Two (Possible? Only Possible?) Failure Modes for AI

I’m an old AI guy — back from the ’70’s and ’80’s — and am often blown away by what “deep learning” and statistical approaches are accomplishing nowadays.  I never would have predicted an AI like Watson that could win at Jeopardy.  Or Go.

But some things are still the same, and when AIs fail, they fail in the same couple of ways.

I’m not talking about them turning the universe into a paperclip factory and eliminating us because we get in the way, a la “SuperIntelligence”  Maybe they’ll kill us someday, but that still seems a long ways off.

Today’s AIs disappoint us in one of two ways.

  1. Explanation.  Sadly, almost no humans will trust an AI’s conclusions without some account of how it reached those conclusions.  And most AI’s can’t account for their conclusions, especially the modern AI’s that are based on statistical weights and neural-style nets.  “I reached the conclusion that your cancer will respond to Treatment Cocktail A because I increased the weights on Nodes 1120-3388-692A-QRST and VVTX-8338-QQ94-AAAA from 10 to 30.”  Yeah, right.
  2. Turing Gulf.  This phenomenon, also called the “uncanny valley”, has a nice explanation here.  It’s often used to talk about an AI that’s kind of creepy because it’s near-human, but it can also be used for an AI that’s almost good enough but peeves you when it fails to make the grade.  Imagine an AI that needs to be “90”, where 90 is “90% accurate” or whatever.  And the AI is only “89” (again, whatever that means).  That AI is useless for the task because it will only peeve and frustrate its users.

Are those the only possible ways AIs can fail?  Welcome your comments.

Leave a Reply

Your email address will not be published. Required fields are marked *