Originally published October 15, 2017
Here is an excerpt:
However machine learning throws up problems of its own. One is that the machine may learn the wrong lessons. To give a related example, machines that learn language from mimicking humans have been shown to import various biases. Male and female names have different associations. The machine may come to believe that a John or Fred is more suitable to be a scientist than a Joanna or Fiona. We would need to be alert to these biases, and to try to combat them.
A yet more fundamental challenge is that if the machine evolves through a learning process we may be unable to predict how it will behave in the future; we may not even understand how it reaches its decisions. This is an unsettling possibility, especially if robots are making crucial choices about our lives. A partial solution might be to insist that if things do go wrong, we have a way to audit the code - a way of scrutinising what's happened. Since it would be both silly and unsatisfactory to hold the robot responsible for an action (what's the point of punishing a robot?), a further judgement would have to be made about who was morally and legally culpable for a robot's bad actions.
One big advantage of robots is that they will behave consistently. They will operate in the same way in similar situations. The autonomous weapon won't make bad choices because it is angry. The autonomous car won't get drunk, or tired, it won't shout at the kids on the back seat. Around the world, more than a million people are killed in car accidents each year - most by human error. Reducing those numbers is a big prize.
The article is here.