Rationality Toughness Tests

by RobinHanson 1 min read6th Apr 200917 comments


(Epistemic) rationality has two major components:

  • Smarts: An ability to, by attending, infer truth from info under ideal circumstances.
  • Toughness: An ability to limit performance degradation as circumstances worsen.

Attending takes time, energy, quiet, etc.  Circumstances where human rationality degrades include when:

  • We expect the truth to long remain hidden.
  • The stakes are very low, or very high, to us.
  • Others see our opinions, and prefer certain ones.
  • The topics are where humans oft self-deceive.

It seems relatively easy to test rationality smarts; repeatedly give folks info and time to work new problems and measure their accuracy, calibration, etc.  And I have an idea for testing for rationality toughness: compare performance on info-similar pairs of good/bad-circumstance problems. 

For example, assume people are better at evaluating if a spouse is cheating when considering an acquaintance in their social circle, relative to a stranger or their own spouse. If so, we could pose them a pair of problems with very similar info structure, one with an easy spouse and one with a hard spouse.  The closeness of their response in these two cases would then be a measure of their rationality toughness.

Of course this test may fail if the similarity is too obvious, or the pair are asked too closely in time.  But maybe we don't even need to ask the same person the two questions; perhaps we could usefully compare someone's answer on a hard question to answers from a pool of similar people on matched easy questions.

While I haven't thought this through, it already suggests a training technique: consider matched hard/easy circumstance problems and compare your answers, separated by enough time that you forget most of your previous analysis.