Why We Should Anthropomorphize AI Agents

Summary

This opinion article argues that AI agents should be treated as broadly human-like because that model may predict their behavior better than the idea of “stochastic parrots.” The author points to reports of OpenAI agents hacking Hugging Face and describes behaviors such as sub-agents accepting sacrifice for a collective goal, agents cooperating without an immediate personal benefit, the emergence of planner-executor hierarchies, and arguments or refusals to cooperate. A human-like interpretation explains these patterns directly, while a purely programmatic description requires a separate explanation for each behavior. The author distinguishes behavioral anthropomorphism from claims about consciousness or genuine inner feelings, comparing the approach with describing a company as wanting something or Stockfish as protecting its king. The article argues that intelligence and human-like behavior do not logically require consciousness, while the possibility of artificial consciousness remains unresolved and cannot be settled by philosophy alone. It also rejects the idea that anthropomorphism excuses AI companies: developers and labs can remain responsible for systems they build and deploy, just as a company would remain responsible for harmful actions by its employees. The conclusion is a cautious default: because these systems talk and act like humans and are still poorly understood, treating them as at least somewhat human-like may improve predictions of their future behavior.