How to ruin Gmail's spam training model at scale (pretty sure this has been happening for a long time):

1. Harvest many millions of legitimate email addresses from the many leaked datasets over the years.
2. Automate subscribing legitimate email addresses to legitimate services all over the internet, whatever is easiest, do it at scale.
3. Produce billions of legitimate transactional emails that nobody wanted, some large subset of them are going to get marked as spam, tarnishing the training data.
4. Bonus: Carefully craft input strings (via signup name) to leverage association of specific strings with spam/not spam.

Result: Much less effective email spam detection.