Comment by Chance-Device

Please answer this question: do you or do you not train your models on anonymized user data, where those users have opted out of such training?

The blog post appears to imply the answer to this is yes, as otherwise I assume it would be impossible for this contamination to have happened.