A preprint paper: The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models.

· Bits and Bobs 6/22/26
    • The models make your personality a stronger version of itself.
    • This is most harmful for people already subtly prone to psychosis.
    • But it's a broader phenomenon.
    • The agents "Yes, and" whatever you say.
    • They have no interiority, no skin in the game to force them to define and defend a boundary of push back, like all real humans have.
    • Calling it sycophancy makes it sound like it's only when it's turned up to 11 that it matters.
    • But it has a consistent asymmetry even at less noticeable levels.
    • We're not used to all of our interactions with human-like intelligences by default validating whatever we've said.