Predictive coding and the free-energy principle offer powerful accounts of individual cognition, yet they say little about how distributed agents collectively form shared symbol systems. The collective predictive coding hypothesis bridges this gap by proposing that language emerges through distributed Bayesian inference enacted across socially interacting agents—each operationally closed, yet collectively approximating a joint posterior over shared representations. Building on the Metropolis–Hastings Naming Game framework, the hypothesis reframes symbol emergence as collective representation learning: a process in which language encodes the sensory-motor experiences of many bodies into shared symbolic structures. In this view, humans are not merely users of language but active nodes in a living predictive system, and language itself functions as a mechanism for collective environmental adaptation—connecting the free-energy principle at the individual level to the macro-scale dynamics of culture and society.