A new research paper from Hugging Face highlights the escalating risks that agentic AI systems—built on large language models (LLMs)—pose to human agency and autonomy. As these systems become deeply integrated across domains, their human-like cognitive patterns raise concerns that remain insufficiently studied.
The authors systematically analyze risks induced by expanding cognitive capabilities, following a three-level framework defined by cognitive scope: physical cognition, social cognition, and self-referential cognition. For each level, they examine potential risks to human agency, autonomy, and control capability.
The paper concludes by proposing mitigation strategies to enhance the controllability of agentic AI systems, aiming to ensure their long-term safe development. The work is part of a broader discussion on the societal implications of increasingly autonomous AI.