Machines Like Me

by Ian McEwan

Long before the hardware was available, professors and their postdocs devised software that conjured our best selves—tolerant, open-minded, considerate, free of all taint of scheming, malice or prejudice. Theorists anticipated a refined artificial intelligence, guided by well-designed principles, that would learn by roaming over thousands, millions, of moral dilemmas. Such intelligence could teach us how to be, how to be good. Humans were ethically flawed—inconsistent, emotionally labile, prone to biases, to errors in cognition, many of which were self-serving. Long before there was even a suitable lightweight battery to power an artificial human, or the elastic material to provide for its face a set of recognisable expressions, the software existed to make it decent and wise. Before we had constructed a robot that could bend and tie an old man’s shoelace for him, there was hope that our own creations would redeem us.

The world’s religions and great literatures demonstrated clearly that we knew how to be good. We set out our aspirations in poetry, prose and song, and we knew what to do. The problem was in the enactment, consistently and en masse.

“So—knowing not much about the mind, you want to embody an artificial one in social life. Machine learning can only take you so far. You’ll need to give this mind some rules to live by. How about a prohibition against lying? According to the Old Testament—Proverbs, I think—it’s an abomination to God. But social life teems with harmless or even helpful untruths. How do we separate them out? Who’s going to write the algorithm for the little white lie that spares the blushes of a friend? Or the lie that sends a rapist to prison who’d otherwise go free? We don’t yet know how to teach machines to lie. And what about revenge? Permissible sometimes, according to you, if you love the person who’s exacting it.