> That works well for about two weeks until users figure out how to exploit this. They form groups that agree to coordinate to report content. Now reports become lower signal than before, but still somewhat useful. We use the reports, but try to limit exploitation.

What about human employees reviewing users reports, especially randomly picked samples from those that resulted in successful takedowns. If the content was found to be sound under the rules, penalise the reporters. Suspend their ability to make reports (probably in the form of silently ignoring them).

Ie some human moderation should still be done at this scale, but using a combo of random sampling and system-flagged-suspicious.