AI Videos Nearly Indistinguishable From Real Videos, Runway Finds

Runway today is releasing its latest AI model, Gen-4.5, that can generate videos from an image. The model, from the startup that specializes in image and video-generating models, is impressive but is likely to underscore growing concerns about AI-generated deepfakes.

To measure the model’s performance, Runway asked over 1,000 people to distinguish real videos from similar videos generated by the model. Participants only guessed correctly 57.1% of the time—not much better than flipping a coin. 

Runway came up with the test when its own employees started having a hard time telling apart videos based on digital footage from the outputs of the new model, said Chief Technology Officer Anastasis Germanidis. “We have probably some of the most trained eyes,” he said, so if Runway’s employees were struggling with the task, they figured people who don’t spend their days scrutinizing AI videos would fare even worse.