OpenAI's own system card names Sora's tested failure modes

The shot

On 9 December 2024, OpenAI published the Sora System Card alongside the wider release of Sora, its text, image and video-to-video generation model capable of clips up to 1080p and 20 seconds. Unlike a marketing page, the document is structured around risk identification, red-teaming, and a named mitigation stack, in the lab's own words, for a model the card states OpenAI considers 'a milestone for achieving AGI.'

What the documents show

The card states that red teamers had access to the product 'starting in September' 2024 with iterations of safety mitigations still in development, testing for 'nudity, deceptive election content, self-harm, and violence,' and that this testing 'revealed the need for stronger classifier filtering' after some non-violative uploads could be modified into prohibited content. In a separate, earlier document, OpenAI's own technical report on Sora as a world simulator states directly that 'Sora currently exhibits numerous limitations as a simulator,' naming inaccurate physics in 'basic interactions, like glass shattering,' objects that fail to show correct state changes when eaten, and 'spontaneous appearances of objects' in long clips. These are the lab's own disclosed technical failure modes, not an independent audit; no external red-team report is cited as a separate check on these specific claims.

The workflow

As the system card describes it, a user brings text, image or video inputs and can 'enhance, remix, and blend their own assets' or generate new content from a prompt, with age-gating restricting access to users 18 or older and additional restrictions on uploading a real person's likeness. The card states OpenAI treats this as an iterative deployment: a 'likeness pilot' was planned for a subset of users with 'active, in depth monitoring,' explicitly excluding uploads containing images of minors.

What the tool does not change

The card frames every mitigation as a control layered on top of a model that still fails at physics simulation in the authors' own examples. Verifying consent for an uploaded likeness, judging whether a remix crosses into someone else's rights, and catching the incoherent long-duration failures the technical report names are review tasks the documents assign to red-teamers, classifiers and, ultimately, human moderators and users, not to the model itself.

• Which specific failure modes does the system card name, and does your intended shot fall into one of them?
• What does the card say about likeness uploads, and does your production have documented consent for any real person shown?
• Is a claim about Sora's capability coming from this system card, or from marketing copy the card itself does not make?

A system card is the lab's own account of what it tested and mitigated, written to accompany a release; it is evidence of what OpenAI says it found, not an independent verification that the named risks were fully resolved.