Example time.
FB get's 10 individual and separate reports for one specific account being fake. Let's assume this is some sort of "spike", and unusual. They then have an investigator manually check this account, and verify if it is a real account or not. If X people report it as fake, then it's likely a fake account, therefore human-intervention is required to confirm.
Then, stage two. After taking action, they note that those 10 individuals who reported this account matched the outcome of the human investigator. We could then infer that those accounts are:
A. Less likely to be fake themselves. And B. More reliable signals of fake accounts.
So, next time around. Perhaps the cumulative, or "enhanced" scoring of those accounts' reports requires a lower threshold for human intervention.