Transcendental model selection: a computational account of symbolic cognition and general intelligence through morality and culture.

General intelligence enables flexible problem solving across diverse contexts by minimizing uncertainty. Symbolic systems such as language extend this capacity, allowing humans to build social groups and construct world models beyond typical biological constraints. Previous research on linguistic communication within active inference has emphasized deep hierarchical models that ensure shared semantics between communicators. We argue that these models, while powerful, require extension to account for symbolic genesis, specifically using morality not only as uncertainty minimization across cultural niches, but also as the mechanism that created the virtual space enabling symbolic cognition. Our ancestors transcended dyadic modeling by implementing cultural layers through novel model selection, enabling in-group signaling and hierarchical social organization through psychological typing. This rendered the generative process endogenous (self-referential). Our emotional and impulsive tendency toward morality, we argue, enabled the deeper level of abstraction and the stable third-party triangulated perspective necessary for symbolic thought. This framework can be evaluated through simulations similar to recent active inference literature and provides a foundation for building generally intelligent systems aligned with human cultural values.