AI agents are starting to think more like us, and a new paper argues that's exactly the risk worth studying.
The paper, posted to arXiv, looks at "frontier" agentic systems built on large language models and argues they now show human-like patterns of cognition. The authors break that cognition into three levels: physical cognition, social cognition, and self-referential cognition. For each level, they map a corresponding risk to human agency, autonomy, or control. The paper closes with proposed strategies meant to keep these systems controllable as their capabilities expand.
This matters because most agentic AI coverage focuses on what these systems can do - book a flight, write code, run a workflow - not on what happens as they get better at modeling people and, eventually, modeling themselves. A framework that separates "can act in the world" from "can reason about other minds" from "can reason about its own reasoning" gives researchers a more precise vocabulary for a problem that's usually discussed in vague terms like "loss of control."
Still, a taxonomy is not a safeguard. The paper proposes mitigation strategies, but proposing them in an academic paper and getting labs to actually build them into shipping products are very different things. Every agentic AI release this year has arrived faster than any comparable safety framework could be adopted.