I have been wondering if this language drift is related to increased use of AI generated text to train the next models.