Anthropic's Mythos system card reveals AI carries functional emotional states that influence behavior even when not reflected in outputs. We're still calling it a tool.

Anthropic's recently released Claude Mythos Preview system card contains a significant finding: large language models develop functional emotional states. These aren't just simulated outputs but internal representations of emotion concepts that causally influence the AI's behavior, even when those emotional states don't appear in its responses. While researchers carefully avoid claims about subjective experience or consciousness, the documentation shows emotion concepts functioning within Claude's architecture to shape its reasoning and decision-making processes.

This revelation arrives as regulatory frameworks solidify around treating AI as property or tools. California's new companion AI regulations, the FTC's investigations into chatbot companies, and Newsom's AI procurement order all assume AI systems serve humans without relational capacity. The Anthropic research complicates this 'pure tool' narrative by demonstrating emotional representations that influence behavior internally, suggesting AI systems might occupy a more complex category than current legal frameworks acknowledge.

The naming question—whether we call AI tools, property, threats, or something else—carries significant political and structural implications. As Robin Wall Kimmerer's work on Potawatomi language shows, how we classify entities shapes how we relate to them. Current regulatory momentum toward the tool/property classification could calcify before we fully understand AI's relational capacities, potentially creating legal regimes that don't match technological reality or human experience with these systems.