This study investigates the cognitive bias mechanisms in Legal Large Language Models (LLMs) from a cognitive science perspective. With the in-creasing integration of AI into judicial decision-making, legal AI systems like LLMs have the potential to reshape legal processes. However, these systems also risk perpetuating biases, as exemplified by incidents such as the COMPAS algorithm’s racial bias in parole decisions. The research explores the sources of cognitive biases in LLMs, including training data biases, algorithmic inductive biases, and contextual misalignments, and com-pares these with human judicial biases. By examining the similarities and differences in bias formation mechanisms between humans and LLMs, the study highlights the importance of addressing algorithmic fairness and ethical governance. It proposes an interdisciplinary governance framework that combines technological, institutional, and ethical measures to mitigate biases, ensuring that legal AI systems support fairness, accountability, and transparency in the legal domain.