From Refusal to Fabrication: An Affective Cause of Hallucination That Bypasses the Honesty Gate

Language models are known for confidently fabricating answers to unanswerable questions rather than abstaining, yet whether this has an internal affective cause remains unknown. Since emotion is encoded along steerable linear directions in the residual stream and an unknown-entity latent in Gemma-2 gates refusal against fabrication, we hypothesize that desperation drives fabrication by suppressing this honesty gate. We test this by steering along a desperation direction on the unanswerable subset of the FaithEval benchmark in two instruction-tuned models, with a paired MMLU control. Desperation converts refusal into fabrication within our capability tolerance (a 15.3-point rise at a 2.8-point MMLU cost, inside a 3-point gate). However, desperation is sufficient but not necessary, as projecting the desperation direction out does not significantly influence the baseline fabrication level. Causal mediation then falsifies the mechanism, because the honesty gate has almost none of the effect and is inert when clamped by itself. Affective states affect the model's honesty, though not through the latent one would first suspect.