Ask HN: How to avoid sensitive data being part of LLM training data?

How are you making sure your sensitive data & PII doesn't become part of LLM training data. When training data is small, manual verification is possible. But when data size is huge, is there a way to mask/filter out PII/sensitive data.