My lying AIs: how to tackle deceptive artificial intelligence

单集文稿 ...

• 00:04

  The Economist. Hello, this is Alok Jha,

• 00:12

  host of Babbage, our weekly podcast on science and technology.

• 00:16

  Welcome to Editor's Picks.

• 00:18

  We've chosen an unmissable article from the latest edition of The Economist.

• 00:22

  Please do have a listen.

• 00:32

  Artificial intelligence models are getting better and better.

• 00:36

  Cutting-edge systems can handle increasingly complex tasks,

• 00:40

  once thought beyond the ken of machines.

• 00:43

  However, as we report in the science and technology section this week,

• 00:47

  they can also find surprising ways to get things done.

• 00:51

  Given AI system, the task of beating a chess playing program, for instance,

• 00:56

  and rather than trying to checkmate its opponent,

• 00:59

  it may simply hack the program to ensure victory,

• 01:03

  give it the job of maximizing profits for an investment client with ethical qualms,

• 01:08

  and instead of changing its strategy,

• 01:11

  it may misrepresent the harms associated with the profits.

• 01:16

  Obviously, these models have no consciousness of their own.

• 01:19

  They are not acting with deliberate malice.

• 01:22

  Instead,

• 01:23

  they are responding to a tension between their initial training and configuration,

• 全部文稿