OpenAI is seeking to reduce the prohibitively high costs of its latest groundbreaking text-to-video tool, Sora, but acknowledged that its potential to spread misinformation during a crucial election year would impact its commercial rollout plans.
Chief Technology Officer of OpenAI, Mira Muratitold the Wall Street Journal that there is still a lot of work left to streamline your powerful tool to do justice to the size and complexity of their image generator DALL-E. These efforts aim to reduce the computational requirements and therefore the price.
“We’re trying to make it available at some point at a similar cost to what we saw with DALL-E,” Murati said, adding that Sora is still “much, much more expensive” than the latter.
Just a year ago, the videos created by some written prompts were amateurish and instantly recognizable. But Sora creates 60-second clips that are so realistic that the untrained eye can hardly tell they’re fake.
While people have gotten better at it recognize telltale signs Because an image is not authentic, we have little experience with advanced AI videos, making them perfect fodder for malicious actors looking to influence public opinion.
Asked whether Murati therefore felt comfortable releasing Sora before November – when the entire House of Representatives, a third of the Senate and the White House are up for election – she said the possibility of misuse of her proprietary technology was a concern.
“This is certainly a consideration that addresses the issues of misinformation and harmful bias, and we will not publish anything that we are not confident about in terms of what impact it could have on global elections,” Murati said, who briefly stepped in as interim CEO in November Leadership crisis.
“We have to sort out these issues”
Take Mark Kaye, a conservative white radio host in Florida.
BBC Panorama reported this month about how he created AI-generated images of Donald Trump surrounded by a crowd of enthusiastic black voters and shared it with his million-plus followers Facebook as proof that the former president was popular among all demographic groups.
However, it’s not just Americans who are voting.
The United Kingdom is likely to hold a general vote as early as May, while European Union citizens will cast their votes for the next EU Parliament in June.
This is likely to be the case even as Indians decide whether Narendra Modi’s Hindu nationalist party continues to lead the world’s largest democracy.
All in all, more than half of the world’s population will vote in 2024 Time magazine has called a “decisive year for democracy”.
However, thanks to AI, voters will have to decide more than ever whether the content they see and hear is real or not.
“We need to resolve these issues before we can safely deploy them widely,” Murati said.
Does OpenAI use data from copyrighted content to train?
This is because OpenAI works with the help of a major shareholder Microsoft has made every effort to create the most compelling generative AI tools around – even if Sora Video doesn’t yet have equivalent audio output.
“Our goal was to really focus on developing the best capabilities,” she said, explaining why Sora is still too expensive to market to consumers. “Now we will start optimizing the technology so that it is cost-effective and user-friendly for people.”
However, when asked about the origin of the datasets used to train Sora, Murati was unable to provide an answer other than to claim they were all publicly available or licensed.
“I’m actually not sure,” she said when asked if they had taken advantage of it Youtube Videos included.
Authors like George RR Martinas well as media like that New York Times, recently sued OpenAI for copyright infringement, arguing that their content was used to develop a commercial product like ChatGPT without their consent without compensation. (OpenAI was also sued by Elon Musk over one separate problem)
When further asked about the source of Sora’s training data, Murati declined to comment.
“I just won’t go into the details of the data used,” she answered awkwardly, “but it was publicly available or licensed data.”