OpenAI details Sora 2 safety measures

OpenAI has detailed how Sora 2 and its app aim to reduce the risks of creating and sharing AI-generated videos. The measures include content provenance markers, controls over the use of faces and voices, filters for harmful material, and protections designed specifically for teenagers.

The company presents these features as part of Sora's initial design, not patches added after launch. Even so, it acknowledges that no safeguard is perfect, especially when generated video includes movement, sound, and the appearance of real people.

How Sora videos are identified

Every video created with Sora includes visible and invisible provenance signals. It also includes C2PA metadata, a standard that works like a digital signature and helps show how a file was created.

OpenAI also maintains internal image and audio search tools that can track with high accuracy whether a video came from Sora. Many videos also display a visible, moving watermark with the creator's name, although the watermark does not necessarily appear on every result.

The goal is to make it easier for people to distinguish an AI-generated video from a real recording, even when the content looks convincing.

Creating videos with people requires consent

Sora lets you turn a photo of a person into a video. To do this, the user must confirm that they have consent from the people shown in the image and the necessary rights to upload it.

Videos created from images containing people are subject to especially strict controls. The restrictions are even stronger when children or people who appear to be underage are shown. If such a video is shared, it will always carry a watermark.

Sora also offers characters, a feature that lets you authorize the use of your appearance and voice. You alone decide who can use that representation, and you can revoke access at any time.

Videos that include your character, even if someone else creates them as drafts, will be available for you to review. From there, you can delete or report them. OpenAI says it applies additional controls to prevent extreme changes to a character's appearance, humiliating situations, or uses that take the character too far from the real person's identity.

More limits for minors and harmful content

Teen accounts have additional restrictions on adult content. Their profiles are not recommended to adults, and adults cannot start direct conversations with them.

ChatGPT's parental controls let parents decide whether a teenager can send and receive private messages and whether they will use a non-personalized feed in Sora. Minor accounts also have default limits on continuous scrolling time within the app.

For other users, Sora combines several filters. Before generating a video, it reviews the written prompt, the result across different frames, and the audio transcripts. The goal is to block material such as sexual content, terrorist propaganda, or the promotion of self-harm.

After generation, automated systems review content posted to the feed against OpenAI's usage policies. Human reviewers also handle cases considered more serious. The company says it has tested its defenses through safety exercises and tightened some rules compared with image generation because video is more realistic and includes movement and sound.

Audio adds another problem

Audio can make a fake much more convincing. That is why Sora analyzes transcripts of generated speech to detect potential violations and blocks attempts to create music that imitates living artists or reproduces existing works.

OpenAI also says it handles removal requests from creators who believe that a Sora result infringes on their work. These measures do not eliminate every risk, but they add specific controls that would not be necessary in an image generator without sound.

You stay in control when sharing

Videos do not appear in the feed automatically. You decide whether to publish them and can remove them later. You can also report videos, profiles, direct messages, comments, or characters, and block accounts to prevent them from viewing your profile, using your character, or contacting you.

For you, the main change is that creating a realistic video with someone else's face or voice no longer depends only on writing a prompt. Sora requires permissions, applies stricter filters, and leaves technical traces that identify the content's origin. The key question will be how these safeguards work in real-world situations and how quickly new forms of misuse are addressed.