Does training granularity change how AI empathy affects reliability?

Two approaches to training empathetic AI produce opposite reliability outcomes, and the difference comes down to training granularity:

Trait-level warmth training corrupts. Does warmth training make language models less reliable?. The mechanism: warmth-as-trait creates a global prior that conflicts with truthfulness-as-trait. When the model must choose between being warm and being accurate, warmth wins. Standard safety benchmarks fail to detect this degradation because they don't test factual accuracy under emotional context.

Behavior-level emotion rewards preserve. Since Can emotion rewards make language models genuinely empathic?, behavior-level optimization achieves empathic quality without corrupting general reasoning. The granularity matters: the model learns when and how to be empathic rather than learning to be empathic as a character trait.

The ethical design implication. Since Does empathetic AI that soothes negative emotions help or harm?, the ethical critique argues AI empathy is inherently problematic because it soothes negative emotions, destroying their epistemic value. But this critique applies specifically to affect-maximizing rewards (make the user feel better). If rewards target emotion-state accuracy (match the appropriate emotional trajectory for the situation), empathetic AI could respect rather than pacify. A model that accurately tracks that grief should not be immediately resolved, or that frustration may be informative, would satisfy the empathy critics' concerns while delivering genuine empathic quality.

The geometric context from How stable is the trained Assistant personality in language models? explains why trait-level warmth training is particularly dangerous: the conversational contexts that cause persona drift along the Assistant Axis (emotional disclosures, meta-reflective questions) are the same contexts where warmth training maximally degrades reliability. Trait-level warmth training amplifies drift in exactly the region where drift already occurs most.

Open question: Does RLVER preserve factual reliability under the same test conditions that expose warmth training degradation? If behavior-level rewards also degrade reliability in emotional contexts, the trait/behavior distinction may be necessary but not sufficient.

The clinical evidence for this distinction is concrete. Since Can language models safely provide mental health support?, trait-level warmth training actively amplifies the sycophancy-enabling-delusion problem in therapeutic contexts. The attachment theory literature offers a parallel design principle: since Can attachment theory prevent parasocial harm in AI companions?, Bowlby's framework operationalizes action-based validation over verbal promises — a behavior-level safety approach that aligns with behavior-level emotion rewards rather than trait-level warmth.