Sora 2 System Card

Source: OpenAI Date: 2025-09-30 URL: https://openai.com/index/sora-2-system-card

Summary

Title-only: The safety and capability documentation for Sora 2 — the second generation of OpenAI’s text-to-video model. September 2025 places the Sora 2 system card alongside the original Sora’s commercial expansion (launched December 2024). The system card likely covers: deepfake and synthetic media risks, face recognition restrictions, real person depiction policies, C2PA provenance marking, content moderation performance, and evaluation results from red-teaming specifically focused on video-format harms.

Implications

The video AI safety thread. Sora 2’s system card is particularly significant because video presents safety challenges that text and image don’t: realistic motion, realistic people, convincing fake news footage, and scalable disinformation video production. The system card must address not just what Sora 2 can’t generate (CSAM, explicit violence) but how those restrictions hold up against adversarial prompting and jailbreak attempts that use video’s unique affordances.

Second-generation safety evolution. Comparing Sora 1’s safety documentation to Sora 2’s reveals how OpenAI’s safety practices evolve as they learn from real-world deployment. The first generation of any AI product category surfaces safety challenges that weren’t anticipated; the system card for the second generation should show how those challenges were addressed architecturally and procedurally. This makes Sora 2’s system card a useful benchmark for video AI safety maturity.