The Lighter Side of Artificial Intelligence

So, remember the 20 Questions game? You think of a person or thing, and the other guy tries to guess the answer using 20 yes-or-no questions. It’s a classic time killer, good for long car rides, roller coaster lines, and waiting out the goddamn Bush administration.

I’m pleased and moderately frightened to report that the game has made a successful transition to the Internet Age. The 20Q initiative is an ongoing project in Artificial Intelligence (AI) that uses an insanely complex Web-based neural network to simulate a game of 20 Questions. You think of something — person, place, thing, abstract idea; doesn’t matter — and start answering questions posed by the 20Q AI. The questions posed can seem impossibly random, some examples might be: “Can you buy it?” “Does it have four corners?” “Does it like to be petted?”

You can answer Yes, No, Unknown, Irrelevant, Sometimes, Probably or Doubtful. After the program has posed an adequate number of questions — often less than 20 — it will attempt to guess the term you’re thinking of. Prepare to be freaked directly out at how smart this thing is.

The 20Q AI was invented in 1988 as a smallish database passed around on floppy disc amongst a group of friends. The game moved online in 1995, and has been growing and “learning” ever since. It makes its guesses based on what it has learned from previous games, and is not programmed with information in any traditional way. As such, it’s more of an autonomous information base that reflects the knowledge of the millions of people who have already played it. According to Wikipedia, the online 20Q AI has about 10,000,000 synaptic connections, with about 10,000 objects in its knowledgebase.

Since moving online, the 20Q technology has been patented and licensed to several third-party organizations. A handheld version for kids, by Radica Games, has won several industry awards, and Burger King licensed a version of the game to power a Darth Vader-narrated promotional tie-in with Star Wars: Revenge of the Sith. 20Q played its 40,000,000th game in July of 2006,

The kids’ version of the game was my first introduction to 20Q, and it came about rather randomly. On a very manly canoe trip with manly college buddies, somebody picked up the game at a local hobby shop as a gift for his daughter. We manly men then proceeded to geek out on the game for about three days straight, eschewing all previous plans of fishing, drinking, poker, etc.

What we discovered was that the game is, as advertised, about 80-90 percent accurate. And this is just the kids’ game, which uses a smaller database and does not “learn” as the online game does. We threw some weird concepts at it, too. Among the terms it accurately guessed: “carburetor,” “e-mail,” “infinity” and, um, “shit” (or “poop,” according to the kid-friendly AI).

Since going online, the 20Q project has added a few satellite AIs, databases specific to cultural categories such as rock and pop music, TV and movies, and sports. Spend a little time here, and you’ll find the game can open up some very interesting pop-cultural and even philosophical issues. Some of the questions posed by the AI can really make you think.

For instance, I thought of the word “president” and came across a few stumpers. Sometimes things can get ambiguous, depending on whether you’re referring to the individual or to the office of the presidency, but I was surprised at how oddly thought-provoking the queries could be. (These are all real questions returned by the 20Q AI):

“Can it be bought?”
“Is it clever?”
“Does it perform?”
“Can you control it?” (I answered “no,” although occasionally Congress gives it the old college try.)
“Can you find it in a church?”
“Does it provide protection?”
“Is it brown?” (Hmm… 200 years of old white guys so far, I’m not holding my breath.)

The 20Q engine eventually got it right, after a few run-throughs. It made some interesting wrong guesses, too, including “clown”, “criminal”, and “ninja.”

I’m here to tell you that the 20Q website is good, creepy fun for all ages. It really is astounding how often the AI gets it right. Starting from “Animal, Vegetable, Mineral or Other,” it successfully guessed such varied items as: “time” (17 questions), “espresso” (17 questions), “mobile phone” (20 questions), and “algorithm” (20 questions). The Pop & Rock AI guessed “Lou Reed” in 18 questions (although for some reason it guessed Paul Stanley first.). The Movies & TV database got “Battlestar Galactica” in 20 questions, and the Sports AI got “The Stanley Cup” in 17.

I’ve long been comfortable with the fact that computers are smarter than I am, but there’s something vaguely disturbing about a computer being better at political and pop culture references. It makes me nervous. Just consider: By asking millions of people 20 questions about whatever crazy-ass thing we want to think up, the 20Q AI can now accurately guess, within 20 questions, any other crazy-ass thing we want to think up. This way madness lies. I’ve considered thinking about my social security number and seeing what it comes up with, but I’m too afraid to try.

You go first.