Correlated Intelligence: AI Adoption and Systemic Risk

Banks delegate decisions to AI models from a common vendor frontier that bundles accuracy with error correlation: the most accurate models share lineage and fail together. The cross-bank cost of that correlation remains unpriced even when lineage is observable, because a credit spread reprices the borrower, not the banks its choice endangers. Model diversity is inefficiently low; under pooled pricing, welfare losses grow quadratically with excess adoption. Frontier improvements can reduce welfare by tipping the market into monoculture. Disclosure makes a correlation surcharge feasible but does not substitute for it.