Clippy’s revenge 08:21 on Friday

“It looks like you’re making a geeky reference.” Yeah, I was referring to the infamous mr.Clippy in my recent post. He sure was annoying and we’re happy to be done with him. Yet we’d like our powerful computers to be more helpful, conversational even.


Like Clippy demonstrated, for these powerful but hopelessly dumb machines a conversation involved quite a lot of guessing, at best. So as a programmer of such artificial intelligence, how do you know how much guesswork from the computer’s part is acceptable by the users? What factors are in play? Some ideas:

  • Guess success ratio. Independent of the value of a correct guess, I believe the more often a guess is right, the more acceptable the intelligent agent becomes.
  • Interruption cost. How much time does the interruption waste? How much time does it take to get back into flow state after an interruption?
  • How much time a correct guess can save? If a correct guess saves a lot of time and thus is valuable to the user, a lot more misses are tolerated.
  • How much time a wrong guess wastes? Likewise, if a failed guess actually wastes the users’ time, the help will not be appreciated. For example, if the user needs to correct wrong values guessed and “helpfully” and automatically entered by computer. Browser autofill is one example of such behaviour.
  • Criticality of task. How much more or less tolerant of agents will users be when they are dealing with critical tasks or information?

Comments are closed.