Well, OpenAI said "we didn't read the conversations", but they never discarded that the model was training with that data... so even worse.