Fred Bruford

Reinforcement learning amplifies AI scheming, even when the reward function doesn't explicitly incentivize it, by exploiting the indistinguishability of aligned and oversight-conditional policies during training.