AFAIK for such decisions human information/intelligence is required. And there are often screw ups/bad decisions. No need for bad machine learning algorithms for this.
Your optimism is misplaced. The USA routinely drone strikes people they have no intelligence on whatsoever on the basis of nothing more than 'they acted like a terrorist', these are called signature strikes, i.e. they matched the "signature" of a terrorist.
Organisations that do this so frequently there's jargon for it are absolutely not going to be slowed down by requiring human intelligence (which is itself full of false positives and duplicity, see how random people were sold to US soldiers as "terrorists" to collect the reward money).
Organisations that do this so frequently there's jargon for it are absolutely not going to be slowed down by requiring human intelligence (which is itself full of false positives and duplicity, see how random people were sold to US soldiers as "terrorists" to collect the reward money).