When AIs are making life and death decisions (because there are only microseconds to act), will they hold their fire in the name of humanitarianism? Would an AI have launched the missiles?
You've misunderstood the story. Petrov didn't defy high-level Soviet goals in the name of humanitarianism: the goal was to fire back if fired upon and _not_ to fire back if not fired upon. Petrov's achievement was making the right judgment call that the US had not actually fired on Russia (i.e. that it was a detection anomaly or something). Given that he wasn't deviating from the high-level goals handed down, I don't see why an AI would be different in that regard.
Correct, it's actually likely that a good AI would also conclude, just like Petrov did, that this is not an attack, given everything that was expected from a real attack at the time. If they simply wired a launch detector to the button that launches nukes, then that's not AI.
