The Sora Watermark

Short answer

The Sora watermark is a visible mark OpenAI placed on videos made with its Sora app, alongside hidden C2PA provenance data. A visible mark is a useful hint, but a missing one proves nothing: marks get cropped, blurred or faked. To check a clip, look for content credentials, visual tells and the original source.

A watermark feels like a verdict. It is not. It is a sticker on the picture, and stickers can be peeled off, covered or copied onto something else. What it does well is tell an honest viewer, at a glance, that the uploader did not bother to hide where a clip came from. What it cannot do is prove anything about clips without one.

Two dates matter here. In its launch post for Sora 2 and the Sora app, dated 30 September 2025, OpenAI said every video generated with Sora includes both visible and invisible provenance signals. The same page now carries a notice that, as of 26 April 2026, the Sora product is no longer available. Clips made before then are still being reposted, which is why people still search for the mark.

What the visible mark looks like

OpenAI's launch post says only that at launch, all outputs carry a visible watermark. It does not publish a size, position or design spec, so be wary of anyone who claims a mark "must" sit in one corner. In practice a visible watermark is a small logo or name drawn over the picture, part of the pixels rather than part of the file's data.

That matters for what it proves. Because the mark is just pixels, three things are true at once:

• A clip with a genuine mark almost certainly came out of the tool that drew it.
• A clip with no mark may still have come from that tool, after a crop or a cleaner.
• A mark can be pasted onto real footage, by someone making a joke or a point.

So read a visible mark as a strong hint about one clip, and its absence as no information at all.

Content credentials (C2PA) and how to read them

The invisible half is C2PA metadata. OpenAI said all Sora videos embed C2PA metadata, an industry standard signature. C2PA is the Coalition for Content Provenance and Authenticity, and its "Content Credentials" are tamper-evident, cryptographically signed data structures that travel with the asset, recording how a file was made and changed.

To read them:

1. Get the most original file you can, not a screenshot or a screen recording.
2. Upload it to the Content Credentials Verify tool.
3. Look at who signed the credential and what tools it lists.

Google's Gemini app also reads Content Credentials. Its help page says that, where present, it shows media composition, edit history and whether AI was used, and it is equally clear about the limit: Content Credentials do not tell you if a file is definitely AI generated.

Platforms read them too. YouTube says it may apply an AI label automatically to content that contains C2PA metadata. TikTok announced in May 2024 that it would read Content Credentials to automatically label AI generated content uploaded from certain other platforms. If a clip came straight from the tool, a platform label may be the credential doing its job.

Cropped, blurred or removed marks

Both layers can be lost. The C2PA FAQ says it directly: metadata can be removed inadvertently or intentionally. Re-encoding, screen recording, editing apps and some upload pipelines all drop it without anyone trying. The C2PA answer is "durable" credentials that use watermarks or fingerprints to find the record again, but you cannot count on every file having one.

Visible marks fail in plainer ways. Watch for these signs that one was taken out:

• An odd crop: the frame is tighter than a normal phone aspect, or the subject is cut off at an edge.
• A smudge or blur patch that sits still while the scene moves behind it.
• A sticker, emoji or caption box parked over the same spot for the whole clip.
• Black bars or a zoom that appears only in a reposted version.

None of these prove AI. People crop and caption real videos all the time. They tell you the clip has been handled, so the mark's absence means even less than usual.

Checking a clip without the mark

When the watermark is gone, check the clip itself and its history.

Look at it. Hands, text, physics, light and sound still give generated video away. The visual tells guide has frames for each, and the nine-check method adds labels and source checks.

Look for other watermarks. Google says its AI video carries an invisible SynthID mark designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression, and its SynthID Detector page says it now checks content from Google and partners including OpenAI. SynthID explained shows where to run that check and what a result means.

Run a file check. The free Sora video checker, like the homepage detector, reads the file in your browser for the same Content Credentials, as set out in the C2PA specification, and for AI labels that use the IPTC digital source type codes, then shows still frames from the clip for you to inspect. You get one of four verdicts: AI generated, Likely AI, Likely real, or No AI signs found. It tells you whether the file says it was made with AI, not which tool made it, and it is free with no signup. "No AI signs found" is one signal among several, not proof the clip is real.

For a frame-by-frame AI score, DeepfakeDetector.ai has a free account with 50 scans a month for clips up to 2 minutes, and it also checks images and voice.

Find the first upload. Reverse search a clear frame, check the account that posted first, and search fact checking sites for the claim. A watermark only ever answers "which tool". The source answers "who, when and why", which is usually the question that matters.

If you think a clip came from Sora, the Sora video checker puts the file check and the watermark advice side by side. Guide maintained by Tessa Ostrowski.

Sources

1. OpenAI, 30 September 2025, with product notice dated 26 April 2026: Launching Sora responsibly
2. C2PA: FAQs
3. Content Credentials: Verify tool
4. Google, Gemini Apps Help: Verify AI-generated images, videos, and audio
5. YouTube Help: Disclosing use of GenAI content
6. TikTok Newsroom, 9 May 2024: Partnering with our industry to advance AI transparency and literacy
7. Google DeepMind: SynthID
8. C2PA: C2PA Technical Specification 2.1
9. IPTC: Digital Source Type NewsCodes

All accessed 9 November 2026.