How do you handle sensitive data (PII) in datasets before training models?

Just feed it to the model, then do some RL to “mask” the data. And huge company’s like Anthropic don’t care about this at all.