generated video. Google DeepMind says the mark is “added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.” The check answers one question, is that pattern present, and nothing else. Gowal, Bunel, Stimberg et al. (DeepMind 2025) frame the whole family this way: “establishing provenance is materially different from detecting AI-generated content.” A SynthID check is a probe for one watermark, not a classifier that reads AI-ness out of the pixels.
Which AI video carries SynthID, and which does not?
Only video from providers that adopted SynthID carries it. Google applies SynthID to its Veo video model and across its consumer generative products, and reports the mark has covered “over ten billion images and video frames” (Gowal, Bunel, Stimberg et al., DeepMind 2025). Every Veo output is marked. Video from any tool that has not adopted SynthID, which is most of the field, returns “not detected.” The footprint is large but cooperative, and a miss is the normal result for non-Google AI video.
What does a hit mean, and what does a miss mean?
| Result | Most likely meaning | Does not mean |
|---|---|---|
| Watermark found | A Google AI video tool (Veo) made it | The clip is unedited or truthful |
| Not detected | No SynthID mark is present | A human filmed it, or it is not AI |
Google’s video output also carries a visible “Made with Google AI” label on some products, but the visible tag can be cropped, while the invisible SynthID is the part a check reads. A hit still means a SynthID-adopting tool generated the clip, not that it is unedited or presented in context.
What about Sora and other AI video?
They use different marks, so a SynthID check cannot see them. OpenAI states that every Sora video embeds C2PA metadata, an industry-standard signature read as set out in how to read C2PA Content Credentials, alongside a visible watermark at launch, which is a provenance system separate from SynthID. Meta released Video Seal, an open-source neural video watermark, under a permissive license, another different mark again. A SynthID check returns “not detected” on Sora and on Video Seal content, not because the video is real, but because it was marked by a different scheme. Reading each requires its own reader, a C2PA viewer for Sora, and none of these tools cross-reads another.
The Sora case also shows why a miss proves little. Independent reporting on Sora’s launch found that C2PA metadata was embedded reliably only when a video was downloaded through official channels, and that clips circulated with the visible mark cropped or absent, so content labeling is, in practice, far from complete. A stripped or missing mark is common in the wild, and expected.
Where can you actually check a video?
Access is limited. Google’s consumer route is to upload the clip to the Gemini app and “ask if it’s been created or altered by Google AI.” Google’s SynthID Detector portal is not a universal public scanner; it is in early testing with journalists and media professionals. There is no single public tool that reads every video provenance mark, so a “not detected” from any one checker means “not detected by this tool, for this scheme,” not “no watermark exists.”
How strong is the mark, and where does it fail?
On its home ground SynthID detection is strong and, per Google, survives cropping, filters, frame-rate changes, and lossy compression. But those are vendor descriptions for video, not an independently published video benchmark, and the payload side is softer even in the image paper: recovered-payload accuracy fell to 71.86% in the worst transformation category, the combined case stacking rotation, cropping, and JPEG (Gowal, Bunel, Stimberg et al., DeepMind 2025). So detection can stay strong while the finer attribution bits degrade under heavy handling, and re-encoding or screen-recording a video is exactly the kind of handling that stresses a mark.
A watermark is one signal, not a verdict
Because SynthID is cooperative, absent from most AI video, and, like any watermark, forgeable, a check is one input rather than a ruling. Golaszewski, Krawetz and Sherman (2026) put it generally: provenance systems provide “provenance signals, not proof of authenticity,” and watermark research has shown marks can be scrubbed and spoofed, with Jovanović, Staab and Vechev (ICML 2024) reporting scrub-and-spoof attacks “for under $50” at an “average success rate of over 80%.” Treat a SynthID hit as one corroborating signal, check for a C2PA manifest as a second, and read a miss as the absence of one cooperating mark rather than a clearance. Why that clean result proves nothing on its own is the subject of does “no watermark” mean an image is real.
Sources
- Gowal, Bunel, Stimberg et al. (2025). SynthID-Image: Image Watermarking at Internet Scale. Google DeepMind.
- Google DeepMind. SynthID.
- Google DeepMind. Veo.
- OpenAI (2025). Launching Sora Responsibly.
- Meta AI. Video Seal: Open and Efficient Video Watermarking.
- Golaszewski, Krawetz, Sherman (2026). Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short.
- Jovanović, Staab, Vechev (2024). Watermark Stealing in Large Language Models. ICML 2024.