Study Analyzes Cognition-Induced Risks in Agentic AI Systems

A new academic paper from arXiv (reference 2608.15304) systematically analyzes risks induced by expanding cognitive capabilities in frontier agentic systems powered by large language models (LLMs). The authors propose a three-level framework based on cognitive scope: physical cognition, social cognition, and self-referential cognition. They examine potential risks to human agency, autonomy, and control capability at each level. The paper also proposes strategies to mitigate these risks and enhance controllability of agentic AI systems for long-term safe development. The study addresses a gap in research on cognitive engagement of AI systems as they integrate across domains. The paper is categorized under Computer Science > Artificial Intelligence and was submitted to arXiv. It includes references, citations, and tools for bibliographic management. The authors emphasize the need for systematic study of risks from human-like cognition in AI. The framework aims to guide future research and policy. The paper is available on arXiv and includes experimental features like arXivLabs for community collaboration.