Government-tailored versions of AI models with reduced safety constraints could support an infrastructure that’s more searchable, correlatable, and predictive, critics say.