What OpenAI says Sora adds
OpenAI says every video generated with Sora carries visible and invisible provenance signals. At launch all outputs carried a visible watermark, and all Sora videos embed C2PA metadata, the industry-standard signed record. OpenAI also says it keeps internal reverse-image and audio search tools that can trace videos back to Sora. Those internal tools are not a public checker, so what you can examine yourself is the visible mark and any C2PA record in the file you have.
Source: OpenAI: launching Sora responsibly (opens in a new tab)
Look at the visible watermark carefully
A Sora export normally shows a Sora watermark on the frame. Its presence is a useful clue, but it is the easiest signal to change: a crop, a zoomed re-recording or an overlay can hide it, and a similar mark can be added to a clip that Sora never produced. If the mark is missing from the copy you have, check whether the frame looks cropped or the clip is a screen recording before drawing any conclusion. Image Evidence does not offer watermark removal and does not help anyone hide provenance.
Check Content Credentials on the original file
C2PA metadata is stored in the video file, so only a copy close to the original export can carry it. Open that file in a C2PA validator such as the Content Authenticity Initiative's Verify tool, which reports the signer and recorded actions it can read; Gemini on the web can also summarize Content Credentials for short clips. The free Image Evidence checker reads images only, so it cannot inspect a Sora video; it can check a still image exported from OpenAI's tools for Content Credentials and AI metadata in your browser.
Source: Content Authenticity Initiative Verify (opens in a new tab)
Re-uploads lose signals
Most viral clips are not original exports. Social platforms re-encode uploads, people screen-record and re-post, and editing apps rebuild the file. Each step can strip the C2PA record, and cropping can remove the visible mark, leaving a copy with neither signal. A clip with no watermark and no credentials is therefore not evidence of a real recording; it is a clip whose origin needs other evidence.
Questions to ask about a viral clip
Find the earliest upload you can and note the account and date. Look for other angles or independent reports of the same event. Read the platform's label and the creator's own description, since some platforms read C2PA data when applying AI labels. Consider whether the claim attached to the clip would matter if it were generated, and avoid sharing a confident verdict that the evidence does not support. Image Evidence has not measured how often any detector identifies Sora videos.
How can you tell if a video was made with Sora?
Look for the visible Sora watermark, then check whether the original file carries C2PA Content Credentials from OpenAI using a C2PA validator. If the copy has been cropped, re-encoded or screen-recorded, both signals may be gone, and you will need the source account, the earliest upload or the creator's own disclosure instead.
Does every Sora video have a watermark?
OpenAI says Sora outputs carry a visible watermark and C2PA metadata, but what reaches you is often a modified copy. Cropping, re-recording or editing can remove the visible mark, and re-encoding can drop the metadata. A missing watermark on a shared copy does not show that the video was not made with Sora.
Is there a free Sora detector?
OpenAI's own tracing tools are internal. A C2PA validator can read OpenAI's Content Credentials when they survive, and general AI video detectors estimate whether footage looks generated, with accuracy that depends on their test data. No public tool reliably identifies every Sora clip after it has been edited and re-shared.