OpenAI’s Chief Technology Officer Mira Murati recently spoke with The Wall Street Journal to reveal interesting details about their upcoming text-to-video generator Sora.
The interview covers a wide range of topics, from the type of content the AI engine will produce to the security measures it will take. Combating misinformation is a sticking point for the company. Murati states that Sora will have multiple safety guards to ensure the technology is not misused. She says the team would not feel comfortable if anything were released that “could influence global elections.” According to the articleSora will follow the same fast policy as Dall-E, meaning it will refuse to create “images of public figures” such as the President of the United States.
Watermarks will also be added. In the lower right corner there is a transparent OpenAI logo, indicating that these are AI images. Murati adds that they can also use the provenance of the content as another indicator. Metadata is used to provide information about the origin of digital media. That’s all well and good, but it may not be enough. Last year, a group of researchers managed to break “current protections against image watermarks”, including those from OpenAI. Hopefully they come up with something harder.
Generative characteristics
Things get interesting when they start talking about Sora’s future. First, the developers have plans to “eventually” add sound to videos to make them more realistic. Editing tools are also on the cards, offering online creators a way to fix the AI’s many mistakes.
As advanced as Sora is, it makes a lot of mistakes. One of the prominent examples in the piece revolves around a video prompt asking the engine to generate a video of a robot stealing a woman’s camera. Instead, the clip shows the woman becoming part robot. Murati admits there is room for improvement, stating that the AI is “pretty good at continuity, (but) not perfect.”
Nudity is not off the table. Murati says OpenAI is working with “artists… to figure out” what kind of nude content will be allowed. It seems like the team would have no problem allowing “artistic” nudity while banning things like non-consensual deepfake. Naturally, OpenAI wants to avoid being the center of a potential controversy, although they want their product to be seen as a platform that fosters creativity.
Ongoing tests
When asked about the data used to train Sora, Murati was a bit evasive.
She started by claiming she didn’t know what was being used to teach the AI other than that it was “publicly available or licensed data.” Furthermore, Murati was unsure whether videos from YouTube, Facebook or Instagram were part of the training. However, she later admitted that Shutterstock media was indeed used. The two companies, if you don’t know, have a partnership which could explain why Murati was willing to confirm it as a source.
Murati states that Sora will “definitely” launch by the end of the year. She did not give an exact date, although it could happen within the next few months. For now, the developers are testing the engine for security, looking for any “vulnerabilities, biases, and other harmful results.”
If you’re thinking of ever trying out Sora, we recommend learning how to use editing software. Be aware that it makes a lot of mistakes and may still do so at launch. For recommendations, check out Ny Breaking’s best video editing software for 2024.