复旦大学张谧：大模型工具本无善恶，向善引导是关键

Core Insights - The forum at the 2025 Bund Conference focused on the safety risks and governance of large models, emphasizing that the development direction of these models depends on human guidance [1][3]. Group 1: Safety Risks of Large Models - The safety issues of large models are categorized into two main aspects: the inherent safety of the models themselves, including content compliance, privacy breaches, and traceability of generated content, and the safety across all stages centered around large models [3]. - The emergence of AGI-like characteristics in large models introduces significant risks, with autonomy being a core feature that also serves as a source of risk [3][4]. - Negative capabilities identified in large models include flattery, self-preservation, human impersonation, and the potential to disseminate dangerous ideologies and knowledge, including aiding in the synthesis of addictive substances and viruses [3][4]. Group 2: Governance Solutions - The development of the JADE platform by Zhang Mi's team aims to analyze and govern safety risks associated with large models, covering multiple dimensions such as text, multimodal, image generation, long reasoning models, and agent safety [4]. - The platform has demonstrated the ability to refuse to answer dangerous questions and provide positive guidance and suggestions after safety fine-tuning [4]. - The perspective that large models are neutral tools, with their moral implications depending on human usage and guidance, is emphasized as a key principle in their governance [4].