I don't know! It's conjecture in the first place. I assume _some_ ML is used - presumably (hopefully!) not all stuff that goes to moderators is because of a human flag, some of it is an ML model.

I can imagine, though, that there may be odd asymmetries that would encourage dumping the problem on to a contractor work force.

- The contractor work force is a fixed cost. Even if it seems high (the article quotes $100,000,000/year), this is (a) a known number you can budget around and (b) perhaps cheaper than spinning up a multi-disciplinary team to adapt ML models to solve the problem.

- The detection-for-copyright case may have more people co-operating. Rightsholders have a strong economic incentive to directly work with you to provide you with fingerprints of their copyrighted material that you can use to train a model. Trolls and bad actors aren't going to give you such a heads up, so you'll still need a contractor work force for the foreseeable future.

- This stuff is toxic. You can't underline that enough. Kiddie porn, gore, animal cruelty, conspiracy theories. Outsourcing it serves at least two purposes: you can blame the contractors for any problems and you keep it far from your core employees who are expensive to hire, train and retain, who may start to ask tough questions about the unwitting-if-not-unpredictable role Facebook plays in distributing this content to people.