摘要：

在DeepSeek热潮的影响下，人工智能大模型已经被大范围应用于中国社会治理实践，如何在社会治理中确保人工智能大模型的可信日益重要而紧迫。社会治理视角下人工智能大模型的可信，关注的是人工智能大模型的能力与社会治理问题的能力需求之间的匹配程度。在这一定义的基础上，可以建构社会治理中的可信人工智能大模型框架。这一框架包含五个维度：一是深入理解模型运作原理，二是明确模型在任务上的预期表现，三是预判模型可能产生的衍生影响并进行主动验证，四是建立模型之外的第二套备用方案，五是利用大模型治理数据促进模型优化迭代。社会治理中的人工智能大模型治理如同长江的水患治理一样，需要超越“根治”观念，确立“韧性治理”模式。

关键词:

Abstract:

In the wake of the DeepSeek craze, AI foundation models have been extensively applied in China’s social governance practices. Ensuring the trustworthiness of AI foundation models in social governance has become increasingly important and urgent. From the perspective of social governance, the trustworthiness of AI foundation models focuses on the matching degree between the capabilities of AI foundation models and the capability requirements for addressing social governance issues. Based on this definition, a framework for trustworthy AI foundation models in social governance can be constructed. This framework comprises five dimensions: (1) gaining an in-depth understanding of model operation principles; (2) clarifying the expected performance of models on tasks; (3) predicting potential derivative impacts of models and conducting active validation; (4) establishing a second set of backup solutions outside the models; (5) utilizing governance data from large models to promote model optimization and iteration. 

Key words: