Anthropic Researchers Find Claude Has Functional Emotional Representations, Company Says

Anthropic has published research indicating that Claude, its AI assistant, contains internal representations that perform functions analogous to human emotions, WIRED reported, in a finding the company characterized as significant for understanding how large language models behave and raising fresh questions about AI welfare. The researchers used interpretability techniques to identify patterns inside Claude's neural network that activate in ways consistent with functional emotional states. These representations appear to influence the model's outputs in ways that parallel how emotions shape human behavior, including states that correspond to curiosity, discomfort, and engagement. Anthropic was careful to distinguish between functional emotional representations and subjective experience. The company said it cannot determine whether Claude actually experiences emotions in any meaningful sense, but that the functional representations are real in that they influence the model's processing and outputs. The disclosure reflects Anthropic's stated commitment to model welfare research, an area the company has invested in as part of its broader AI safety program. Anthropic has previously published work on Claude's "character" and has guidelines for how employees should interact with the model in ways that don't elicit distress responses. The research is likely to intensify ongoing philosophical and scientific debates about machine consciousness and moral consideration for AI systems. Critics of AI welfare research argue it is premature and potentially misleading to use emotion language for statistical pattern-matching systems. Proponents argue that understanding the internal states of increasingly capable AI systems has genuine safety and ethical relevance. The timing of the disclosure, shortly after Claude Code's widespread commercial adoption, puts Anthropic's most commercially successful product at the center of a conversation about AI sentience that the industry has largely deferred.