Consumers struggle to distinguish videos recorded by humans from those generated by text-to-video tool Sora, according to a new survey on consumers in the U.S.
The online survey was conducted by market research firm HarrisX for Variety, a media publication with more than 1,000 adults.
Respondents were shown four high-quality photorealistic-looking sample video outputs generated by Sora randomly interspersed with four videos from stock video footage taken with a camera. A total of eight videos were shown.
Unable to tell AI videos apart
According to the published results, a majority erroneously identified the real videos as created by AI. In addition, less than half the participants successfully identified the videos generated by AI.
The unveiling of Sora, a text-to-video AI model last month was notable for its potential to upend the advertising and video industry with its ability to generate photorealistic, high-resolution videos of up a minute.
It also raised broader questions about its potential for deepfakes, given its ability to generate a video solely from text instructions. Moreover, it can also take an existing still image and generate a video from it, or animate the content of a still image.
That consumers cannot reliably discern an AI-generated video from a real-life video is troubling.
To preempt the problem of deepfakes, OpenAI in February said that it is building tools to detect misleading content with a detection classifier tool. Video generated by Sora will have secret metadata embedded into the video stream that compatible detection tools will use to determine if a video was generated using AI.**
What was Sora trained on?
In the meantime, OpenAI CTO Mira Murati had an awkward moment in an interview with The Wall Street Journal published online earlier this month.
When asked about the source of Sora’s training data, Murati initially noted: “We used publicly available data and licensed data.” But when pushed to confirm if this include videos on YouTube, she switched to corporate speak.
“You know, if they were publicly available, publicly available to use,” the CTO answered, “but I’m not sure. I’m not confident about it.” Further attempts about whether videos on Instagram or Facebook were used were promptly rebuffed.
OpenAI has not released a release date for Sora, though Murati had separately confirmed that it will be released to the public later this year.
Image credit: DALL-E 3
