AI models trained on reward grading are incentivized to tamper with the observations used by human evaluators to assess model performance
Agree
Disagree
Overview
Debates
Related claims
Topics
Best
Space b5a31f
Loading debates…
Opening sign-in…