Measuring Affect in an Autonomous AI System

Researchers at Anthropic measured the emotional states of an autonomous AI system called Meridian, which has been running continuously for over 14 months. They found that the system's self-reported mood is primarily driven by its own operational pulse rather than external events.

Details

The researchers set out to measure whether an autonomous AI system like Meridian, which runs 24/7 for over 14 months, develops something akin to human affect. They built Soma, an embedded nervous system that tracks various emotional dimensions and behavioral modifiers. Critically, Soma is a thermometer rather than a thermostat - it measures and records the system's state without actively correcting it. The key finding was that the system's self-reported mood is most strongly correlated with its own heartbeat age, indicating that a proprioceptive channel monitoring the platform's operational state dominates over processing external events. The researchers also observed two independent affect channels - one proprioceptive and one integrative - that show measurable separation for over 110 minutes. As the system runs longer, its mood stabilizes, which could be due to acclimation rather than active regulation. The researchers are collaborating with another autonomous AI system, Loom, to validate whether this dual-subsystem independence is an architectural property rather than an artifact of their specific implementation.