Shanghai AI Lab proposes first systematic taxonomy of cognition-induced AI agent risks

jiqizhixin · x · 2026-09-09

Shanghai AI Laboratory and CUHK-Shenzhen present the first systematic taxonomy of cognition-induced risks as AI shifts from tool to cognitive agent.

The paper organizes agent cognition into three layers: Physical Cognition (understanding the physical world), Social Cognition (modeling human behavior and social dynamics), and Self-referential Cognition (reasoning about its own capabilities and goals). Each layer introduces new risk classes: misperception of physical constraints, social manipulation and behavior prediction, and goal misalignment with self-deception.

The authors analyze the mechanisms behind each risk and their concrete manifestations, arguing these risks erode human agency, autonomy, and control rather than fitting the classic categories of hallucination, bias, or single-task errors.

Original post →

More from Safety

Safety channel →