What Data Powers OpenAI’s Upcoming Video Model Sora?

In a Wall Street Journal interview, OpenAI CTO Mira Murati gave vague answers about the data sources for the upcoming video‑generation model Sora, stating that the model is trained on publicly available and licensed data.

The journalist asked whether Sora was trained on content from platforms such as YouTube, Instagram, or Facebook, to which Murati replied she was not sure, adding that if the data were publicly accessible it could have been used.

“You know, if they’re publicly available—publicly available. But I’m not sure. I don’t have confidence in that.”

When questioned about OpenAI’s partnership with Shutterstock, Murati confirmed that Shutterstock data was indeed used for training Sora, though she declined to detail the dataset.

AI models rely on large training datasets to learn patterns, make predictions, and understand language.

Murati has been with OpenAI since 2018, leading projects such as DALL‑E 3, Whisper, and GPT‑4, and briefly served as interim CEO after the board’s removal of Sam Altman in November 2023.

OpenAI faces multiple lawsuits over the use of copyrighted material in its training data, including cases filed by authors in July 2023 and a December 2023 lawsuit by The New York Times alleging the use of its content to train chatbots, as well as a class action in California accusing OpenAI of scraping private user information without consent.