Tests show OpenAI's Sora can closely mimic Netflix shows, movies, TikTok videos, and Twitch streams, suggesting it was trained on versions of such content
Tests by The Post suggest the training data for OpenAI's video generators Sora included versions of movies, TikTok clips and Netflix shows.
Context & Ripple Effects
Sora began as a limited research release capable of generating short text-to-video clips, and early outside testing found stronger facial realism than rival tools. The Post's mimicry tests move the story from model capability toward the provenance of the material behind those capabilities.
That provenance question becomes more consequential as OpenAI expands Sora's product surface: the later Sora 2 launch emphasized multi-shot instruction following, while the app subsequently added character-cameo and clip-stitching tools.
First-order effects
- OpenAI faces sharper questions from Netflix, TikTok, Twitch and other rightsholders over whether source material was used in Sora training; the reported tests are evidence of resemblance, not by themselves proof of a particular training dataset.
- Media owners and creators gain a concrete basis to scrutinize Sora outputs for recognizable styles, formats and scenes, raising the compliance burden around use of the tool.
Second-order effects
- Video platforms and studios may tighten their approach to licensing, data access and technical controls if publicly available or platform-hosted video can be replicated by competing generative tools.
- Sora competitors will face greater pressure to explain training-data provenance and to offer rights-management or provenance features, rather than competing solely on output quality.
Third-order effects
- The dispute points to video becoming both a distribution product and an inference input, with the commercial value of catalogs and creator archives increasingly tied to their availability for model development.
- If resemblance-based investigations continue to surface across models, generative-video commercialization is likely to hinge more on traceable licensing and auditable data practices; the eventual balance will depend on technical evidence and rights-holder responses.
The trend: Generative video is shifting from a race over visual realism toward a contest over who can lawfully source, document and commercialize the video data that makes that realism possible.