In this second selection of speculative fiction, and excerpt from a forthcoming novel, David Brin asks how we will keep our machine mind progeny loyal.
...
Complete entry
Posted by
Greg Hullender on 01/28 at 02:52 PM
I've worked on machine learning and natural language since about 1979. I recently went back to school to get a masters in computational linguistics. I made a study of Dialogue Systems, which are software systems meant to be conversational. A "General Dialogue System" would, by definition, pass the Turing test. "Practical Dialogue Systems" exist today (I actually wrote one for a class project) and let you converse as long as you stick to tasks the computer can do. (E.g. make a plane reservation.) If you haven't, you ought to read about dialogue systems.
That said, I spent a lot of time thinking about why we can't make a general dialogue system. I concluded that the problem is with volition. WHY does the system do or say anything? I could create a system that correctly parsed and "understood" everything a person said, but then what? What does it do with that info and why? The only answer I could find is that it will only "want" to do what people have programmed it to.
So it would never be a crime to create an AI--just to give one motivations that are contrary to human interests. Much as it's no crime to use dynamite, as long as you don't try to hurt people with it.
I could visualize highly paid teams of "Volition Designers," who construct the rules that govern deep behavior of an AI. (Asimov's laws of Robotics are a very, very simple form of this; the Robot's only volition is to a) save people b) obey orders c) protect itself. Otherwise, it'll just sit there.)
So could you have a self-modifying AI? Meaning one which could change its own volition template? I almost want to claim that that's logically impossible; whatever code decides when and why to change the template is now the real volition system.
So could a system kill people? Sure. It could have bugs (like HAL 9000--everyone's favorite General Dialogue System), or a malicious programmer could build one, but I can't see one evolving. Would they get rights? I hope not, but people can be dumb. Someone could easily make an AI that mimed human feelings, swaying the public into pushing for such rights. (I already see people get attached to characters in video games, and those have almost no ability to communicate.)
One thing's for sure; we have made zero progress toward a machine that "wants" to do anything. Something about human intelligence is very, very different from anything we have ever put on any machine ever made. It may not seem that way to someone outside the field, but everyone doing serious work in the area knows it.
IEET Blog