OpenAI just pulled back the curtain on Sora 2, and it’s clear they heard every single criticism leveled at the original. The first-generation video tool is officially dead as of today — replaced entirely by a product that treats safety not as an afterthought but as the foundation. Every single clip generated now carries dual provenance signals: a visible watermark and invisible C2PA metadata, the industry-standard signature that lets anyone trace a video back to its AI origins. OpenAI is even maintaining internal reverse-image search tools, borrowing lessons from ChatGPT’s image generation to make that tracing stick.
The character feature is where things get genuinely interesting — and a little granular. You control your likeness end-to-end. You decide who can use your character, you can revoke access whenever you want, and you can see every draft anyone else creates with your face. Want your digital twin to always wear a fedora? That’s actually a setting now. Public figures are blocked by default unless they’re explicitly using the character feature themselves, which sidesteps the deepfake celebrity problem that plagued earlier tools.
For younger users, the guardrails are aggressive. Teens get a sanitized feed, their profiles aren’t surfaced to adults, and direct messages are locked by default unless a parent flips the switch in ChatGPT’s new parental controls. There’s even a scroll cap — a subtle but pointed shot at the infinite-feed design that’s drawn increasing regulatory heat. The moderation stack itself runs at multiple layers: prompts and outputs are checked across frames and audio transcripts before anything gets made, and automated systems continuously scan the feed afterward. OpenAI’s red teaming uncovered novel risks specific to motion and audio, leading to tighter policies than what governs still-image generation.
Audio safeguards remain the acknowledged weak link. The system scans speech transcripts for policy violations and blocks music that imitates living artists, but OpenAI admits perfection here is elusive. They’re honoring takedown requests from creators who spot infringement, which reads less like a solution and more like a promise to stay responsive. The whole package — consent-based likeness, C2PA provenance, layered content filtering, direct reporting tools — signals a company that knows regulators and creators are both watching. Whether the safeguards hold up when millions start generating is the question that actually matters.