As I understand it, that setting does not prevent them training on user data, just which derivatives are used (i.e. just PII scrubbed vs certain types of synthetic summarization)