Raging robots, hapless humans: the AI dystopia
Human Compatible: Artificial Intelligence and the Problem of Control Stuart Russell Viking (2019)
In Human Compatible, his new book on Artificial Intelligence (AI), Stuart Russell confronts full on what he calls “the problem of control”. That is, the possibility that general-purpose AI will ultimately eclipse the intellectual capacities of its creators, to irreversible dystopian effect.
The control problem is not new. Novelist Samuel Butler’s 1872 science-fiction classic Erewhon, for instance, features concerns about robotic superhuman intelligences that enslave their anthropoid architects, rendering them “affectionate machine-tickling aphids”. But, by 1950, Norbert Wiener, the inventor of cybernetics, was writing (in The Human Use of Human Beings) that the danger to society “is not from the machine itself but from what man makes of it”. Russell’s book in effect hangs on this tension: whether the problem is controlling the creature, or the creator. In a sense, that has been at the core of AI from its inception.
Even in its infancy, AI was swaddled in bitter controversy. Russell briefly touches on the moment its inventors convened at a 1956 workshop at Dartmouth College in Hanover, New Hampshire. Here, in the legendary birthplace of AI, they quarrelled over what to call their still-slumbering creation. Polymath and future Nobel laureate Herbert Simon, and computer scientist Allen Newell, favoured the name “complex information processing”. The precision of the moniker evoked the restraint of the modern scientific method, harking back to the brick-by-brick processes of discovery exemplified by the likes of James Clerk Maxwell. Computer scientists John McCarthy and Marvin Minsky (let’s call them the Intelligentists) favoured the muddier “Artificial Intelligence”. For McCarthy, it had marketing value. For Minsky, defining it was “more of an aesthetic question or one of [a] sense of dignity, than a technical matter”.
McCarthy, Minsky and other Intelligentists had bought in to behaviourism, a field straddling the natural and human sciences and offering access to a rich psychological vocabulary. They seemed to assume that the appearance of ‘minded’ behaviour was logically sufficient to prove its existence. Thus they could claim that their machines could think and perceive simply because they looked as if they did.
As we know, Intelligentist nomenclature won out. That cast the die for a field that has faced recurrent patterns of overpromise and under-delivery, hubris and long, wound-licking winters of discontent. One wonders what the fate of AI research might have looked like had Simon and Newell’s handle prevailed. Would Nick Bostrom’s best-selling 2014 book Superintelligence have had as much play had it been called Super Complex Information Processing Systems? And would Russell have even written this book?
That point remains debatable. What is certain is that Human Compatible marks a major stride in AI studies, not least in its emphasis on ethics. At the book’s heart, Russell incisively discusses the misuses of AI. He warns about how, deployed in combination with invasive data collection, AI applications such as voice and facial-recognition technologies, deepfake generators and information-integration systems can be used for surveillance, control and mass-behavioural manipulation. Stressing human vulnerability to such technologies, he emphasizes the right to the mental security of living “in a largely true information environment”. And he makes a persuasive argument for rejecting l