A New Manual of French Composition "provides a guide to French composition aimed at university students and the higher classes in schools. "
The LINGUIST List is dedicated to providing information on language and language analysis, and to providing the discipline of linguistics with the infrastructure necessary to function in the digital world. LINGUIST is a free resource, run by linguistics students and faculty, and supported primarily by your donations. Please support LINGUIST List during the 2016 Fund Drive.
Date: Wed, 8 Jun 2005 13:46:28 -0400 From: Larry Smith Subject: The Turing Test: Verbal Behavior as the Hallmark of Intelligence
EDITOR: Stuart Shieber TITLE: The Turing Test SUBTITLE: Verbal Behavior as the Hallmark of Intelligence SERIES: Bradford Books PUBLISHER: MIT Press YEAR: 2004
Larry H. Smith, U.S. National Center for Biotechnology Information, Bethesda, MD.
This book is an anthology of twenty famous papers on the Turing test related to the question of whether machines can be said to think. Each chapter of the book was previously published, except for the final chapter which is an essay by Noam Chomsky appearing in print for the first time. Gathering these papers into a single book aims to dispel some misconceptions about the often referenced Test and its famous criticisms.
Summary of Chapters
The first three chapters contain precursory writings from Descartes, demonstrating that animals cannot possess souls (which for Descartes means the same thing as "capable of thought"), and a rebuttal from the 18th century philosopher Julien Offray de la Mettrie, who vehemently argued the opposite in "Man a Machine."
The next four chapters are papers by Turing, beginning with the 1950 essay "Computing Machinery and Intelligence" in which he first proposed his test. There are many sensible re-interpretations of the Turing test, but it should be remembered that Turing originally framed it as a challenge put to a computer to imitate a human communicating by teletype. The remaining papers further clarify Turing's thinking on the origin and meaning of the test, and include the transcript of a 1952 BBC interview of Turing and some of his eminent peers.
The remaining chapters can be grouped into opponents and proponents of the Turing test. Pinsky offers a humorous parody of the test designed to uncover human neurotic thinking. Gunderson rejects Turing's proposal because it is based on logic that could also be used to imply that rocks are capable of imitating human toe-stepping behavior. Purtill claims that a machine that merely recalls stored responses cannot be called intelligent, to which Sampson gives a rebuttal. Millar argues that the concept of intelligence is necessarily and exclusively applicable to humans only. French then argues that because human behavior is entirely conditioned on human experience (or, what he calls subcognition), that it is impossible for a non-human ever to pass a Turing test.
Then comes Searle's famous "Chinese room" article, in which it is argued that an English speaking person following instructions for carrying on a conversation in Chinese cannot possibly be judged intelligent on the basis of those conversations. This is followed by a similar refutation by Block using his "Aunt Bertha" machine, also called a "Blockhead", which is able to imitate human behavior perfectly for a given time period by searching a database of all sensible responses (typical say, of his Aunt Bertha) stored for that time period.
Dennett rebuts Block by insisting that we must apply the label "intelligent" when it appears to fit, and that we are not qualified to judge a thing on the basis of its composition. Moor makes a similar argument, that an unrestricted test is strong evidence supporting the theory that a machine does indeed "think." Stalker then criticizes Moor for his willingness to apply a human theory of mind when an engineering theory of computers seems more appropriate. And Moor gets the last word by pointing out that actually both theories are valid. Finally, in a previously unpublished paper, Chomsky argues that it is idle to debate whether a machine can exemplify such a non-transferable human concept as "thinking", quoting Wittgenstein that "We only say of a human being and what is like one that it thinks."
Summary of Essay
Shieber's comments begin in the introduction and continue throughout the book as introductory material between chapters, filling approximately sixty pages. He prepares the reader for each stage of dispute and suggests terminology to help tie the papers together. The chapters are ordered by relevance so that each chapter precedes its extensions and rebuttals, even when they are not chronological. Shieber subtly encourages the reader to accept Turing's defenders, to reject his critics, and to appreciate his originality and insight.
The papers in this anthology (except for Shieber's essay and Chomsky's unpublished paper) are well known historical documents. Likewise, Shieber's essay nobly avoids taking a stand for or against any of the arguments, which are evaluated in subsequent papers anyway. I therefore confine my remarks to the general ideas expressed in the book.
Despite the subtitle, this book contains nothing that might be called linguistics. The authors seem to agree implicitly that the ability to use language is simply a convenience; any behavior that can exhibit intelligence could substitute in the test (though, to be consistent with Turing's insight, the behavior should be digitized). Also, although Shieber starts the book by quoting humorous pop references to the Turing test, there are no further references, and no analysis of the social impact of the Turing test.
I began this book with excitement, which quickly faded when I came to the responses following Turing. The opposing views made me feel like I was reading a "blog" or some opinionated newsgroup. The light of reason only returned with Dennett and the subsequent articles culminating with Chomsky. It is the unprofitable duty of professional academics to respond with carefully reasoned, rational, point by point rebuttals to misleading pronouncements of their peers, and we should thank de La Mettrie, Sampson, Dennett, Moor, and Chomsky for speaking up. Readers' experiences may differ, but anyone who makes the effort to consider each paper carefully and in the given sequence will develop or at least sharpen their opinion of machine intelligence and the Turing test in particular.
Praise for Turing
Turing's intuitions are penetrating. He argues that we shouldn't even try to assess whether machines think, but instead we should ask whether machines are capable of intelligent behavior. This was not meant to substitute for the original question; he meant to dump the original question completely. He also realized that in assessing intelligent behavior, people are likely to have human prejudices, which might be overcome if the behavior is digital from the start. And finally, it was obvious to Turing that computers would eventually be called "intelligent," and I suspect he believed this was a fact of human psychology and not philosophy.
There are two questions that are hopelessly confounded, (1) can machines can think? and (2) does passing a Turing test prove that a machine thinks? Shieber calls the second one the "big question." Proponents of Turing understand that it is ridiculous to argue about whether a machine can think, and consider observable behavior instead. While critics build their arguments on tacit assumptions that machines cannot think. Not once do the critics attempt to explicate the concept of "thinking" nor how it differs from "intelligent behavior."
A common strategy, from Descartes to SearlE, is to propose a counterexample, a "wedge" as Schieber calls it, of an "unintelligent" machine that appears to behave intelligently. The problem is that the arguments given for why the proposed machine isn't intelligent, an argument Shieber would call the "spark," when they are given, are not convincing. The opponents seem to rely on sympathetic emotional prejudices of the reader. To take one example, Gunderson quotes Turing, word for word, but with the computer replaced by a device that drops a box of rocks, and the goal to imitate a person stepping on someone's foot. Gunderson apparently hopes the reader will have the gut reaction that "rocks can't imitate people." But as later authors point out, the behavior describes the system, not the rocks or other isolated components. By providing terminology for it, Shieber calls attention to the defect, but his analysis is too forgiving.
In fact, there is a human capacity to recognize "thinking" and "intelligent behavior", but it is explained by cognitive psychology and not philosophy. It's unfortunate that none of the authors seemed to take this seriously.
An alternative thesis
There is something to be said for the relative neutrality of Shieber's essay, however the book might have been more stimulating if it had taken a stronger position. For instance, the philosophical debate of the Turing test was a historical mistake which had the unfortunate consequence of leading many a good mind down a garden path. That might have been avoided if Wittgenstein had been taken seriously, and if these philosophers had accepted that there is an essential human psychology in the concepts of "thinking" and "intelligence." Chomsky, a linguist, puts the endeavor on the right foot, but then he advocates dumping the question altogether. On the other hand, the introduction of the Turing test into pop culture may have been harmless, and may even have been beneficial. But that is a sociological question which is outside the scope of this anthology.
The papers in this anthology should be well known to anyone who wishes to express a "professional" opinion on the Turing test. I would also recommend this book to students and lovers of philosophy, especially those with an interest in AI. The selection and ordering of the chapters, as well as the relative neutrality of the essay, are well designed to guide the reader to their own opinion, even though the conclusion is subtly biased.
ABOUT THE REVIEWER:
ABOUT THE REVIEWER
Larry Smith has a PhD in math, and is a research consultant at the U.S. National Center for Biotechnology Information where he works on information retrieval and extraction techniques for the large public collection of biomedical text maintained there. His interests include statistical, syntactic, and semantic language modeling.