Re: Thinking about the future...

From: N.Bostrom@lse.ac.uk
Date: Tue Sep 03 1996 - 12:47:14 MDT


          About assuring that an >AI won't harm humans, Robin Hanson
          said that the police would hunt it down if it misbehaved.
          
          OK, let's assume that an >AI is positively maliscious. It is
          easy to see what would happen to a moderate >AI in a
          laboratory if it misbehaved. It would be switched off. It
          could be more problematic if it were out there in the real
          world, perhaps with the responsibility for a banking system.
          Economic losses could be great, but we would hardly risk
          total destruction unless we gave it unrestricted power over
          the US nuclear arsenal or such.
          
          But when transhumanists talk about >AI they hardly mean a
          moderate >AI -like a very brilliant human and then some. We
          speculate about a machine that would be a million times
          faster than any human brain, and with correspondingly great
          memory capacity. Could such a machine, given some time, not
          manipulate a human society by subtle suggestions that seem
          very reasonable but unnoticeable affects a general change in
          attitude and policy? And all the time it would look as if it
          were a perfectly decent machine, always concerned about our
          wellfare...
          
          How likely is it that a malicious >AI could bring disaster
          to a human society that were initially determined to take
          the necessary precautions?
          
          Society is a quasi-chaotical system: a small perturbation of
          initial conditions will often lead to large unexpected
          consequences later on. But there are also regularities and
          causal relations that can be to some extent predicted. The
          manipulative power of an >AI would depend on the existence
          of such sociological regularities that would be obvious to
          the >AI but would look like irrelevant coincidence to human
          observers. It is a question of how much sociology and
          psychology there is to be discovered between what we know
          now and what an >AI could learn in the course of few years.
          
          Exactly how much manipulation it would take depends on how
          close we are to the nearest road to disaster. Perhaps we are
          close. A deadly virus that would happen to be produced as a
          by-product of some medical experiment; a little incident
          that would lead to escalating hatred between China and USA;
          the list goes on...
          
          My contention is that with only one full-blown >AI in the
          world, if it were malicious, the odds would be on the side
          that it could annihilate humanity within decades.
          
          Nicholas Bostrom n.bostrom@lse.ac.uk



This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 14:35:44 MST