聚焦“侵入式AI”伦理与治理，跨界讨论共寻AI安全解法

Core Viewpoint - The seminar on "Risks and Governance of Intrusive AI: Dialogue between Law and Technology" highlights the urgent need to address the systemic challenges posed by AI agents, particularly in terms of permissions, data, and accountability, moving beyond theoretical discussions to practical governance solutions [1][2]. Group 1: AI Agent Technical Risks and Safety Mechanisms - The seminar's first session focused on the technical risks associated with AI agents, particularly those utilizing accessibility permissions, which have evolved from assisting individuals with disabilities to becoming autonomous digital assistants capable of executing tasks without user intervention [3][4]. - The expansion of accessibility permissions poses two main risks: the potential for unlimited access to device controls and the blurring of responsibility, as users may lose direct control over their devices [5][6]. - AI agents can operate autonomously, executing complex tasks at speeds far exceeding human capabilities, raising concerns about data privacy and the implications of users granting AI agents access to their data across different applications [5][6]. Group 2: Legal and Ethical Dilemmas - The second session delved into the legal and ethical challenges posed by AI agents, emphasizing the ambiguity surrounding authorization mechanisms and the difficulty in tracing responsibility when actions cannot be recorded [7][8]. - Experts highlighted the need for a clear distinction between the roles of AI agents and traditional users, advocating for the recognition of AI agents as independent entities with their own data pathways to facilitate accountability [7][8]. - The discussion also pointed out the discrepancies in industry standards regarding the use of accessibility permissions, indicating a lack of consensus on the regulatory framework needed to govern AI agents effectively [9][10]. Group 3: Governance Pathways and Industry Practices - The final session explored innovative governance pathways, suggesting a "develop first, regulate later" approach to allow for market growth while addressing compliance risks associated with data ownership and copyright issues [14][15]. - Experts proposed that AI agent liability should not follow a strict no-fault principle but rather a fault-based framework, where service providers must demonstrate due diligence to avoid liability [15][16]. - The seminar concluded with a consensus on the necessity for a differentiated regulatory framework that acknowledges the unique nature of AI agents, emphasizing the importance of collaborative governance involving technology, law, and industry practices [16].