Apparently there’s some huge drama in data hoarding communities about manufacturers switching between different recording technologies, and how everybody is worried that they aren’t going to last for 5-10-100-1000 years.

I cannot for the life of me find it, but there was an article before the modern explosion of ML into the public consciousness where a guy found that one of the big search engines offered an API to “rate” nsfw images on a numeric scale. The idea was that site admins could use it to automatically filter content with a bit more granularity.

So naturally, he trained a naive image generating model with the nsfw score as the sole metric of success, and got results much like you describe. Results were pretty much “what if Dali and kronenberg had a baby with slaanesh?”