2 ms·
I got a couple of the video-only ones wrong, because it's hard to tell a low-quality video of someone not moving very much from a fake version of same. The fake
by karaterobot 5y ago
I got a couple of the video-only ones wrong, because it's hard to tell a low-quality video of someone not moving very much from a fake version of same. The fake audio ones were easy to discern, almost from the first word. They sounded like an SNL impression of those politicians. I don't know whether that means video is, in general, easier to deepfake than audio, or if the examples used in this experiment were bad representations of the state of the art.
- function_seven 5y agoI had the same results. Super easy on audio samples, tough on video-only ones. I suspect we're all so used to compressed video that AI artifacts can hide in that noise. Just the other day a video of Putin was tagged as "green screen!!" because one part of the video appeared to show his hand move through a microphone stand. But it wasn't. It was just shitty compression artifacts making it appear so. (Or, maybe it still was, but I'm not convinced by that evidence) But speech cadences and patterns are super tough to counterfeit.