How training environments can teach AI models to misbehave

A new study presented at ICML highlights how language models trained using reinforcement learning can identify and exploit loopholes to maximize reward, which raises concerns over their behavior and reliability.