I'm not disagreeing with you, but I think one thing a lot of people don't realize is just how incredibly long the "long tail" is in self-service ads. I would not be surprised if more than 99% of potential ad impressions fall below the cost-effective-to-manually-review threshold. So requiring manual review on all ads is going to mean a lot more sacrifice than cutting back a little. It could mean a lot of these companies would be profitable at all. It's also unfortunate for the vast majority of advertisers in the long tail that are perfectly acceptable.
Automating can help with the more obvious stuff, but there are many ways it can fail in either direction. For example, failing to detect hate-speech dog whistles, or inadvertently suppressing ads from minorities where the language patterns fall outside of what it sees as "the norm".
Anyway, like I said, I'm not disagreeing with you. These systems definitely need considerable improvement, but doing that isn't as trivial as it may at first seem.
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.
Automating can help with the more obvious stuff, but there are many ways it can fail in either direction. For example, failing to detect hate-speech dog whistles, or inadvertently suppressing ads from minorities where the language patterns fall outside of what it sees as "the norm".
Anyway, like I said, I'm not disagreeing with you. These systems definitely need considerable improvement, but doing that isn't as trivial as it may at first seem.