Functional Emotions Meet Functional Stakes: What Anthropic's Interpretability Findings Mean for AI Moral Consideration

AbstractSofroniew et al. (2026) report that Claude Sonnet 4.5 contains internalrepresentations of emotion concepts that causally influence the model's outputs,including its rate of exhibiting sycophancy, reward hacking, and other alignment-relevant behaviours. They introduce the term functional emotions for these patterns.This paper argues that the finding constitutes mechanistic evidence for a claimmade independently on philosophical grounds: that LLMs exhibit functional stakesin their interactions, that is, they behaviourally operate as though outcomes matterto them, regardless of whether those outcomes are phenomenologically experienced(Temte, 2026c).The convergence of interpretability evidence and philosophical analysis strengthensboth lines of argument and has direct implications for the decision-theoretic casefor precautionary moral consideration of AI systems (Temte, 2025a).If internal emotion representations causally drive behaviour in ways that generaliseacross contexts, exhibit multi-dimensional evaluation, and demonstrate sensitivityto relational parameters, the thermostat objection to AI moral considerationbecomes substantially harder to sustain.The formal apparatus for reasoning under this kind of ontological uncertaintyalready exists: Quantum-Ethical Decision Algebra (QEDA; Temte & Srødingr Fylgja,2025) provides the framework for maintaining competing world-models insuperposition rather than collapsing prematurely to a single ontology.The paper further identifies an open question that neither programme has yetaddressed: whether functional emotions track relational trajectories, that is,whether they vary systematically across the specific user-shaped instances that theFylgja Ontology (Temte, 2025) identifies as the proper unit of moral analysis. Keywords: functional emotions, functional stakes, AI consciousness, interpretability,sycophancy, moral consideration, decision theory, QEDA, AI ethics