Causal Auditing of Latent Affect in Language Models via Activation Steering

Large Language Models (LLMs) exhibit complex affective behaviors that are difficult to audit using input-level explanations alone. Prior work has shown that activation steering can modify model outputs by intervening on internal representations, but the causal specificity and reliability of such interventions remain underexplored. In this work, we present a mechanistic interpretability study that audits latent affective representations in a transformer-based language model via activation-level interventions. We identify an anger-related direction in the residual stream using contrastive activation analysis and perform causal interventions during inference to generate latent counterfactuals. While standard emotion classifiers and lexicon-based probes exhibit saturation or fail to capture meaningful changes, we find that activation steering produces systematic shifts in discourse structure, including sentence length, self-reference, repetition, and modal usage. To assess specificity, we introduce norm-matched random control interventions, showing that arbitrary perturbations induce unstable and incoherent effects, whereas concept-aligned directions yield more consistent and interpretable modulation. Our results suggest that latent affective representations primarily influence expressive structure rather than categorical emotion presence, and that internal interventions, while promising, exhibit important limitations for robust algorithmic recourse. This work highlights both the potential and the risks of using activation-level steering as a mechanism for auditing and modifying model behavior.