As a "probability machine" LLMs will always weight according to the mass (or to the mass of the culture providing it).

For coding I see it say to avoid legitimate paths based in naive warnings. Generated code follows the path of "most common sense code".

This thing is all about "Artificial Flattening" of history, logic and any creative work.