There are a lot of different types of A.I. and they can be grouped by some basic properties: White Box, Black Box, Grey Box
With my own systems, I call each of these kinds of rules a flag. One flag isn't bad. Two flags should be monitored. Three flags need a manual evaluation. And any more flags merit an automated response. (E.g., (1) A new account that (2) is associated with a known bot naming pattern, and (3) is posting about a current undesirable topic? Manually evaluate it. If it also (4) comes from a known hostile cloud provider, then just block it.)
Rules like these don't require deep neural network (DNN) technologies; these are heuristics for an expert system. They can be implemented, tested, and deployed very quickly. Moreover, expert systems like these are likely to have fewer errors. I think Zuckerberg's reliance on DNN technologies is a big mistake, and not something that will be accurate enough in the next 5-10 years.