Over the last few days the algorithm thinks I want to see police officers either being mean to children or rescuing kittens or old people because a dog did the "Timmy's down the well" thing and led them to where the rescue was needed.
AI (at the level available to makers of slop videos for social media at least) continues to struggle with lettering, but an interesting thing I've noticed is that continuity of things that are not central to video is very shaky. The officer remains visibly the same "person" throughout, but for example his shoulder patch often changes in every shot. (In this video, it also should match the sticker on the door of his car, but it is also different.)
And the ol' AI bugaboo of hands is still a problem -- it's gotten much better at stills hands, or hands moving slowly in a video, but when the hand moves quickly this often happens:
I suspect that's the AI compromising between "this should be a hand" and "but hands in my training data moving this fast are motion blurred," resulting in this monstrosity.
Other than a continuing recognizable "hyper-realism" look to these things, it is getting to the point where zooming in and freezing a frame is necessary in a lot of these to spot the "tells." All of the images in this post are zoomed in.
In another one, the police officer does stuff that makes no sense, then goes to a pet store to buy food for a presumably hungry dog. The prompt must have said that he was to buy the expensive "Blue Buffalo" brand dog food, as the bag of food looks like their packaging -- but of course real Blue Buffalo packaging has actual words on it, not just stuff that looks sort of like writing -- the AI decided to put "Blue Buffalo" on the officer's shoulder patch!
( I am sorry, I've lost the link to the first video, the second with the dog food is here:
Source: https://www.facebook.com/reel/1462658448897141
)
PS: Though correct in this video, American flags are still a useful thing to check. for a moment there the number of stripes was often wrong, and this still causes problems if the flag is at an odd angle. Getting the number of stars right is still a struggle though! Presumably other flags that have any degree of complexity would also have problems, but I don't see them much in these videos.