Briefs
Briefs
Today
OpenAI released a Sora safety document on March 23 covering C2PA watermarking, dual-layer content filters, and teen protections — one day before shutting the app down.
On March 23, 2026, OpenAI published "Creating with Sora Safely," a comprehensive account of the guardrails built into its Sora 2 video generation model. The document outlined C2PA provenance metadata embedded in every output, a dual-layer content filtering system, consent-based "Characters" likeness controls, and additional protections for teen accounts — published one day before OpenAI announced it would shutter the standalone Sora app, Sora.com, and the Sora API entirely.
The timing transforms the safety document from a proactive framework into a retrospective account. Sora's closure — driven by an estimated $15 million per day in inference costs at peak and a download decline from 3.3 million in November 2025 to 1.1 million by February 2026 — shows that even well-documented safety infrastructure cannot compensate for a product unable to achieve commercial sustainability. The collapsed Disney deal, which would have brought 200 licensed characters to the platform, removed a potentially stabilizing revenue stream.
Two content filtering layers operated in sequence: a text model blocked harmful prompts before video generation began, and a visual model reviewed generated frames as a secondary check. Every output embedded C2PA metadata — an industry-standard provenance signature — along with visible, dynamically moving watermarks on videos featuring real people. The Characters feature gave users control over their likeness and voice, while the system blocked depictions of public figures outside that consent framework.
Google Veo, Runway, and Luma AI now inherit Sora's market position. OpenAI's exit signals that high inference costs remain the central unsolved problem for consumer AI video — a lesson that will recalibrate how competitors price their services and time their go-to-market decisions. OpenAI has indicated the underlying Sora research team will continue focused on world simulation for robotics, meaning the technology is repurposed rather than abandoned.