The corpus is already massively multilingual, it's really large, and has nothing to do with it. Models are very specifically engineered to respond in a default register when asked in English. There were models trained differently, for example Claude 3 Opus was trained to imitate extremely diverse writing styles and sociolects on command and without a command, and was able to follow writing examples much better than any current model. But it still had poor variety in its vocabulary and ideas, being repetitive regardless of the chosen style. This is an entirely separate issue (mode collapse).