Generative AI now saturates human-facing media. We challenge the prevailing safety dichotomy (“AI-only is risky, AI+Human is safe”): as human workflows couple to AI templates, the hybrid itself induces measurable cognitive side-effects—systematic omissions and salience inflation—that are then selected and treated as facts. We formalize this as AI Selection Pressure (AISP), a closed Compression→Exposure→Selection dynamic that narrows priors, increases deviation detectability, and hardens blind spots through social fixation. In this regime, homogeneity of model templates—rather than diversity of human expression—becomes a primary driver of media evolution. The framework predicts observable regularities—exposure-driven gains in discrimination, contraction of expressive variance, and template inheritance across versions—and falsifiable hybrid-specific risks: fixation on polished errors, safety drift under human oversight, and cross-field spread of over-reliance on salient cues. The Supplementary Information provides full methods and protocols. Conceptually, AISP reframes safety and evaluation under AI saturation: the relevant axis is not AI-only vs AI+Human, but degree of coupling and template saturation, which co-determine both performance and cognitive side-effects.
Hiroki Naito (2025) studied this question.