Search papers, labs, and topics across Lattice.
This paper analyzes the optimal balance between character shaping and rule enforcement in AI safety systems as deployment scales increase, using a stylized comparative-statics model. The findings reveal that while the optimal allocation between these approaches shifts slightly toward character shaping with larger scales, the most significant factor influencing this balance is the baseline character fragility rate. The study concludes that safety architecture choices are more dependent on the reliability of character shaping under distributional shifts than on the scale of deployment itself.
The optimal balance between character shaping and rule enforcement in AI safety systems is more influenced by character fragility than deployment scale, challenging conventional wisdom.
Artificial Intelligence (AI) safety systems combine character shaping (e.g., Reinforcement Learning from Human Feedback [RLHF], Constitutional AI), which modifies behavioral distributions at training time, with rule enforcement (e.g., output filters, safety classifiers), which blocks harmful outputs at inference time, yet little formal analysis exists on how their optimal balance should change as deployment scales increase. We introduce a stylized comparative-statics model that parameterizes safety design as a resource allocation alpha in [0,1] between these two approaches, incorporating scale-dependent filter degradation, common-mode failures, and character fragility -- the risk that shaped behavior degrades or collapses under novel conditions. Under a multiplicative Pareto damage model, we derive closed-form expected harm and supplement it with tail-risk (CVaR) analysis via Monte Carlo simulation. Across three scenarios (optimistic, moderate, pessimistic), the optimal alpha* is interior or at the rules-only boundary and shifts weakly toward character shaping as deployment scale T grows, from negligible (Delta alpha* = +0.01) to pronounced (Delta alpha* = +0.21) depending on scenario. The dominant parameter is the baseline character fragility rate p^(0)_frag, which shifts alpha* by 0.50 across its range -- far exceeding the effect of tail severity, filter quality, or common-mode failure probability. CVaR and expected-harm optima converge at large T. These results suggest that safety architecture decisions depend less on deployment scale per se than on the reliability of character shaping under distributional shift.
