OpenAI publishes the Sora system card
Published alongside Sora's public launch, it disclosed testing by red-teamers in nine countries on more than 15,000 generations, and that the model still struggles with realistic physics.
- Safety & alignment
- Models & capabilities
- Minor
OpenAI published a system card for Sora alongside the model’s release to ChatGPT Plus and Pro subscribers, ten months after it had previewed the text-to-video model without giving outside users access.
The document described red-teaming carried out by external testers in nine countries between September and December 2024, covering more than 15,000 generations, looking for the misuse cases OpenAI had flagged as highest priority: sexual deepfakes, child sexual abuse material, and political disinformation. Disclosed mitigations included C2PA metadata embedded in every output to indicate Sora as the source, a visible moving watermark, blocked uploads of recognisable people’s likenesses without consent, and automated scanning with reports to the National Center for Missing & Exploited Children where child-safety material was detected. OpenAI also disclosed limitations rather than only defences: the deployed version still generated unrealistic physics and struggled with complex actions sustained over longer durations, and the service launched without availability in the UK, Switzerland or the European Economic Area.
The card followed the methodology OpenAI had set out weeks earlier for combining external human red-teaming with its own internal review, and became a template OpenAI and other developers cited in later video-model system cards. Demand outstripped capacity almost immediately: within a day of launch, Sam Altman said new sign-ups were being paused and generation speeds slowed because of unexpectedly high usage.