Wednesday, December 30, 2020
Try the Gregg test instead of the Turing test
Random thought about AI bots after reading this funny account of a failed Turing test.
You could get a sharper test by requiring both the "applicant" and the tester to be a type of person who isn't well represented on the Web, and who isn't likely to be known by the programmers of the bot.
Ideally the type should belong to a densely defined subculture that isn't easily accessible or understood by academics, eg Amish or Hmong.
Good human actors can realistically simulate people who aren't well described in media, types they've never encountered in life. I'm thinking of two radio actors in the '30s and '40s: Virginia Gregg and John Brown. Each was able to "be" all sorts of EXTREMELY DIFFERENT characters with complete accuracy. Some of the characters were close to my experience, so I could judge their accuracy, but I could NEVER recognize the actor as Gregg or Brown until I read a description of the program.
Most actors fail this test. When Lucille Ball or Jack Webb played different characters with different dialects, you could tell immediately that it was Ball or Webb.
Labels: AI point-missing
¶ 4:56 AM
