The brazilian Miku video makes weird noises, in fact it sounds the same way it would sound if the mic was in fact just right next to the fabric she moved, which isn't the case. She doesn't even touch the mic with her shoulder or hair, yet the sound is the same it would make if you decided to "caress" the phone with fabric.
Makima video is believable because the hair doesn't act like real hair, it acts like poorly maintained wig. Which is a nuance a public generative AI without fine tuning would struggle to understand. Only weird thing is her mouth suddenly opening with just one frame.
Frieren one (I assume it's Frieren) has weird movement and this sharpness a lot of generative AIs seem to go far by default, but I can't say for sure.
I kinda disagree about the "thumb wrong side" one, it's her pinky, shortened by perspective and her glove. Her hair is heavily edited tho, but I wouldn't say it's AI just for that, everyone edits.
As for AI above editing, I wonder which cosplayers you "know", because unless they self-host, it doesn't cost less than editing at all. Free models exist but most of them will recreate a whole image which will look like the original, not edit. Localized editing is something some model would propose, but I have no knowledge of anyone doing it for free. Even then, it would lack the fine-tuning necessary to make it okay-ish, and fine-tuning takes way more time than editing manually.
What's funny, as Ambrosia said, is that you didn't provide comparison. Also a funny thing, trying to tell us you "know a bunch of cosplayers doing it" while I somehow only know cosplayers not doing it, or using it in a very conservative way. You're right tho, sometimes heavy editing can make someone look AI-generated, remember Mersweb...