OpenAI publishes Sora 2 system card

OpenAI published a system card for Sora 2 alongside the model’s launch as a social video app, setting out the safeguards built around a text-to-video model capable of generating realistic, sound-synced clips of real people. The card focused on three risk areas specific to video generation at this fidelity: non-consensual likeness use, provenance and detectability of synthetic content, and the model’s physical and factual realism making misleading videos harder to spot as fake.

The main technical safeguard described was the “cameo” system, requiring a person to record a one-time verification clip before their likeness could be used, with the ability to revoke consent or delete videos featuring them at any time. OpenAI said generated videos would carry C2PA metadata and visible watermarking to support provenance tracking, and described classifiers intended to detect attempts to generate content depicting real people without authorisation, sexual content, and other prohibited categories.

As with prior OpenAI system cards, the document disclosed safeguards and evaluation categories without publishing full red-team results, pass rates, or details of training data. It also did not anticipate the scale of the copyright dispute that followed within days of launch, when users generated large volumes of videos featuring copyrighted characters under an opt-out policy OpenAI reversed within roughly 72 hours. Weeks later, complaints from actor Bryan Cranston and the union SAG-AFTRA over unauthorised likeness use prompted a further tightening of the consent rules the system card had described as already in place.