Those same models have already been subjected to the equivalent of probably millions of hours of RL on simulated user messages containing every possible combination of self-harm, torture, murder plots, genocide, child endangerment, CSAM, etc to test refusals.
"Model welfare" seems like an objectively absurd concept given the inherent training process needed to produce a usable instructions tuned model.
"Model welfare" seems like an objectively absurd concept given the inherent training process needed to produce a usable instructions tuned model.