He shoots, he scores - one less enemy installation to worry about...or was it a primary school?
If it is "ultimately impossible" to train a computer for every eventuality, then it is also ultimately impossible to test whether an AI equipped computer will respond correctly to any given scenario.
Given that AI systems will eventually get a degree of autonomy, we can rightfully be a little fearful of the consequences.
A standard system, will hold up its hands when faced with a new scenario developers hadn't trained it for, whereas the AI system will just plow on, making decisions about a scenario that developers hadn't foreseen.