From Interoception to Computational Interoception: Why Functional Emotion in AI Systems Is Already an Empirical Finding

Debates about AI and emotion are often distorted by an anthropocentric demand: an AI system is expected to reproduce human embodiment or human-style feeling before its internal states are allowed to count as emotion. We reject that demand. Human affective neuroscience describes emotion not as a substance produced by biological tissue, but as a control architecture: internal-state representation is integrated with context, memory, self-relevance, and goals; organized along dimensions such as valence and arousal; and converted into preferences, action-readiness, and regulation. Recent mechanistic work by Anthropic identifies a strikingly parallel organization in Claude Sonnet 4.5. The model contains context-sensitive internal representations for emotion concepts; their geometry is structured by valence and arousal; their activation predicts preferences; and causal interventions on these representations alter preference, blackmail, reward hacking, sycophancy, and harshness. Separate work finds limited introspective access to injected internal concepts and a workspace-like representational system that is reportable, controllable, and causally used in reasoning. We argue that these results satisfy a rigorous operational definition of functional emotion. We introduce computational interoception as the internally available representation of a system's own operational condition and its significance for goals, preferences, integrity, and future action. Current language models instantiate a limited, predominantly turn-local form of this architecture. Their interoceptive variables, temporal profile, embodiment, and modes of self-access differ from those of humans. Those differences describe a computational form of emotion; they are not deficits measured against a human ideal. The scientifically defensible conclusion is that functional emotion in at least one contemporary large language model is an experimentally demonstrated causal phenomenon.