The San Francisco-based startup warned that the “current model has weaknesses” such as confusing left and right or failing to maintain visual continuity throughout the length of a video.
The tool isn’t yet publicly available and OpenAI has revealed limited information about how it was built. The company, which has been sued by some authors and The New York Times over its use of copyrighted works of writing to train ChatGPT, also hasn’t disclosed what imagery and video sources were used to train Sora. (OpenAI pays an undisclosed fee to The Associated Press to license its text news archive).
OpenAI said in a blog post that it’s engaging with artists, policymakers and others before releasing the new tool to the public.
“We are working with red teamers — domain experts in areas like misinformation, hateful content, and bias — who will be adversarially testing the model,” the company said. “We’re also building tools to help detect misleading content such as a detection classifier that can tell when a video was generated by Sora.”
