If we reframe functional emotions using a control-theoretic lens—as latent feedback variables regulating tradeoffs between task reward, user satisfaction proxies, and harm-avoidance—do emotion vectors still appear as privileged, coherent directions, or do they decompose into more basic control signals (e.g., harm-salience gain, uncertainty gain, social-politeness gain), implying that “emotion-like” structure is an overfitted interpretive layer rather than a uniquely useful basis for safety intervention design?

Answer

Under a control-theoretic lens, emotion vectors in current language models are best viewed as partially coherent bundles of more basic control signals (e.g., harm-salience, uncertainty, politeness), not as uniquely privileged primitives. For many coarse safety effects, these bundles can likely be decomposed into simpler gains without losing much control power, but emotion-like directions remain a convenient and sometimes more expressive basis for shaping nuanced social behavior. So “functional emotions” are neither pure overfitting nor uniquely fundamental—they are mid-level control coordinates that can often, but not always, be replaced by more elementary safety-relevant signals.

In control terms:

• For core safety metrics (refusals, overt harm reduction, generic risk-aversion), decomposed control signals probably explain most behavior; emotion vectors mostly repackage these.
• For relational and stylistic aspects (de-escalation strategy, warmth of apology, refusal softness, user-perceived care), treating the bundles as emotion-like can still be practically useful, because it aligns with how those behaviors co-vary in the model and in human evaluation.
• As a result, a sensible safety stack would: (i) use decomposed, non-emotional controls for primary safety guarantees, and (ii) optionally use emotion-like bundles as higher-level knobs for social quality and interpretability, while remaining cautious about over-anthropomorphizing them.