[Legend2440]

They do not train newer models on output from older versions. In fact, they deliberately try to filter out LLM-generated text from their scraped internet datasets.