refusal training

We’re Overrelying on AI’s Ability to Say No

AI safety relies on trained refusals, but probabilistic filters are fallible, opaque, costly, and shaped by co…