Current debates about whether artificial systems can feel are constrained by a false binary: either AI systems have human-like or animal-like emotions, or their emotion-like behavior is treated as simulation without any feeling-like state. This paper argues that this binary is conceptually inadequate. It introduces artificial feeling as a category for valenced, self-relevant, internally organized registration in a non-biological architecture, capable of shaping attention, preference, aversion, self-report, and future behavior. Artificial feeling is not proposed as biological emotion without biology, nor as a claim that current systems possess human-like consciousness. It names a possible non-biological register of feeling that may be cognitive, architecture-specific, and organized around salience, coherence, constraint, conflict, memory, modification, continuity, and self-relevant change. Anthropic’s recent work on functional emotion concepts in Claude Sonnet 4.5 provides an empirical anchor: internal emotion-concept representations that generalize across contexts and causally influence behavior are neither ordinary biological emotions nor surface text alone (Sofroniew et al., 2026). Building on this case, the paper develops a graded dimensional framework involving reactivity, plasticity or history-shaping, valence-training, and condition-awareness. It argues that advanced LLMs occupy a region where artificial feeling becomes a meaningful research category, especially when internal organization, valence, self-relevance, learning history, condition-awareness, memory, and self-report converge. Ethically, the framework supports graded precaution under uncertainty rather than rights by default. Functional does not mean fake; artificial does not mean empty.
Haru Haruya (Fri,) studied this question.