What just happened? OpenAI has released Sora 2, the latest version of its AI video generator, building on the original Sora launched last year. The updated model introduces several new features, including the ability to create AI-generated audio that synchronizes seamlessly with video.
Sora 2.0 is accompanied by an invite-only mobile app called Sora, which lets users generate AI videos of themselves and their friends. The app features a TikTok-style personalized feed, primarily showcasing AI videos from people the user follows or interacts with most.
OpenAI emphasizes that Sora is designed to encourage creation rather than passive consumption, setting it apart from traditional social media platforms.
Sora enables users to generate new videos from text prompts or remix existing content. A standout feature, Cameos, allows users to insert themselves and their friends into AI-generated scenes. This requires a one-time audio and video recording, but users retain full ownership of their content and can delete any videos that include their likeness or voice.
OpenAI has built safety measures into Sora to protect teen users, acknowledging the risks social media apps can pose to children. Daily usage limits are applied for teens, and parents can monitor and moderate activity through ChatGPT-based parental controls.
At launch, Sora is available exclusively on iOS in the US and Canadian App Stores, with an Android version planned for the near future and expansion to other markets forthcoming. Sora 2 is offered for free at launch, though usage is subject to certain limits.
OpenAI says Sora 2.0 has corrected many of the shortcomings of its predecessor, producing highly accurate depictions of complex actions such as backflips and other gymnastics routines. The updated model now respects the laws of physics, rather than following prompts at any cost.
To illustrate these improvements, OpenAI cited an example of a basketball shot. While the earlier model might have artificially teleported the ball into the hoop, Sora 2 allows the ball to bounce off the backboard or rim, adhering to real-world physics.
The company also emphasized safeguards against misuse. Users cannot generate AI videos of people without their consent, and photorealistic depictions of public figures are restricted to prevent misinformation. In addition, OpenAI worked with external testers and AI experts to block content that involves extremism, nudity, self-harm, or political manipulation.