Evolved AI Agents Use Self-Talk for Behavioral Regulation

A new paper from Joshua Nunley (arXiv:2606.02840) challenges the traditional view of communication as purely information transfer. In a minimal predator avoidance task, pairs of evolved Continuous-Time Recurrent Neural Network (CTRNN) agents developed three distinct communication strategies across over 2,000 evolutionary runs: safety calling (39% of agents signal from safe cover), alarm indication (22% vocalize when a threat is present but don't rely on self-hearing), and self-regulatory calling (20% depend on hearing their own call to sustain escape behavior). The study found that 47% of agents calling during an active threat depend on self-hearing, compared to only 10% of those calling after reaching safety.

This asymmetry reveals a fundamental difference in causal order: safety callers act first then communicate, while self-regulatory callers communicate in order to act. When self-hearing was removed, self-regulatory callers' fitness dropped to 0.40 (p<10^-9), while safety callers remained functional at 0.90. These findings demonstrate that communication can evolve to serve the caller's own behavioral regulation—talking to oneself to sustain action—rather than merely signaling to others. The work has implications for understanding the evolution of language, animal communication, and designing autonomous systems that use internal signaling for coordination.