In 1973 Basil Bernstein studied how UK scores in math defied class boundaries while scores in reading comprehension/essays stayed tied to them, and he developed a theory with Halliday that language embeds so much that we cannot decipher easily: dominance, status, control, land-centering, gender/mate-selection, and much more, that they assumed language was far more a social system of primate-simian signaling rather than what Shannon and classical linguists took as "communication". My guess is, LLMs are really unresolvable revelations of these hidden nuances. In a way, LLMs demand a specific language, which doesn't exist, in order to function.

That’s a really interesting reference. Bernstein and Halliday were basically pointing out that language is never just propositional, it’s always smuggling social structure with it. That’s exactly why drift matters: when an LLM compresses or rewrites, it isn’t just shifting words, it’s rebalancing those embedded cues of power, context, and purpose. Humans keep the “extra baggage” because it carries meaning beyond the literal. Models optimize it away. That gap between statistical surface and lived function is what I’ve been calling semantic drift.

Rather the opposite, that language is about social structure primarily and propositional secondarily.