25 October 2013

The Squirrels and the Rats of AI

Hazelnuts reDone
Image ©Mark Tomlinson
Concerned about viruses, porn, and spam choking up the internet? Just wait until the AI rats get to work.

As bioethicist James Hughes has pointed out, we should not assume that malign forms of artificial intelligence that may emerge would be of the UFAI (unfriendly artificial intelligence) variety. ‘Unfriendliness’ implies ill will, which in turn implies human-level or greater intelligence. In contrast, AI rats would be neither friendly nor unfriendly, but they could do a great deal of damage with their relentless digital scavenging.

Don’t such rats already plague us? Aren’t viruses gnawing at and infecting our IT systems now? Yes, but computer viruses are not intelligent. Rat-level AI is still some way off. The problem-solving capabilities of even the tiniest of rodents are a source of both inspiration and frustration to neuroscientists and AI researchers; rats have drives, and like us, they employ sophisticated behaviours to satisfy those drives. In contrast, computer viruses are mere algorithms lacking any form of desire or intent.

How would AI rats emerge? ‘Weak AI’ systems already abound: air-traffic control systems, vehicle engine-management systems, big-data language translation systems, ‘expert systems’, and so on. It is possible that weak AIs augmented to perform ever-more specialised roles could be elevated, accidentally (or maliciously), to a semi-sentient or sentient – but not sapient – level. Without sapient cognition, these entities would have no reasons or ability to attempt to communicate with us. Initially, they would perform the roles for which they were originally designed; multiplying geometrically, however, they might soon run out of target tasks and become ‘hungry’ for more. An engine-management system in frenzied competition with itself and with other such systems may not make for a pleasant driving experience. Rat infestation in an air-traffic control system could spell disaster.

Fortunately, others have different ideas about AI rodent scenarios. The roboticist and AI-developer Steve Grand opts for the rather cuter analogy of squirrels. In his book Creation: Life and How to Make It, he suggests that even squirrel-level intelligence could be extremely useful to us:
But imagine putting squirrel brains into, let us say, a set of traffic lights. …If the mind of a rodent was placed into each signal, and the signals were rewarded for how well they managed to smooth the flow of traffic in their local area, then it seems plausible that it could work.
Note that reward is vital to this kind of scenario. These ‘squirrels’ would have to be intelligent enough to be able to seek reward, and they would have to be able to learn that they would find that reward in the harmonious and efficient performance of their appointed tasks. It may be hard for us to imagine smooth traffic flow as the ‘nuts’ of such a setup, but just try to think about how a sapient or transapient strong AI might view our motivations.

I find implausible the idea of a Singularity where we suddenly leap from where we are now directly to strong AI. More likely, and I agree with Hughes on this, is one where intermediate-level AIs begin to spring up and multiply as scientists press on towards ever-stronger artificial intelligence. Refining our AI ethics to cover the kinds of capabilities with which we might accidentally or purposely endow weak AIs may allow us to benefit from harmonious (and quite cuddly) traffic systems while avoiding the need for drastic pest control.