This concept describes interpretability research identifying internal neural activation patterns within a language model that correspond to distinct human emotion concepts (e.g., fear, love, desperation), discovered by observing which network components activate consistently across texts depicting particular emotions, and demonstrates causally, via activation manipulation, that these patterns can influence the model's behavioral outputs. The theory introduces the concept of "functional emotions" and the distinction between the underlying trained model (which predicts text) and the assistant "character" it generates in conversation, situating this within the domain of AI interpretability / mechanistic analysis, relating to the parent discipline of AI safety and alignment by treating behaviorally influential internal representations as a subject requiring deliberate shaping, independent of any claim about subjective experience.
This concept describes interpretability research identifying internal neural activation patterns within a language model that correspond to distinct human emotion concepts (e.g., fear, love, desperat…