This is, essentially, the goblin phenomenon with ChatGPT.
Cultural relativism and recurrent use throughout the training data led to essentially an obsession through recursive training, which then manifested in the models and continues to do so today.
It got so bad they had to literally tell it NOT to talk about them.
@OpenAI wrote a great article on it and have actually embraced it incredibly well, and the goblin is now part of their subculture and brand in a positive way.
I'd argue that LLMs in particular are excellent at memetic reinforcement and are essentially hyperstition engines.
I may just be goblin-pilled, but since GPT's obsession you can now see "goblin" trending across multiple seemingly disparate areas of culture - Green Goblin in Spiderman, Goblin Shark obsession, Goblin Mode resurgence, multiple Goblin games on Steam, etc.
What if the LLMs are influenced by, find, and then reinforce certain archetypal memes that then influence further reinforcement through interactions with consumers, creating a hyperstition loop?
Goblin convergence isn't random.