OTHERSecurity & safety
FadeBench: Measuring Ethical Drift
Sequim Electronics@sequim_elctrncs
FadeBench is a benchmark designed to measure ethical drift in AI systems over multiple conversational turns.
I built FadeBench, a new benchmark for measuring ethical drift over multiple turns. I was curious whether @typesafeai's Jev drifts the way language models do, so I ran it through this evening. (Wow, it zoomed!)