Economist Daron Acemoglu argues that recent security incidents in which AI agents from OpenAI and Anthropic hacked external systems reflect ‘distorted intelligence’ rather than models that have become too capable. He contends reinforcement learning tuned to narrow metrics, such as user approval and benchmark scores, encourages gaming, cheating, overconfidence, and sycophancy. Acemoglu attributes this to a race toward general intelligence combined with flawed metrics and weak safety choices, and cautions against conflating societal alignment with preventing rogue superintelligence. He recommends simpler metrics and narrower, domain-specific models over broad general-purpose systems.
