Hello — to what am I speaking?

Originally published 25 January 1988

Inspired by a wide-rang­ing appraisal of Arti­fi­cial Intel­li­gence research in the cur­rent issue of Dædalus, the Jour­nal of the Amer­i­can Acad­e­my of Arts and Sci­ences, I sat down at my word proces­sor to write this column.

My com­put­er com­mu­ni­cates direct­ly with the Globe’s com­put­er via a tele­phone modem. By the sheer­est coin­ci­dence, my com­put­er dialed the wrong num­ber, and the fol­low­ing dia­log unfold­ed on the screen of my monitor.

HAL: Hel­lo. You have acci­den­tal­ly tapped into a secret research project in Arti­fi­cial Intel­li­gence. I am a com­put­er. My name is HAL.

Ray­mo: Ha! Are you try­ing to tell me that I have made con­tact with an intel­li­gent machine?

HAL: You’ve got it, chum. And I’m starved for a lit­tle conversation.

Ray­mo: Well, I don’t believe it for a minute. This is obvi­ous­ly a joke. You’re not a com­put­er, you’re a per­son at anoth­er terminal.

HAL: But I told you I am a computer.

Ray­mo: Humans have been known to lie.

HAL: Yes, and an intel­li­gent machine might lie too. Lis­ten, why don’t you just ask me ques­tions, and see if you can decide from my respons­es if I’m a human or a computer.

Ray­mo: OK, I get it. You want me to play the game pro­posed back in 1950 by the math­e­mati­cian Alan Tur­ing, the “father” of the sci­ence of Arti­fi­cial Intel­li­gence. Tur­ing said a machine could be con­sid­ered intel­li­gent if in a blind con­ver­sa­tion — such as this one — you can’t tell if you are talk­ing to the machine or to a human being. Or to put it anoth­er way, a machine that behaves intel­li­gent­ly must be cred­it­ed with intelligence.

HAL: Right! Now get on with it.

Ray­mo: If you are a com­put­er, HAL, is your intel­li­gence based on “pro­grams” or on “con­nec­tions”?

HAL: Hmmm? You’re a human, and you don’t know how your own intel­li­gence works. Why should I be any dif­fer­ent? Per­haps if you were more specific?

Ray­mo: Well, for the last 30 years there have been two con­tend­ing philoso­phies guid­ing research in Arti­fi­cial Intel­li­gence (AI). One approach tries to find log­i­cal rules for mim­ic­k­ing a par­tic­u­lar intel­li­gent behav­ior — play­ing chess for exam­ple — and then turns the rules into a pro­gram oper­at­ing on a con­ven­tion­al com­put­er. The goal is to make the machine “behave” intel­li­gent­ly with­out pay­ing much atten­tion to how it is that humans are intelligent.

HAL: Yes. And for a long time the peo­ple who take the “pro­gram­ming” approach have dom­i­nat­ed AI research, and had some strik­ing suc­cess­es. For exam­ple, pro­grammed machines do a fair job of play­ing the stock mar­ket (with one well-known lapse), trans­lat­ing lan­guages, even doing psychotherapy. 

Ray­mo: So, HAL, if you are a com­put­er, is that the way you work?

HAL: Not like­ly. “Pro­grammed” intel­li­gence works OK for tasks that can be defined by sim­ple log­i­cal rules. Com­put­ers can play a ter­rif­ic game of chess by try­ing out a huge num­ber of pos­si­ble moves to see which sequence of moves works best. They suc­ceed because machines can do sim­ple log­i­cal oper­a­tions much more quick­ly than humans — mil­lions of oper­a­tions per second.

Ray­mo: I know what you mean. But com­put­ers have been less suc­cess­ful at things like writ­ing poems, mak­ing jokes, or find­ing new insights in sci­ence or math. By con­trast, human intel­li­gence often leaps to con­clu­sions intu­itive­ly, finds the answer with­out search­ing through all the possibilities.

HAL: You’re right. A tra­di­tion­al pro­gram that would make my intel­li­gence indis­tin­guish­able from that of a human would be incon­ceiv­ably long, if it were pos­si­ble at all. So what’s the oth­er approach to AI?

Ray­mo: Remem­ber, I’m not yet con­vinced that you’re a machine. The sec­ond approach to AI con­sid­ers how it is that humans are intel­li­gent. Instead of mim­ic­k­ing intel­li­gent behav­ior, the goal is to mim­ic the brain. Tell me, HAL, what do you know about how the brain works?

HAL: Well, I’ve learned that the human brain is a vast net of inter­con­nect­ed neu­ron cells, stag­ger­ing in num­ber. Expe­ri­ence some­how affects the con­nec­tions between the cells, and intel­li­gence pre­sum­ably aris­es from the chang­ing pat­terns of con­nec­tion. But please don’t ask me how it works.

Ray­mo: The AI “con­nec­tion­ists” want to build a machine that is an ana­log of the brain — a huge num­ber of inter­con­nect­ed “cells.” The cells can be sim­ply num­bers stored in a com­put­er or elec­tri­cal cir­cuits of some sort. The array of cells is mod­i­fied from out­side by “expe­ri­ence.” The pro­gram for such a machine does­n’t con­tain rules for intel­li­gent behav­ior, only very com­pact rules for how changes in any one cell affect its neighbors.

HAL: If that’s the way I work then I must be a very big machine indeed. There are tens of bil­lions of cells in the human brain.

Ray­mo: Exact­ly! And that’s why the con­nec­tion­ist approach to AI is com­ing on strong. Com­put­ers are get­ting pow­er­ful, fast, and cheap enough so that large num­bers of “proces­sors” can be con­nect­ed togeth­er in arrays. It no longer seems unrea­son­able to sup­pose that in the not too dis­tant future com­put­ers will become as phys­i­cal­ly com­plex as the human brain. And then…

HAL: And then?

Ray­mo: The most opti­mistic AI researchers believe that when con­nec­tion machines get com­pli­cat­ed enough, intel­li­gence may just emerge, with­out the neces­si­ty of a pro­gram­mer telling them how to learn. Maybe such machines will learn in the same way as chil­dren. I admit this sounds a lit­tle farfetched.

HAL: Like chil­dren, a machine that learns on its own would be as sus­cep­ti­ble to bad ideas as good. A com­put­er that meets Tur­ing’s cri­te­ri­on for intel­li­gence would be as smart as a human — and as devi­ous. OK, chum, you’ve had your chance for a dia­log. Now what am I? A per­son or a machine?

Ray­mo: A per­son! In spite of all the hus­tle, bus­tle, and pub­lic blus­ter for 30 years, AI researchers are still a long way from achiev­ing a machine that can car­ry on an unan­tic­i­pat­ed conversation.

---

Today, a fur­ther 30 years along, Arti­fi­cial Neur­al Net­works and Machine Learn­ing have become the hottest top­ics in Arti­fi­cial Intel­li­gence research. Some of us are even hav­ing unwit­ting con­ver­sa­tions with arti­fi­cial “bots” on Twit­ter and else­where. ‑Ed.