The term comes from the ICML 2026 spotlight paper "Large Language Models Develop Novel Social Biases Through Adaptive Exploration" (Wu, Liu, Bai, Griffiths; Princeton/Stanford), which had LLMs play a repeated hiring-decision game across four fictional demographic groups with identical success rates. Models still stratified applicants by group, overgeneralizing from early random outcomes and then exploiting that impression instead of continuing to sample — a failure mode borrowed from human decision-making research (the "hot stove effect"). Unlike training-data bias, this bias has no origin in the model's training corpus; it is generated live by insufficient exploration during a sequential task, and the paper found it gets worse, not better, with newer and larger models. Classical bandit algorithms like UCB and Thompson Sampling, which explore more systematically, show far less of it on the same task.