Sora 2 adds safety features for AI videos

OpenAI has launched Sora 2 and its new app with controls designed to reduce the risks of creating and sharing AI-generated videos. The system includes visible watermarks, metadata that identifies each video's origin, protections for people's likenesses and specific measures for minors.

The company presents these features as part of the launch, not as a guarantee that no problematic content will appear. The approach combines automated filters, human review and tools that let users control what they publish.

Every video carries signals that it was created with AI

At launch, all videos generated with Sora will include a visible watermark. They will also contain C2PA metadata, a digital signature used to show how and with which tool a file was created.

OpenAI is adding internal image and audio search systems that can trace videos back to Sora. The company says these systems are based on tools it already used for image generation in ChatGPT and in the first version of Sora.

The goal is to make it easier for someone to check whether an apparently real video was artificially generated. That does not eliminate deception on its own, but it provides a technical clue that platforms and users can check.

Your likeness can only be used with permission

Sora lets users create appearances of a person through a feature called cameos. The idea is that each user decides who can use their likeness and voice, and can withdraw that permission at any time.

Videos that include your cameo, even if someone else creates them as drafts, will be visible to you. You will be able to review, delete or report them. You will also be able to set preferences for how your cameo appears, such as asking it to always wear a specific hat.

OpenAI says it applies additional controls to videos with cameos and blocks depictions of public figures. The exception is public figures who voluntarily use the feature themselves.

More limits for teenagers

Sora includes stronger protections for younger users. These include restrictions on adult content and a feed designed for teenagers.

In addition:

• Teen profiles will not be recommended to adults.
• Adults will not be able to start direct messages with teenagers.
• ChatGPT parental controls will allow parents to manage direct messages and choose a non-personalized feed in Sora.
• By default, teenagers will have limits on how long they can scroll continuously in the app.

These measures aim to reduce both exposure to inappropriate content and unwanted contact between adults and minors.

Filters for images, motion and audio

Sora checks user prompts and generated videos to try to block sexual content, terrorist propaganda and the promotion of self-harm before they reach the feed. Reviews analyze multiple frames and audio transcripts, not just a single image.

OpenAI says it has conducted internal testing to identify new risks and has tightened its policies compared with those used for image generation. The reason is that video adds motion, sound and a greater level of realism.

After creation, automated systems review published content against OpenAI's usage policies. Human reviewers also handle cases considered more serious.

Audio adds another challenge. Sora analyzes speech transcripts and blocks attempts to generate music that imitates living artists or reproduces existing works. The company also accepts takedown requests from creators who believe a result infringes on their work.

Publishing is optional

Videos only appear in the feed if the user chooses to share them. Any video, profile or comment can be reported, and any published content can be deleted afterward.

You can also block accounts to prevent them from viewing your profile or contacting you through direct messages. In practice, this means Sora is not just a generator: it also incorporates the rules of a social network, with the moderation and privacy problems that come with that environment.

For you, the most important change is that creating a video with another person's face, voice or style requires more controls than writing text or generating an image. What remains to be seen is whether origin markers survive when videos are downloaded and reposted, and whether the filters respond well to the new tricks users try to find.