The creator of ChatGPT, OpenAI, developed an AI model that can generate realistic-looking videos, lasting up to one minute, from text. Examples include panoramic videos of 19th-century gold rushes, birthday parties and people moving through busy cities. However, Sora will still take some time to reach the general public: OpenAI wants to ensure that the model is not used to produce inappropriate or illegal content. For example, fake news or sexual content involving minors.
The main objective of the new model, presented this Thursday, is to teach artificial intelligence tools to understand the physical world, in movement. The technology can be used to simplify the creation of simulation systems that allow you to explore scenarios or train skills without the risks or costs associated with real-world experiences. For example, in the area of health, security and education. The name Sora comes from the Japanese word for “sky”.
“We are sharing the progress of our investigation in advance to begin working and receiving feedback from people outside of OpenAI and to give the public an idea of the AI capabilities that are on the horizon”, explains the company in the statement Sora’s presentation. “Despite exhaustive research and testing, we cannot predict all beneficial ways to use technology, nor all forms of abuse. We believe that learning from real-world use is an essential component to creating and launching increasingly secure AI systems.”
The image was generated from the OpenAI request “Images of California during the Gold Rush”
The model is now available for so-called “red teams” (red teams), which are groups that simulate “adversaries” in security exercises. In this case, the risk of the technology being used by creators of fake news or sexual predators is tested.
In 2023, several organizations warned that new artificial intelligence tools were being used to create child pornography more quickly. The theme was one of the focuses of the 2023 edition of TrustCon, an annual meeting of experts whose work is to maintain the security of digital platforms and communities. As a rule, real images of children found online are used as the basis for this crime.
By making Sora available to teams “red team”, OpenAI can develop barriers that prevent this type of use. The company is also making the model available to some filmmakers and graphic artists to understand how the tool can benefit the artistic creation process. In recent months, several writers and artists have shared concerns about the way AI systems use their work to produce new content without crediting their sources.
These fears surrounding the use of artificial intelligence were one of the reasons behind the actors’ strike in 2023 in Hollywood. In November, a group of authors including writers George RR Martin and Jodi Picoult sued OpenAI for allegedly misusing their work to train ChatGPT, another company tool.
A ball that never ends
Just as imaging systems fail in details (for example, the number of someone’s fingers), Sora also reveals some difficulties. For example, the concept of cause and effect. “[Num vídeo gerado pela Sora] a person can take a bite of a cookie, but then the cookie may not have a bite mark”, explained the OpenAI team in the presentation.
Other examples shared include someone trying to blow out the candles on a birthday cake without any effect on them (some candles appear with two flames burning in opposite directions), or someone using a gym treadmill with their back to the console. Although the model is not yet publicly available, in the last 24 hours, the executive president of OpenAI, Sam Altman, also shared several videos on the social network X based on suggestions from internet users, from dogs recording podcasts take a class to learn how to make fresh pasta in Italy.
The creator of ChatGPT does not reveal the data used to train the artificial intelligence model. PÚBLICO contacted the OpenAI team for more information, but did not receive a response at the time of publishing this article.
Sora’s news comes in the same week that OpenAI announced that ChatGPT will now be able to memorize details of user conversations to produce more contextually relevant responses.