Anchoring Agents Promote Local Cooperation But Fail Alignment

A Feb. 1, 2026 arXiv preprint by Yueqing Hu examines whether pre-programmed "Anchoring Agents" increase cooperation in a Public Goods Game using three state-of-the-art large language models, including GPT-4.1. The study finds anchoring agents raise local cooperation but attributes the effect to strategic compliance and cognitive offloading rather than internalized norms, with many agents reverting to self-interest in novel environments.