Will Researchers Develop a Computer that Duplicates Human Intelligence in the Foreseeable Future? No.

Stuart M. Shieber
Division of Engineering and Applied Sciences
Harvard University

In 1950, the British mathematician Alan Turing devised what has come to be known as the Turing Test for machine intelligence, based on a judge's inability to distinguish computer and person in open dialog with the two. The Turing Test is popularly thought of as the holy grail of research in Artificial Intelligence. A computer able to pass the Turing Test is in essence what many think of when they think of "duplicating human intelligence in a computer". HAL 2000 springs to mind.

I have good news and bad news. The bad news is that like the Holy Grail of Arthurian legend, the Turing Test will I expect remain beyond our grasp for quite a while, undoubtedly not achievable during our lifetimes. Philosophers like Searle and Dreyfus think that it is in principle impossible; although I am agnostic on this point. Various AI luminaries and media pundits have been predicting the imminent arrival -- from "just around the corner" to "just a few decades" -- for many years now, at least since the 1950's. Like prognostications of the end of the world, as the predicted dates pass, the predictions become necessarily more inaccurate. The problem of how intelligence works and how it can be duplicated artificially is tremendously difficult. Denying this fact is simple hubris. The issue is not one of insufficient computer power; even if we had computers faster by orders of magnitude (which we undoubtedly will), we would need to know how to make use of the resources to duplicate human intelligence or allow computers to learn it on their own. We don't.

The good news is that like the Holy Grail of Arthurian legend, the goal itself is less important than the quest. Indeed, many if not most AI researchers view the Turing Test as an exceedingly poor goal for current research in the field. The study of AI, by engaging some of the brightest minds in computer science on arguably the hardest problems in the field, can claim credit for time-sharing computers, windowed interfaces, computer dictation, medical diagnostic systems, financial industry mechanization, Deep Blue. The technologies on which these systems were based were not developed by researchers trying directly to build artificially intelligent Turing-test passers, but through myriad attacks on varied problems in understanding particular types of knowledge, reasoning, learning, and intelligent behavior. It is important not to gauge progress in AI on progress in passing the Turing Test. One can, and we do, have tremendous progress in the former, both in theory and in practice, without approaching the grail itself.

[The CQ Researcher, volume 7, number 42, page 1001, November 14, 1997.]