The Singularity Q&A

Q: If AI has not been created yet, why does it seem like all new software claims to possess it?

A: If you went to the grocery store to buy laundry detergent and saw a box labeled, "Hyper Ultra-Brite:  Now with Cleantechâ„¢ Scrubbing Spheres!" would you be more inclined to buy it than your usual detergent?  You might, even though you don't really know what "scrubbing spheres" refers to.  The name sounds exciting, like it will somehow do more for your clothes than the proverbial "leading brand."

If your box of software has a list of features on the back that includes "Advanced AI," you might similarly be more inclined to purchase it.  When you see the term "AI", you probably think of super smart robots in movies who can respond to your commands with something at least approaching the insightfulness of a human secretary.  Well, we are accustomed to hype in marketing, so when talking into your mouse doesn't get your latest version of WidgetWorks 9.0 to do anything, you probably forget all about the "AI" and go on muddling your way through the menus.  There might be some sort of AI in there, just as your laundry detergent might have some sort of "scrubbing spheres," but whatever this feature is doing, it's not obvious.

The problem lies in the way the term "Artificial Intelligence" has been constantly used by market-savvy software publishers to refer to many different things.  Some of these uses have become so commonplace that anyone attempting to refer to the really smart kind of AI -- the kind that hasn't been created yet -- often finds it necessary to be more specific.  I am no exception, and will talk about Artificial General Intelligence (AGI), real AI, genuine AI, strong AI, etc.  But rather than simply lamenting the hijacking of the AI name, we might as well discuss briefly the various types of so-called weak or narrow AI, the things they are admittedly very useful for, and the reasons they fall short of the original AI goal.

Most generally, AI is often used to refer to any kind of software that can learn something about a user and adapt to it.  For instance, speech and handwriting recognition programs examine samples of your words in order to learn your own particular style, and will generally increase with accuracy the more you use them.  Additionally, many common programs today include "wizards" to help you set up or troubleshoot common types of documents, files, or networks, by asking questions about what you want to accomplish.

The most advanced kinds of wizards are another type of narrow AI called expert systems:  specialized knowledgebases used to make decisions from complicated data.  Some of the most important expert systems are used in medicine to diagnose certain ailments.  Patients are typically asked a lengthy series of questions about their symptoms, and doctors asked to give certain other statistics.  The expert system crunches this data and gives a diagnosis based both on the rules it was initially programmed with and on past experience (the system is eventually told whether or not each diagnosis proved correct).  Expert systems that have built up vast libraries of experience have been known to significantly outperform human specialists, developing rules of such complexity that no human could hope to make use of them.  Nevertheless, the expertise of these systems is very narrow; they are useless for any subjects outside their specialty. 

One variation on expert systems is another type of narrow AI known as the immobot.  The name is short for "immobile robot," since these machine-control systems typically have no moving parts of their own.   Many people find the thought of a motionless robot to be something of an oxymoron (why not just call it an on-board computer?), but these systems have begun to reach a level of complexity deserving a new name.  Found inside everything from high-end copy machines to automobile engines, they are akin to on-board engineers, constantly tweaking performance and actively responding to trouble.  Some of these systems can learn from past experience, and all of them will generally outperform any human trying to do the same task;  they can make decisions hundreds of times a second based on dozens of different factors. 

But, sadly, you can't yank an immobot out of your car and expect it to play a game of dominoes with you.  For games, we need still other types of software that may or may not include features found in most expert systems.  The reference to a computer controlled opponent in a video game may actually be the oldest diluted definition of "AI" around, and game programmers are generally regarded as some of the most creative narrow AI builders out there.  The mission of a computer opponent can vary tremendously from game to game, and in simple or older games the AI will not generally adapt at all to the tactics of human players.  But the computer has a number of advantages over humans:  it can react instantly, control many different units simultaneously, calculate the odds or utility of certain conditions exactly, and, if allowed, cheat.  So, you might find that an AI opponent on the highest difficulty setting will clean your clock in a game like Warcraft, but you may also find that you can use some simple trick to defeat it over and over again because it can't learn from its mistakes.  Over the years, AI opponents have generally grown more cunning and adaptable, even if little loopholes like this are still very common. 

In fact, in many games, the best humans are no longer a match for the best AI, the victory of IBM computer Deep Blue over world chess champion Gary Kasparov being the most publicized example.  Deep Blue played a very different kind of game than Kasparov, however, and ultimately could be said to have triumphed through brute force;  the universe might have ended before Kasparov could have thought through all the moves his opponent did.  And yet, as a result of their differences, all the power at the disposal of Deep Blue would be useless in a game of Pokémon without a whole new design, while Kasparov might be able to learn the game in less than an hour.

Narrow AI is, as the name implies, narrow.  It is a master of one trade and jack of none.  You will know general AI when you see it, because you will be able to take it off one task and teach it to do something completely new without bending over backwards or having an aneurysm.  Be warned:  lately, companies have been creating or announcing software that they refer to as general-use AI.  What these publishers really mean is that they have created a type of AI that can function in any environment meeting certain, specific standards.  Although these programs will no doubt be very useful, the onus of adaptation will be on the user rather than the AI, since the environment must be kept AI-ready.  In a way, it is as though IBM were marketing Deep Blue as an AI that can play any game -- so long as it is presented as a game of chess.

Don't be fooled by imitations.  If an AI can't convince you of its general, flexible intelligence, it's not the real thing.

[Back to Futurism] [Back to Main]  ©2002 by Mitchell Howe