I do some of my best thinking in the shower. In last night’s shower, I was thinking about artificial intelligence. A.I. systems have a come a long way since folks like Alan Turing, Marvin Minsky, and others first begin thinking about them. Many systems today seem to provide at least the illusion of passing the Turing test. I think of Siri and Alexa as just two examples. I’m sure some readers will debate whether or not these pass the Turing test. Let me state that for my purposes, they give me the illusion of doing so–so much that I am often compelled to thank them for their tasks. When my parents taught me to say thank you, they said you always say thank you. They didn’t make exceptions for artificial intelligence.
What I wondered while rinsing the shampoo from my hair last night was this: is there some obvious task I can give to a purported AI that would convince me it is, in fact, intelligent? People often point to chess (a game for which I have no skill or patience) and the IBM AI that can beat just about any human. That is an amazing feat of programming, but I’m not sure it convinces me. Is there something else that would convince me?
I allowed my mind to drift. Showers these days are pretty much the only time my mind really drifts, completely disconnected from the world. I thought about chess, about games, and puzzles. I thought about crossword puzzles, which led me to think about some of the great clues in crosswords that are occasionally recycled. My favorite is the 3 letter word for “Little giant,” which of course is “Ott” for Mel Ott, the 5′-9″right fielder for the New York Giants from 1925-1947.
Given a blank crossword, and the first clue “Little giant”, could an A.I. complete the crossword? Could it figure out through techniques of machine learning and neural networks that “Little giant” = “Ott”?
Perhaps it could, but it seems it would be quite a challenge because it involves a clue containing not just a metaphor but a pun. Moreover, the answer is a proper noun and the correct answer is not the first name (also three letters) but the last name. If more of a challenge were need it, could an AI complete a crossword puzzle “in ink”–that is, without having to correct itself once it put down an answer? Thus, if it answered “Mel” instead of “Ott” it would be stuck with it.
I’m better at crosswords than I am at chess, but I’m still in the lower ranks of crossword solvers I know. I’ve picked up on patterns along the way, however. When the clue is plural the answer is plural, for instance. Would an AI pick up on subtleties like that? Many crossword clues or answers are based on some kind of pun, homonym, or other quirk of English. How do you teach the concept of puns to an AI?
I emerged from my shower with the idea for a new test for AI’s. Instead of the Turing Test, I’d call it the Shortz Test.
After writing this post, I did something I don’t often do: I did a Google search to see if anyone had written on this topic before. I typed “Can AI solve a crossword puzzle” into Google. There were plenty of matches, so I guess this thought isn’t original with me. And clearly, I am not the only one who thinks that crossword puzzles may prove a particular challenge for AIs.
I wonder if they got their idea in the shower, too?