Creating with Sora Safely

OpenAI is trying to get ahead of the deepfake problem before Sora 2 turns into a headline for the wrong reasons. The company's latest post lays out what it calls a foundation of concrete protections built directly into both the video model and the new Sora app, rather than safety features bolted on after launch.

The timing is not an accident. Sora 2 is a meaningfully more capable video generator than its predecessor, and pairing that with a consumer-facing social app changes the risk calculus entirely. A single model producing realistic clips is one thing. A feed where people can remix each other's likenesses and share the results widely is another, and OpenAI is clearly aware that the second scenario invites far more trouble than the first.

So the emphasis lands on consent and control rather than just content filters. Features like cameo, which lets someone insert their own likeness into a generated video, are being framed around explicit permission — you decide if and how your face gets used, and presumably can revoke that later. That is a direct response to the obvious worry: without guardrails, a tool like this becomes a machine for putting words and actions into people's mouths and bodies without them ever agreeing to it.

OpenAI is pitching this as an ongoing project rather than a finished product. Given how quickly people find creative ways to misuse generative tools, that framing is probably honest rather than just corporate hedging. The real test won't be the policy language in a blog post, it'll be what happens once millions of users start pushing on the edges of what cameo, remixing and sharing actually allow.