AI Transmission Kernels: Modelling Generative Systems as Operators in Human Cultural Transmission

Abstract

Generative AI is commonly treated as a source of information or an intervention applied to an individual. This paper proposes a different formal object: the AI transmission kernel, defined as the conditional probability that a particular model, language and interaction configuration transforms one human epistemic variant into another during transmission. The kernel places the generative system inside the mechanism of cultural evolution rather than treating it as an exogenous exposure. It permits direct comparison between human-only, shared-model, heterogeneous-model and diversity-preserving transmission environments and separates three effects that are otherwise conflated: preservation of existing variants, directional transformation towards dominant variants, and innovation of new defensible variants. The paper outlines how kernels can be estimated from iterated human transmission chains using structurally coded explanations rather than lexical similarity, how their parameters can be linked to delayed transfer and minority-variant survival, and how model concentration changes population dynamics when the kernel is repeatedly applied. Competing explanations include simple copying, task-induced convergence and social influence. The proposal is falsifiable: if estimated kernels do not predict held-out transmission trajectories better than human-only transition models, generative AI has not been shown to act as a distinct selective operator. The framework offers a bridge between human–AI cognition and formal cultural-evolution models.