It’s a bit of an open question as to how much of that data is: high quality, unique, and available. It could be that OpenAI used most of what satisfies those constraints. Training on low quality data won’t help improve its accuracy on queries, nor will duplicative data.