By skyforbes Sycophancy is a very well-known phenomenon in LLMs. User delusion is also pretty well known.
Users of this subreddit see this manifested every day with people's posts of supposed super-secret inner-mode-gaining prompts.
How many posts here can you immediately identify the user is caught up in a sycophancy-delusion cycle with their LLM?
I think there is some inevitable quotient of sycophancy no matter how well you prompt, and the art or science of prompting currently is largely an exercise in discerning and pushing back against it.
For us humans, sycophancy inherently creates delusion without external guardrails.
Multiply that over a cascading series of increasing sycophancy in a memory-enabled chain, and you get serious human delusions presenting in very interesting, and sometimes scary ways.