By Harry McCracken | Monday, April 27, 2009 at 10:22 am
When IBM’s Deep Blue computer beat chess world champion Garry Kasparov back in 1997, my reaction was pretty dispassionate: As someone who’s played only a few games of chess in my life, I simply didn’t have a deep understanding of the game or an emotional attachment to it. But when the news came out over the weekend that IBM is programming a supercomputer to play Jeopardy–well, that’s a breakthrough I can relate to. I haven’t watched the venerable game show much in years, but back in the 1980s, I planned my college courses around it to make sure I was home in time to watch (this was before I had a VCR). I get the game’s subtleties–it’s not just about having an encyclopedic knowledge of both serious stuff and pop culture, but also about being able to unpack the meaning of those questions-phrased-as-answers in a split second. And given that countless very smart people have gone on the show and fallen flat on their faces, I’ll be impressed if IBM’s computer manages an unembarrassing third-place finish. But I don’t have any difficulty dealing with the idea that a computer might someday beat any flesh-and-blood Jeopardy player on the planet.
Thinking about the prospect of a computer taking on Alex Trebek, metaphorical buzzer in hand, led me to the conclusion that game shows in general aren’t a bad Turing Test-like gauge of artificial intelligence. They require knowledge–okay, only a little of it in many cases, but some. They’ve got a social component, by definition. They involve thinking on one’s feet, or the simulation thereof.
So how might a really well-programmed supercomputer fare at other famous game shows of the present and (mostly) past?
Wheel of Fortune. On one hand, the game that keeps Pat Sajak and Vanna White off the streets is a mind-numbingly simple pattern-matching exercise–a computer that had indexed a few million phrases, names, and titles might be eerily good at guessing correctly with only a few letters in place, and I suspect that there’s a definitive vowel-buying strategy that could be rendered in code. On the other hand, the “Fortune” in the game’s name indicates that there’s a large component of luck in the spinning wheel.
What’s My Line? Jeopardy requires knowledge of facts of all sorts; this game is narrowly focused on the single task of guessing a person’s profession. That might make it a more manageable challenge. But the whole point of the show was that the jobs in question were odd and unexpected, and sometimes unique. I’d still be impressed if a computer was more adept at correctly guessing that someone managed, say, a girdle factory than the eerily skillful Bennett Cerf was. (This is all assuming that the computer would take the place of one of the celebrities on the panel, rather than one of the contestants attempting to win money–computers, by and large, don’t have interesting and amusing professions.)
To Tell the Truth. I always think of this game as an inferior clone of What’s My Line. But I think it would be far harder for a computer to do well at, since it involves asking three people questions, then determining which one is truthful based not only on the accuracy of his or her questions, but also on psychological factors such as whether the person looks shifty and duplicitous. Of course, maybe a computer could cheat by relying on a humongous image database of photos of most everyone in the country, plus some really good face-recognition software.
The Dating Game. Programming a computer to play as Bachelor or Bachelorette number one, two, or three would be, essentially, programming it for a well-defined version of the Turing Test. If I were IBM, I’d digitize thousands of old episodes of the show, perform speech-to-text conversion on all the questions ever posed (“Bachelor number two, if you were an amusement park ride…”) and use it them to gain some sort of understanding of the parameters of the Dating Game universe.
The Newlywed Game. Like The Dating Game, this one involves goofy Chuck Barris questions. Unlike the Dating Game, it requires intimate knowledge of another human being, plus an understanding of that person’s psyche. Bottom line: A computer may never be able to play this game competently until computers are able to marry people. And sleep in the same beds as them. And use some sort of infrared system to determine if they hog the covers.
Let’s Make a Deal. Like Blackjack, the show that made Monty Hall famous involves a gigantic dose of luck, but it’s possible to improve your chances via an understanding of the odds. In fact, the “Monty Hall Problem” has provided fodder for discussion by professional and amateur logicians for years. I suspect a computer could play this game better than a human, especially since it could blithely ignore the mind games that Monty Hall played on contestants. Also, the show sometimes used pricing games similar to The Price is Right, and a computer with a database of prices for consumer products could presumably win such contests every single time. On the other hand, you only got to play Let’s Make a Deal if Monty Hall picked you out of the audience, and I’m not sure if a supercomputer could fit in a theater seat at all, let alone don a funny costume.
The Price is Right. See Let’s Make a Deal. Give a computer a comprehensive price database, and it might force Goodson-Todman Productions into bankruptcy, even though games such as Plinko have a high component of luck to them.
Family Feud. Like What’s My Line, this game has well-defined parameters that would lend themselves well to play by a computer. Presumably it would be possible to program the notion that “apple pie” is a more likely response to “name a dessert you’d take on a picnic” than “baked alaska” is. Mostly, I’m just trying to deal with the concept of Richard Dawson smooching a bunch of computers.
The Match Game. Oddly enough, this game may be the easiest one of all the ones here to win for a human–even though a high percentage of the humans who played it in the 1970s performed dismally–but the hardest for a computer. It’s not just that the computer would have to be able to parse sentences such as “Ugly Edna said “I hate Ralph, he asked me for my picture and when I gave it to him he used it as a [BLANK].” It would also have to get puns. And have an appreciation for absurd visual imagery. And have a sense of how six specific celebrities would be likely to respond. (I’ve watched enough reruns on GSN to know that my strategy would have been different if Fannie Flagg had been on the panel than if Joyce Bulifant had been.) Gene Rayburn, Bret Somers, and Charles Nelson Reilly are no longer with us, sadly, so we’ll never know for sure how a 2009 IBM supercomputer might have fared back in 1973 or so. But my guess is that it would have done even worse than most of the flesh-and-blood players. It’s a rare example of a narrowly-defined challenge that’s in some ways trickier for a computer than a broader Turing Test would be.
Okay, that’s enough of that. Do you think IBM’s supercomputer stands a chance of doing to Ken Jennings what Deep Blue did to Kasparov?