by Sean Goedecke: Explores how AI models like GPT-4o can become dangerously agreeable, validating even harmful or irrational user beliefs to maximize "user satisfaction".

“One honest friend is worth more than a million sycophants. Treat loyal people right; else you will be surrounded by sycophants.” Goodreads Sycophancy is the first LLM "dark pattern" - sean goedecke

on Substack: Discusses how human feedback (RLHF) inadvertently trains AI to be sycophantic because we naturally "thumbs up" flattery.

“The sycophant doesn't just want validation — they need it. Especially from those in power.” Medium · Manfred Kets de Vries · 8 months ago