Ella,

your analogy between mind and machine is a bit misguided since it compares different levels. because transistors are VERY VERY reliable and fast deep learning algorithms reuse one to support billions of deep net linkages. Mother nature could never trust neurons in this way.

Still the interesting comparison is between deep net linkages and neural linkages. At this level you will find that both mind and machine share the same basic traits of reliability even when many units are destroyed or altered.

So even today our current algorithms share these very human attributes. Indeed it is merely an (big) engineering challenge to even replicate the lowest level of the deep net directly in hardware... that would mean even at the lowest level one could destroy parts and the whole would continue. But there is no need to do that, as the present algorithms behave exactly as these 'native' ones would.

And I don't share your certainty regarding AI being nice. Present day AI is not programmed, it is RAISED with lots of data and or interaction with the world.

Guaranteeing anything about the ultimate behavior of these systems seems as hard as guaranteeing you can conceinve and raise a human child in a way that guarantees they will never be a murder.

YOU CAN'T do the latter, and I fear you cannot do the former either.