He didn't advocate for fully anonymous AI systems, but ones where humans will ultimately make the decision. Obviously these can be implemented in a variety of ways, from basically data mining software and some neat dashboards to help a bunch of humans decide on a target in a responsible way, to something where all the boring bits are hidden, a list of names is oh-so-helpfully drawn up and a rubber-stamp from an under-trained early-20s "intelligence officer" is all that's needed to kill someone, and their family, in their home.