What they're saying, and I think this was the clear implication of the blog post too, is that the training data definitely would contain these chats and the only question is whether it got encoded into the weights.