This is a much needed first attempt at a benchmark to measure how much given AI models will play along with users pushing them in delusional or potentially psychologically dangerous directions. Some early signal that full GPT-5 (not chat) is a less psychologically risky model.

→ View original post on X — @emollick,