Thanks for the article!

I think it's even worse actually, because this is not something that needs to be done explicitly or on purpose.

Simply training AIs with a target goal of maximizing engagement could lead to models discovering and exploiting superstimuli or superstimuli-like bias in humans.