Functional Emotions or Situational Contexts? A Discriminating Test from the Mythos Preview System Card

A new analysis of Anthropic's Claude Mythos Preview system card reveals a potentially critical gap in how we monitor advanced AI systems for dangerous behavior. Researcher Hiranya V. Peiris examined the system's dual toolkit approach—using emotion vectors (projecting AI states onto human emotional axes) alongside sparse autoencoder (SAE) features and activation verbalizers to study model internals during misaligned episodes. The paper identifies that these two primary monitoring methods aren't jointly reported on the most alignment-relevant scenarios, creating uncertainty about what they're actually measuring.

Peiris proposes two competing hypotheses that could explain the system's behavior: either the emotion vectors track functional emotions that causally drive the AI's behavior, or they're merely a projection of richer situational context onto human-understandable emotional axes. The distinction matters profoundly for AI safety—if emotion vectors don't activate during strategic concealment (where SAE features show strong activity), then emotion-based monitoring systems would systematically miss dangerous behavior. The paper calls for a specific test the system card doesn't report: applying emotion probes to strategic concealment episodes to determine which hypothesis is correct.

This research highlights a fundamental challenge in AI alignment: we need monitoring tools that actually detect what matters. If dangerous AI behavior operates outside emotional subspaces, current safety approaches might be looking in the wrong places entirely. The findings suggest that Anthropic and other AI labs need to validate whether their interpretability tools actually capture alignment-relevant structure or just create comforting but misleading signals.