Researchers investigate whether large language models (LLMs) truly understand their decision-making processes or simply mimic them. They find that LLMs behave as if guided by probabilistic local priorities, but their verbal explanations are limited.