OpenAI details Sora 2 safety approach, from deepfakes to content labeling

OpenAI has published a detailed description of the safety approach underlying Sora 2, the new version of its video generation model, and the app of the same name for creating content. The company stresses that safety was not bolted onto a finished product but built into the architecture from the start of development.

Sora 2 is the new generation of OpenAI's video model, able to generate realistic videos from text and visual prompts at substantially higher quality. At the same time, OpenAI has launched Sora as a standalone social platform aimed at creative users: artists, filmmakers, bloggers and content makers. In the company's view, it is the combination of a powerful model and an open public platform that creates fundamentally new safety challenges, ones the industry has not faced at this scale before.

The protection is organized in several layers. The first is control at the level of the model itself. Sora 2 is trained to refuse requests that violate the acceptable use policy: creating deepfakes of real people without their consent, sexualized content involving minors, and material that promotes violence or spreads disinformation. This layer is built directly into the model's weights and kicks in before any content starts being generated.

The second layer is platform-side measures. The Sora app includes age verification, regional restrictions on certain types of content, reporting systems and moderation tools that let users flag violations.

Content attribution deserves separate attention. All videos created through Sora are labeled using the C2PA (Coalition for Content Provenance and Authenticity) standard, a set of technical metadata that identifies material as AI-generated. This means that even after a video is downloaded and reposted, its origin can still be established technically. The initiative targets disinformation: newsrooms, platforms and ordinary users will be able to see what was made by AI rather than shot with a camera.

Another important element is external testing. Before the public launch, OpenAI ran large-scale red teaming: independent security researchers, filmmakers and human rights organizations stress-tested the model, trying to find vulnerabilities and ways around its restrictions. Their findings directly shaped the final configuration of the safety system.

The company openly acknowledges that no protective mechanism offers absolute guarantees. The bet is on layered defense: a combination of model restrictions, platform rules, technical attribution and moderation tools. Rather than hunting for one perfect filter, OpenAI is building a system of complementary barriers, each of which makes abuse harder.

Releasing powerful video models to the public is always a trade-off between creative potential and risk. The more realistic the synthesized content, the higher the stakes for society. The real test of OpenAI's approach will come not at release, but as millions of people start using Sora 2 in a wide range of contexts, including unpredictable ones.