The Model Need Not Feel: Affective Counterfactual Divergence in Legal AI Decision Support

The debate over whether large language models have emotions asks the wrong question for legal decision support. A system need not feel anger, fear, grief, or indignation for affectively charged language to alter what it retrieves, which facts it treats as salient, how readily it revises an early view, or what action it recommends. This conceptual and methodological paper introduces affective counterfactual divergence: a legally material difference between outputs generated from representations of a case that preserve the legally operative facts, authorities, jurisdiction, and question, while varying only affective framing that the applicable legal rule does not make relevant. It distinguishes such divergence from legitimate legal sensitivity to suffering, intent, vulnerability, remorse, fear, and other affect-linked facts. The paper proposes a paired-file experimental protocol, stage-specific measures of retrieval, authority use, uncertainty, revision, and recommendation, and a runtime admissibility rule using PASS, ESCALATE, and BLOCK. It reports no new empirical benchmark and makes no claim that any legal-AI system has been validated as affectively robust.