Automation + human classifiers + user reporting. The stuff automation gets wrong or isn't confident about goes into the training set and the system improves. E

Automation + human classifiers + user reporting. The stuff automation gets wrong or isn't confident about goes into the training set and the system improves. Even if they lose money by manually reviewing a small ad that was flagged as possibly political, it doesn't mean it costs more to review 1/100 low-volume ads. They're also paying to classify useful data to train their model.