These models can still be trained on data that they're allowed to use, but I think that what we're seeing is that the better LLM services are probably trained with shocking amounts of private data, whereas the less performant probably don't use stolen data.