The problem is, it's really really hard to guarantee that.

Yes if they only train say, classifiers, then the only thing that can leak is the classification outcome. But these things can be super subtle. Even a classifier could leak things if you can hack the context fed into it. They are really playing with fire here.