关键词: 生成式人工智能, 证据推理, 法律推理, AI幻觉

Abstract:

As the application of generative AI in the judicial field continues to deepen，intelligent reasoning models such as DeepSeek-R1are gradually revealing their inherent rational limitations，which are primarily manifested as a conflict between technical rationality and judicial rationality. The strict standards of proof and severe legal consequences in criminal cases demand a high degree of certainty in criminal legal reasoning. This has led to a distinctive pattern of reasoning in criminal law：a heavy reliance on empirical evidence in factual reasoning and an emphasis on value-based balancing in the application of law. However，the probabilistic and statistical reasoning mechanisms underlying generative AI carry inherent uncertainty. They lack an understanding of empirical knowledge and the ability to make substantive judgments regarding reasoning and values，making them ill-suited for the reasoning tasks involved in complex criminal cases. To expand the application prospects of generative AI in criminal legal reasoning，it is necessary to keep a “human in the loop.” At the institutional level，the auxiliary role of generative AI in reasoning must be clearly defined，and its involvement in core reasoning during the trial phase must be restricted. At the technical level，a criminal law knowledge graph should be constructed to promote the explicit application of rules of thumb in judicial rulings and to attempt to infuse generative AI with empirical knowledge through supervised fine-tuning. At the same time，it is necessary to enhance algorithmic explainability，improve the fidelity of reasoning chain outputs，and reduce the risks of hallucinations and bias. Ultimately，this will achieve a balance between the characteristics of criminal judicial reasoning and the advantages of generative AI，thereby realizing the synergistic development of technological empowerment and judicial fairness.

Key words: generative artificial intelligence, evidence reasoning, legal reasoning, AI hallucination