Legal AI models have gone through several stages of development. Early legal AI models focused mainly on information retrieval. With the rise of machine learning technology, legal AI models began to be applied to legal text processing. With the development of deep learning technology, BERT and GPT models have revolutionised legal AI models. Legal AI based on BERT or GPT models not only improves the efficiency of solving repetitive and simple legal problems, but also reduces the occurrence of illusions and biases in legal AI. Notwithstanding, legal AI still falls short in providing satisfactory solutions to complex legal issues. The resolution of complex legal matters demands that legal AI not only possess a profound comprehension of legal language but also an understanding of the factual context. However, owing to the complexity, ambiguity, and potential obsolescence of legal language, legal AI lags behind legal experts in terms of understanding legal language. To enable more precise reasoning and knowledge application by legal AI, in the short-term, it is imperative to apply legal AI in appropriate fields. In the long-term, legal AI must enhance legal databases to ensure its capacity for unsupervised pre-training using vast amounts of legal data. Secondly, at its core, legal AI needs to grasp the values underlying the law and integrate these values into its decision-making processes.