While AI videos once looked artificial and unprofessional, that is no longer the case. The internet is being flooded with increasingly convincing AI videos, and they are often becoming part of advertisements and professional campaigns. This makes it increasingly difficult for the average viewer to distinguish between what is real and what was created by algorithms. This was also confirmed by new research by Runway. For most people, AI-generated video is almost indistinguishable from reality.
The American company Runway specializes in the development of generative artificial intelligence. Its most acclaimed products include the Gen-4.5 video model, which excels at creating hyperrealistic AI videos. And these were presented to a group of more than a thousand people. So what did the study look like?
The experiment involved 1,043 randomly selected participants
The company named its study The Turing Reel and its goal was to determine whether ordinary people could distinguish artificial content from real content. Participants were shown 20 five-second videos, 10 of which were real and 10 of which were generated using artificial intelligence, specifically the Gen-4.5 model. To keep the experiment objective, none of the AI videos were edited or repeatedly generated to achieve better results.
After each video, participants had to decide whether it was real or AI-generated. The results revealed just how well ordinary people can distinguish authentic content from artificially created content. The study considered individuals with a score above 75% to be successful participants.
The type of video also affected the results
The results showed that the days when we could identify an AI video at first glance are long gone. More than 90% of participants were unable to reliably distinguish between real and generated videos. Only 9.5% achieved statistically above-average accuracy, or 99 people out of more than a thousand.
Detection success was also affected by the type of video. People were better at correctly identifying footage that featured people. When faces, hands, or human interaction appeared in a clip, the success rate reached 58–65%.
Viewers performed considerably worse with videos of animals, nature, or urban environments. In those cases, the results ranged from 45-47%. This means that the generated videos appeared more convincing to them than reality itself.
A call for new standards
The results of the experiment clearly show that we have reached an entirely new point in how we perceive videos. The increasing quality of generative models now makes it almost impossible to detect AI content, reducing the chances of verifying its authenticity. This can be highly problematic, especially when it comes to the spread of fake news in fields such as science or journalism.
As protection against the spread of misinformation, all AI videos should contain C2PA metadata that can be used to verify their origin and authenticity. According to Runway, however, this protection is insufficient, and entirely new norms and standards must be established for synthetic videos. Only then can a safe environment be created in which users can reliably identify authentic content and technology companies can continue to push the boundaries of creativity.
The results of The Turing Reel study clearly show that the line between reality and artificial intelligence in video content is already becoming completely blurred. This is ushering in a new digital era in which critical thinking and verification skills will be more important than ever.
Do you think you could do better on the test? Try Runway's experiment for yourself.
