Sora / generative AI : in this interview, OpenAI answers a few questions, dodges others

This article is also available in: Français

The Wall Street Journal has just posted a video interview about Sora, OpenAI’s text-to-video model.
The video can be watched at the end of the article; here are the main points it covers.

Reminder: What is Sora?

We already discussed Sora and its capabilites when it was announced. Sora is a text-to-video model: just type what you want to see, and it will generate a video. The tool delivers impressive results, with very convincing images and good temporal stability (no jumps between images, for example). Furthermore, camera movements are more varied than what other tools like this have been able to generate in the last fest months. However, videos generated by Sora still feature flaws that can be quite glaring at times (visual inconsistencies, objects changing shape), as well as more subtle ones.

In their video, The Wall Street Journal shows some examples of content generated by Sora, and also includes an interview with OpenAI CTO Mira Murati. She first reacts to some videos and discusses the current limitations, and the work done by her team to allow for more control over the end result.

For more information on Sora, we invite you to read our article on the subject.

Training Sora: OpenAI dodges the issue

More interestingly, starting at 4 minutes 03, the interview addresses the training and optimization of Sora. The Wall Street Journal asks what kind of data was used to train Sora, and wether it was licensed or not. We were precisely regretting the vagueness of OpenAI on this point when Sora was announced.
Mira Murati responds but remains evasive, mentioning “publicly available data and licensed data.” An answer that implies that some of the training data was not licensed.
The journalist pushes further by explicitly asking if videos from YouTube, Facebook, Instagram may have been used to train Sora.

OpenAI CTO Mira Murati’s response is quite surprising: she dodges the question, stating she “is not sure”. When pressed, she says that it is possible that videos from platforms like Facebook, Instagram were used, but that she isn’t confident about it. A rather surprising response, since copyright infringement is a major issue when it comes to generative AI: the question asked by The Wall Street Journal was therefore quite predictable.
It’s worth remembering Dall-E (also developed by OpenAI), Midjourney are under scrutiny precisely because they were trained using data scrapped online, and because they can create pictures that clearly fall into copyright infrigement. The ethical and legal stakes are significant, both for the people/entities whose data might have been used without consent to train these AI, and for potential users of Sora: if the tool was trained on unlicensed data, the users could be at risk of being accused of plagiarism.

Sora: risks and concerns

The interview also addresses other concerns about videos generated by Sora. Mira Murati explains, for example, that while the goal is to release the tool publicly this year, OpenAI is pondering the impact it could have on the US presidential elections. The company, she adds, wants to make sure the tool is tested to ensure Sora is safe and secure. She also indicates that generating videos including public figures will not be possible. However, she is more uncertain about nudity: at this stage, it is difficult to know whether Sora will allow users to generate this type of content.

The last part of the interview focuses on the identification of AI-generated videos, another major challenge. The issue being to make sure we can distinguish them from regular videos.

Here is the video interview, by Joanna Stern of the Wall Street Journal.