Stuart
M. Shieber
Aiken Computation Laboratory
Division of Applied Sciences
Harvard University
April 15, 1993
(Revision 5)
This report appeared in Communications of the Association for Computing Machinery, volume 37, number 6, pages 70-78, 1994. Also available as cmp-lg/9404002 and from the Center for Research in Computing Technology, Harvard University, as Technical Report TR-19-92.
The English logician and mathematician Alan Turing, in an attempt to develop a working definition of intelligence free of the difficulties and philosophical pitfalls of defining exactly what constitutes the mental process of intelligent reasoning, devised a test, instead, of intelligent behavior. The idea, codified in his celebrated 1950 paper ``Computing Machinery and Intelligence'' [28], was specified as an ``imitation game'' in which a judge attempts to distinguish which of two agents is a human and which a computer imitating human responses by engaging each in a wide-ranging conversation of any topic and tenor. Turing's reasoning was that, presuming that intelligence was only practically determinable behaviorally, then any agent that was indistinguishable in behavior from an intelligent agent was, for all intents and purposes, intelligent. It is presumably uncontroversial that humans are intelligent as evidenced by their conversational behavior. Thus, any agent that can be mistaken by virtue of its conversational behavior with a human must be intelligent. As Turing himself noted, this syllogism argues that the criterion provides a sufficient, but not necessary, condition for intelligent behavior. The game has since become known as the ``Turing test'', a term that has eclipsed even his eponymous machine in Turing's terminological legacy. Turing predicted that by the year 2000, computers would be able to pass the Turing test at a reasonably sophisticated level, in particular, that the average interrogator would not be able to identify the computer correctly more than 70 per cent of the time after a five minute conversation.
On November 8, 1991, an eclectic group including academics, business people, press, and passers-by filled two floors of Boston's Computer Museum for a tournament billed as the first actual administration of the Turing test. The tournament was the first attempt on the recently constituted Loebner Prize established by New York theater equipment manufacturer Dr. Hugh Loebner and organized by Dr. Robert Epstein, President Emeritus of the Cambridge Center for Behavioral Studies, a research center specializing in behaviorist psychology. The Loebner Prize is administered by an illustrious committee headed by Dr. Daniel Dennett, Distinguished Professor of Arts and Sciences and Director for Cognitive Studies, Tufts University, and including Dr. Epstein; Dr. Harry Lewis, Gordon McKay Professor of Computer Science, Harvard University; Dr. H. McIlvaine Parsons, Senior Research Scientist, HumRRO; Dr. Willard van Orman Quine, Edgar Pierce Professor of Philosophy Emeritus, Harvard University; and Dr. Joseph Weizenbaum, Professor of Computer Science Emeritus, Massachusetts Institute of Technology. (Dr. I. Bernard Cohen, Victor S. Thomas Professor of the History of Science Emeritus, Harvard University, chaired the committee at an earlier stage in its genesis, and Dr. Allen Newell, U. A. and Helen Whitaker University Professor of Computer Science, Carnegie-Mellon University, and the prize establisher Dr. Loebner served as advisors.)
The prize committee spent almost two years in planning the structure
of the tournament. Because this was to be a real competition, rather
than a thought experiment, there would be several computer
contestants, and therefore several confederates would be needed as
well.
It was decided
that there would be ten agents all together. In the event, six were
computer programs. Ten judges would converse with the agents and
score them. The judges and confederates were both selected from the
general public on the basis of a newspaper employment advertisement
that required little beyond typing ability, then screened by interview
with the prize committee. They were chosen so as to have ``no special
expertise in computer science''.
The committee realized early on that given the current state of the art, there was no chance that Turing's test, as originally defined, had the slightest chance of being passed by a computer program. Consequently, they attempted to adjust both the structure of the test and the scoring mechanism, so as to allow the computers a fighting chance. In particular, the following two rules were added to dramatically restrict Turing's test.
As will be seen, these two rules - limiting the topic and tenor of the discussion - were quite problematic.
The prize committee specified that there be independent referees
stationed in several locations: several in the rooms with the judges
and confederates to answer questions concerning interpretation of the
above rules, and one in the auditorium to serve as a sort of roving
ombudsman. I was a referee in the confederates' room, and can vouch
for the fact that my and my colleagues' efforts there were hardly
needed; the confederates performed admirably. Reports from the other
referees indicated the same for the judges.
Dr. Loebner placed only two restrictions on the setting up of the competition by the prize committee: that a competition be held each year, and that a prize be awarded at each competition. The prize at this first competition was a nominal $1500, although Dr. Loebner has reportedly earmarked $100,000 for the first computer program to pass the full Turing test at some later running of the competition. (Costs for the running of the competition itself were paid for by grants from the National Science Foundation and the Sloan Foundation.)
Figure 1: Mock-up of the form used to implement the scoring method for
the first Loebner competition. The judge writes the letters
corresponding to the terminals in order from least to most human-like,
and draws a line purporting to separate the computer contestants from
the human confederates. In this case, the line has been drawn such that
three of the terminals (F, H, and G) were deemed to be connected to
humans.
To determine the prize-winner, an ingenious scoring mechanism was devised. The Turing test involves a single binary decision, which is either right or wrong. But to determine a winner, the contestants had to be ranked, so each judge was required to place all of the agents in order from the apparently least human to most human. This alone induced the ordering on the basis of which the prize would be awarded. The contestant with the highest average rank would be deemed the winner of the tournament. However, this does not allow a direct reconstruction of the results of the 100 implicit binary decisions that might be made: which of the agents were humans, and which computers. To allow for this to be deduced as well, each judge was requested to place a single line separating the ranked agents into two groups. Those to the right of the line were claimed by that judge to be humans, those to the left computers. (See Figure 1.) The judges were told that at least two of the agents were human confederates, and at least two computer contestants, thus limiting the number of places that the line could be (rationally) placed. The binary decisions could then be read off of the rankings by noting on which side of the line each agent fell. This demarcation process was not used in the awarding of the prize, but was carried out for its informational value alone.
The tournament was to begin at 1 pm on the scheduled Friday. One room of the computer museum was set up with ten terminals for the judges, each labeled with a code letter and the specified topic for conversing with the associated agent. In a back room, hidden from the publicly accessible part of the museum for obvious reasons, five computers had been set up to serve the four confederates. (One terminal was intended to be a backup, and in case it was not needed, to be connected to a publicly accessible terminal so that press and the public could interact with it as a sort of separate Turing test.) In a large auditorium, the ten conversations were projected each on its own screen around the perimeter of the room, and A. K. Dewdney provided running commentary.
Unfortunately, there were serious technical difficulties with the rented computer equipment that had been set up for the confederates. None of the three IBM computers could be made to appropriately interact over the prepared lines with their companion terminal in the judges' room. (The two DEC workstations seemed to work fine.) After almost two hours of unsuccessful last minute engineering, the prize committee decided to begin the competition with only two confederates in place (just the number that the judges had been told was the minimum), reducing the number of agents to eight. The time that each judge had to converse with each agent was shortened from approximately fifteen minutes to approximately seven in order to accommodate the press's deadlines.
The topics chosen by the six contestants were of the sort appropriate for a cocktail party venue (burgundy wines, dry martinis, small talk, whimsical conversation, dissatisfactions in relationships) or perhaps, a child's birthday party (second grade school topics). The two participating confederates chose to converse on Shakespeare and women's clothing. In the end, and perhaps unsurprisingly, the average rankings placed the two human confederates as ``more human-like'' than the six contestants. The highest-ranked contestant, Joseph Weintraub's program (topic: whimsical conversation) was awarded the $1500 prize.
The surprises, such as there were, were in the bipartite classifications. Five judges ranked the top contestant as human, and there were eight instances of such misclassifications of computers as humans overall. (For the record, the contestant discussing relationship problems, the entry of Dr. Kenneth Colby, well known for his PARRY program, which simulates paranoid responses, was twice rated human. The simulated seven-year old accounted for the final misclassification.)
Ms. Cynthia Clay, the Shakespeare aficionado, was thrice misclassified as a computer. At least one of the judges made her classification on the premise that ``[no] human would have that amount of knowledge about Shakespeare.'' [26] Ms. Lisette Gozo was honored as the most human of the agents for her discussion of women's clothing, although one judge rated two computer programs above her.
Perhaps the most conspicuous characteristic of the six computer
programs was their poor performance. It was widely recognized that
computer experts could readily distinguish the contestants from the
confederates. Indeed, many of the techniques being used by the
programs were easily spotted by those familiar with the ELIZA
program that prize committee member Weizenbaum developed in 1965. The
repetition of previous statements verbatim (subject only to pronominal
adjustments, sometimes wrong), answers transparently keyed to trigger
words, and similar tricks of the ELIZA trade were ubiquitous. For
example, the following example from the whimsical conversation program
is illustrative of the regurgitation technique:
01:11:42 Judge 10
It has been a long day! Now I think I'm ready for some fun. The weather isn't very pleasant today. Do you have any jokes that could lift the clouds?
01:13:01 Terminal 5
Everybody talks about the weather but nobody seems to do much about it. Why do you need to tell me now you think you're ready for some fun? Groannnnnn...............
Note that the first sentence in the reply, which concerns the weather, seems to be triggered by the keywords ``weather'' or ``clouds'' in the judge's contribution. The second sentence is comprised of the judge's second sentence, with pronouns transposed, and preceded by the introductory ``Why do you need to tell me''.
Such techniques are subject to well-known and obvious errors, as can be exemplified in the transcripts of interactions with the winning computer entrant. For instance, a naive application of pronoun transposition interacts poorly with tag questions like ``aren't you''.
01:17:52 Judge 10Similarly, repetition of a compound phrase is conversationally bizarre.
You're trying to make this difficult for me aren't you? I just explained that I was ready for fun!
01:18:53 Terminal 5
Would you pre]fer it if it was not true that IO re trying to make this difficult for you aren't you? I just explained that i was ready for fun! I am, therfore I compute.... What makes you think I am trying to make this difficult for you aren't i?
01:35:26 Judge 6
I can't say I understand the question...but the whimsy of a poet is one thing and the whimsy of most of us is another.
01:36:51 Terminal 5
How do you know you can t say you understand the question...but the whimsy of a poet is one thing and the whimsy of most of us is another.
One reporter noted that Weizenbaum himself was ``disturbed'' by how easily people were fooled by these programs [26], and more than one of the judges reported that they were disappointed in the programs' capabilities after their expectations had been raised by interacting with ELIZA in the interviewing process. Dr. Epstein, in a speech after the event, noted that he had learned from the day's proceedings that ``little progress has been made in the last twenty-five years'', that is, since ELIZA. (We address this dubious conclusion below.)
The obvious question, then, is how to reconcile the apparent success of the programs in fooling judges with their patently low technology. Clearly, part of the answer relies on the phenomenon that P. T. Barnum used to amass a fortune. People are easily fooled, and are especially easily fooled into reading structure into chaos, reading meaning into nonsense. This accounts for the popularity of newspaper horoscopes and roadside psychics. This is not a flaw in the human mental capacity. Sensitivity to subtle patterns in our environment is extremely important to our ability to perceive, learn, and communicate. Clouds look like ships, and Rorschach blots seem like vignettes. How much different is interpreting non sequitur as whimsical conversation?
Ned Block, a professor of philosophy at MIT (and by coincidence a
referee at the competition, stationed with the judges) has argued that
the Turing test is a sorely inadequate test of intelligence because it
relies solely on the ability to fool people
[3]. Certainly, it has been known since
Weizenbaum's surprising experiences with ELIZA that a test based on
fooling people is confoundingly simple to pass.
People are even more easily fooled when their ability to detect fooling
is explicitly vitiated, for instance, by a prohibition against using
``trickery or guile''. When I
asked Mr. Weintraub during the post-contest press conference how he
himself would have unmasked his program, his response - typing
gibberish in to see if the program spat it back verbatim at a later time
a la ELIZA - was certainly outside the established rules. In fact,
the referees had discussed that very technique the previous night at a
meeting with the prize committee to calibrate our collective
understanding of the rules. I pointed out to Mr. Weintraub that his
response fell under the ``trickery and guile'' prohibition, and he took
another stab at the question. His second attempt to specify a winning
strategy against his program succumbed to the same problem. (It
involved repeating questions multiple times.)
Weintraub's problem in answering the question points to the craftiness of his solution to the Loebner prize puzzle. His entry is unfalsifiable independent of its performance and solely on the basis of the choice of topic. As almost everyone has noted who was familiar with the rules, whimsical conversation is not in fact a topic but a style of conversation (at least as practiced by Weintraub's program). And whimsical conversation in the mold of Weintraub's program is essentially nonsense conversation, a series of non sequiturs. Thus, when Weintraub's program is unresponsive, fails to make any sense, or shows a reckless abandonment of linguistic normalcy, it, unlike its competitor programs, is operating as advertised. It is being ``whimsical''. At those times when, by happenstance, the program trips over an especially suggestive response, a judge can grab at it as the real article. (The strategy is reminiscent of that used by the program Racter to create ``free verse'' poetry, another unfalsifiable genre.) Weintraub's strategy was an artful dodge of the competition rules. He had found a loophole and exploited it elegantly. I for one believe that, in so doing, he heartily deserved to win.
We might call this winning strategy ``PARRY's
finesse'', after Kenneth Colby's previously mentioned PARRY program
[4]. PARRY was designed to engage in a dialogue in the
role of a paranoid patient. The program was perhaps the first to be
subject to an actual controlled experiment modeled on the Turing test
[5], in which psychiatrists were given transcripts of
electronically mediated dialogues with PARRY and with actual
paranoids and were asked to pick out the simulated patient from the
real. The fact that the expert judges, the psychiatrists, did no
better than chance, has been credited to the fact that
unresponsiveness and non sequitur are typical behaviors of paranoids.
Joseph Weizenbaum's response to the experiment - in the form of his
own model of a deviant mentality - parodies PARRY's finesse
succinctly:
The contribution here reported should lead to a full understanding of one of man's most troublesome disorders: infantile autism.... It responds exactly as does an autistic patient - that is, not at all.... This program has the advantage that it can be implemented on a plain typewriter not connected to a computer at all. [29]
Post hoc thinking of this sort can go a long way to rationalizing the various misclassifications of the whimsical conversation program or, in the same vein, the program that talks at the level of a second-grader. (Who could fail to give a seven-year-old child the benefit of the doubt?) It leads to noting other insidious forms of scoring bias that crept into the competition. One possible source of such bias, for instance, follows from the technical problems that caused two of the confederates to be eliminated. Once the number of confederates had been reduced to the announced minimum, it became impossible for a judge to rationally place the demarcation line between ``humans'' and ``computers'' in such a way as to rate a human as a computer without also rating a computer as a human. Of course, the converse was not true. This might have accounted for one or two more of the errors. Dr. Epstein points out in response to this observation that ``(1) Two of the ten judges drew the line after just one entry, in spite of our instructions. (2) Three of the 5 judges who mistook Weintraub's program for a person rated it above one or both confederates. (3) Two judges mistook a confederate for a computer. In fact, in two (and only two) cases could our instructions have forced the judge to mistake a computer for a person.'' (personal communication to Harry Lewis, 1992) The third point is, of course, irrelevant, the first hardly gratifying, the second accounted for by Weintraub's use of PARRY's finesse, and the final comment is exactly my point.
But post hoc rationalization, like telling your boss off, may be enjoyable at the moment, but is, in the long run, ungratifying. The important questions do not involve microanalysis of the particular competition as run several months ago, but the larger questions of the purpose, design, and even existence of the Loebner prize itself.
There is a long history of argumentation in the philosophical literature opposing the appropriateness of the Turing test as a litmus test of intelligence. Certain arguments against the effectiveness of the test in answering questions about the intelligence of computers or the possibility of human thought center around the behaviorist nature of the test. Intelligence, it may be claimed, is not determinable simply by surface behavior. Variants of this argument have been given by Block [2], Gunderson [15], and Searle [24][23]. Others have suggested that the Turing test is not sufficient in that the behaviors under adjudication are too limited [10][15]. On the basis of such counterarguments, Moor [18] has argued for a drastically limited view of the Turing test, not as an operational definition of intelligence at all, but rather as a mode for accumulating evidence leading to an inductive argument for the intelligence of the machine. (See the reply by Stalker [25] and a later clarification by Moor [19] for further arguments.) Moor [20] provides a good introduction to these issues. French [11] provides a strong argument that as a sufficient condition for intelligence, the Turing test is so difficult as to be uninteresting. Nonetheless, none of these sorts of presumptive counterarguments to the use of a Turing test are the basis for the discussion in the remainder of this paper. The issue of whether an operational definition of intelligence is appropriate, and whether the particular definition codified in the Turing test is too narrow, though important questions, can be taken as resolved in favor of the Turing test for the purposes of the present discussion. Thus, we will side with the behaviorist interpretation favored by the organization administering the prize, the Cambridge Center for Behavioral Studies. Nonetheless, these arguments do provide another strong basis on which to question the appropriateness of the Loebner prize. A full discussion is, unfortunately, well beyond the scope of this paper, but readers are urged to consult the cited literature. Having sided, for the nonce, with the philosophical appropriateness of Turing's design as a test of intelligent behavior, we turn to the question of whether the Loebner prize competition is itself an appropriate enterprise.
Prizes for technological advances have existed before, and much can be
learned by comparison with previous exemplars.
Just as humankind has dreamed of mimicking the human power of thought,
so have we longed to possess the avian power of flight. Human-powered
flight entered the mythology of the ancient Chinese and Romans, the
designs of da Vinci, yet was only accomplished within the last
generation as a direct result of a prize set up for the express
purpose of promoting that technology. The Kremer prize, established
in 1959 by British engineer and industrialist Henry Kremer, provided
for an award of £5000 for the first human-powered vehicle to
fly a specified half-mile figure-eight course. It was awarded in
1977, less than twenty years later, to a team headed by Paul Macready,
Jr., for a flight by Bryan Allen in the Gossamer Condor.
The success of the Kremer prize depended on two factors.
The
task was just beyond the edge of the current technology.
Unfortunately, since our ability to dream far outstrips our ability to
build, the establishment of tests of ridiculous difficulty is not
difficult to imagine. At a time when an award-winning human-powered
flight was one of one meter at an altitude of 10 centimeters (the 1912
), the Paris newspaper La Justice established
a prize for the first nonstop human-powered flight from Paris to
Versailles and back. (It was never won.)
The history of human-powered flight indicates that only when the
purpose of the prize is clear and the task is just beyond the edge of
current technology is a prize an appropriate incentive. The Kremer
prize is a prime example of a prize that meets these criteria. The
Loebner prize is not.
We turn first to the goals of the Loebner prize. It was, according to the formal statement in the competition application, ``established...to further the scientific understanding of complex human behavior.'' Along these lines Dr. Loebner has been quoted as saying ``People had been discussing the Turing test; people had been discussing AI, but nobody was doing anything about it.'' [17] The several thousand members of the American Association for Artificial Intelligence may be surprised to learn that nobody is doing anything about it.
Others have argued that the prize will serve to publicize the Turing test, thereby increasing the public's awareness and understanding of artificial intelligence. Increased public understanding of AI is certainly a laudable goal, especially since the regular appearance of superficial popularizations in the press serves more to mislead the public by alternately raising and dashing expectations than to inform it by cogent coverage of actual results. A flurry of the standard stories in the press like ``Computer fools half of human panel'' [13] and ``Test a breakthrough in artificial intelligence'' [16] was certainly one of the side effects of the Loebner prize competition, but perhaps not a laudable contribution.
Overselling of AI by the media (and, occasionally,
practitioners) has, in its brief history, been a repeated and persistent
problem, and the hubristic claims of the organizers of the Loebner
prize that they are ``confident that within 10 to 20 years a system
will pass this electronic litmus test'' [27] perpetuates
the hyperbole. Robert Epstein, in his recent article describing the
event, its genesis, and his speculations as to its importance,
constructs a standard claim of this sort:
Thinking computers will be a new race, a sentient companion to our own. When a computer finally passes the Turing Test, will we have the right to turn it off? Who should get the prize money - the programmer or the computer? Can we say that such a machine is ``self-aware''? Should we give it the right to vote? Should it pay taxes? If you doubt the significance of these issues, consider the possibility that someday soon you will have to argue them with a computer. [][emphasis in original]epstein
Not surprisingly, the winner of the Loebner prize has jumped on the
publicity bandwagon by taking out an advertisement pushing his program
as the ``first to pass the Turing Test''. Conversely, a prize whose execution convinces
fellow scientists mistakenly that little progress has been made in a
quarter century does little to promote the field. In summary, there
is a difference between publicity and increased public understanding.
Events of this sort - and the Loebner competition has been no
exception - tend to generate the former rather than the latter.
Dennett has hinted at a completely different goal for the Loebner prize. ``It is useful to have the demonstration of the particular foibles that human beings exhibit in 1991.... We won't learn much about AI from the Loebner prize for a long time, but we will learn some non-negligible things about social psychology, perhaps, in the meantime.'' (Dennett, personal communication) For instance, the competition might be justified ``as a proving ground for the environmental conditions necessary to permit the Turing test to someday occur. In other words, the Loebner competition can tell us what we need to know about how humans behave in computer mediated interactions.'' (Dranoff, personal communication) This line of teleology for the Loebner prize, that it serves not as a test of the abilities of the computers but of the psychologies of the various participants, has often been proposed informally. Such a ``conspiracy theory'' of the prize as a vast psychology experiment executed on unwitting and unconsenting adults is as unlikely as it is disturbing. Of course, there is already an extensive literature on how humans behave in computer-mediated interactions, and the Loebner competition is not likely to contribute to it; it was not designed or executed as a controlled scientific experiment, nor was that its apparent intention, despite the hopes of Dennett and Dranoff that firm conclusions in psychology might be gleaned from it.
Thus, it is difficult to imagine a clear scientific goal that the
Loebner prize might satisfy. Turing's test as originally defined, on
the other hand had a clear goal, to serve as a sufficient condition
for demonstrating that a human artifact exhibited intelligent
behavior. Even this goal is lost in the Loebner prize competition.
By limiting the test, it no longer serves its original purpose (and
arguably no purpose at all), as Turing's syllogism
fails. It is
questionable whether the notion of a Turing test limited in the ways
specified by the Loebner prize committee is even a coherent one. The
prize committee spent some time with the referees attempting to
explicate the notion of ``natural conversation without trickery or
guile''. It was suggested that a criterion be used as to whether you
might say the utterance in conversation with a stranger seated next to
you on an airplane. For instance, what might a competition judge
legitimately ask on the topic of Washington, DC? Certainly, the
question ``Are there any zoos in Washington?'' is the kind of thing
you might ask a stranger when flying to the capital for the first
time, whereas ``Is Washington bigger than a breadbasket?'' is just as
certainly a trick question. What about ``Is there much crime in
Washington?'' Undoubtedly acceptable. ``Are there any dogs in
Washington?'' An odd question for an airplane conversation. ``Are
there many dogs in Washington?'' Sounds better. ``Are there many
marmosets in Washington?'' Odd. ``Are there many marmosets in the
Washington zoo?'' Okay again. The exegesis of such examples begins
to sound like arguments about angels and sharp objects.
Similar problems accrue to the notion of limiting the topic of discourse. Is the last question about Washington, DC or marmosets? (One of the referees in fact thought that this and similar questions should be ruled out as it was not strictly on the topic of the city alone.) How about ``Are the buildings in Washington very modern?'' Perhaps a question about architecture, as the following question surely is: ``Do you know any examples of neo-Georgian architecture in Washington?'' Are culinary topics ruled out, as in ``What foods is our nation's capital best known for?'' Such issues are not idle in the context of the Loebner competition. Cynthia Clay, the Shakespeare expert, was asked why Mario Cuomo has been referred to recently as ``Hamlet on the Hudson''. The question caused much consternation among the referees peering over Ms. Clay's shoulder. Her response was ``His brooding'' after which she coolly changed the topic back to Shakespeare. Or had it ever left?
The reason that Turing chose natural language as the behavior definitional of human intelligence is exactly its open-ended, free-wheeling nature. ``The question and answer method seems to be suitable for introducing almost any of the fields of human endeavor that we wish to include.'' [page 435]turing-mind In attempting to limit the task of the contestants through limiting the domain alone, the prize committee succeeded in doing neither.
The distinction between domain and task is crucial. Finance is a domain, but not a task; withdrawing money from a bank account is a task, one that is achievable through both human and computer intermediaries these days; taking dictation of a funds-transfer request is a task that only humans can currently undertake with reliability. Had Babbage limited his differential analyzer to multiply only even numbers, the design would have been no more successful. This is a limitation of domain that does not yield a concomitant limitation in task.
It is well understood in the field that natural-language systems must
be tested using a constrained task. Currently, standard limited tasks
can be found in evaluation of natural-language database retrieval
systems (like withdrawing money from a bank account on the basis of a
natural-language request) and speech recognition systems (like
transcribing a spoken funds-transfer request). The tasks, typically
undertaken with limited vocabulary, are easily quantifiable along
several dimensions (for example, technical notions of precision,
recall, overgeneration, perplexity) independently of the subjective
judgments of lay judges. In addition, they can be adjusted to sit
just at the edge of technology (a topic we return to below) unlike the
Turing test itself. The natural-language-processing research
community has used such tests for some time now, and there has been
increased interest in issues of evaluation of systems (primarily at
the behest of funding agencies) over the last few years; whole
conferences have been devoted to the subject (see, for instance, the
report by Neal and Walter [21]).
In summary, the Loebner prize competition neither satisfies its own avowed goals, nor the original goals of Alan Turing. In fact, it is hard to imagine a scientific goal that establishment of the Loebner prize provides a better route to than would be provided by other uses of Dr. Loebner's $1500, his $100,000 promissory note, and the $80,000 in ancillary grants from the National Science Foundation and the Sloan Foundation. (Nonscientific goals are much easier to imagine, of course.)
Now to the second criterion for an appropriate technology prize, that
the task be just beyond the edge of technology. Imagine that a prize
for human-powered flight were set up when the basic science of the
time was far too impoverished for such an enterprise, say, in da
Vinci's era. The da Vinci prize, we shall imagine, is constituted in
1492 and is to be awarded to the highest human-powered flight. Like
the Loebner prize, a competition is held every year and a prize must
be awarded each time it is held. The first da Vinci competition is
won by a clever fellow with big springs on his shoes. Since the next
competition is only one year away (no time to invent the airfoil), the
optimal strategy is universally observed by potential contestants to
involve building a bigger pair of springs. Twenty-five years later,
the head of the prize committee announces that little progress has
been made in human-powered flight since the first round of the prize
as everyone is still manufacturing springs.
Of course, a lot of progress had been made in human-powered flight in those twenty-five years. Da Vinci himself was studying human physiology and anatomy and the flight of birds, and - although his own work directly on the topic of human-powered flight, ornithopter design, was essentially meritless beyond its decorative qualities - the apparently tangential work was, in the long run, pertinent to the technologies that would eventually enable the Gossamer Condor to be constructed. (See, e.g., Gibbs-Smith [12].) However, over that period, and indeed at every point during the following four centuries, the kind of progress that needed to get made to solve the problem was not directly observable at that time in improvement in solutions to the problem, the kind of improvement that might be observable in an annual contest. Nonetheless, tremendous scientific progress was made between the fifteenth and twentieth centuries. The Gossamer Condor and the digital computer are two outgrowths of this progress.
The field of artificial intelligence is in that kind of
state. The AI problem, like the problem of human-powered flight
in the Renaissance, is only addressed directly and dismissed as
imminently solvable by those who underestimate its magnitude.
Progress on restricted tasks in limited domains is well documented in
the literature on applications of artificial intelligence. But
progress on the underlying science that has been made in the last
twenty-five years, important though it is, is not of the type that
allows incremental advantage to be demonstrated on the big problem,
the full-blown Turing test, nor should this be seen as a failing of a
field addressing a problem of the scope and magnitude of human
intelligence. (And like all scientific endeavors, a lot of time can
be spent on fruitless avenues of attack; ELIZA, as a discipline for
natural-language processing, was such a fruitless avenue. It was
quite fruitful in other areas, however, as cogently argued by
Weizenbaum himself.) Indeed, one aspect of the progress made in
research on natural-language processing is the appreciation for its
complexity, which led to the dearth of entrants from the artificial
intelligence community - the realization that time spent on winning
the Loebner prize is not time spent furthering the field.
Twenty-five years of progress in the fields associated with the Turing test - artificial intelligence, computational linguistics, and natural-language processing - cannot be summarized in a single program, but is captured in the many small results, some of which, some day, at an unpredictable time in the future, may lead to a dramatic demonstration of apparently intelligent artificial behavior. To expect more is hubris. What is needed is not more work on solving the Turing test, as promoted by Dr. Loebner, but more work on the basic research issues involved in understanding intelligent behavior. The parlor games can be saved for later.
Given that the Loebner prize, as constituted, is at best a diversion of effort and attention and at worst a disparagement of the scientific community, what might a better alternative use of Dr. Loebner's largesse be? The goal of furthering the scientific understanding of complex human behavior is no less laudable now than it was before the competition, but clearly, a direct assault on a valid test of intelligent behavior is out of the question for a long time; even the prize committee well appreciates that. Thus, any award or prize based on a behavioral test must use a limited task and domain, so that the envelope of technology is pushed, not ignored. The efforts of the Loebner prize committee to design such a test have failed in that the test that they developed rewards cheap tricks like parrying and insertion of random typing errors. This is an (indubitably predictable) lesson of the 1991 Loebner prize competition.
This problem is a general one: Any behavioral test that is sufficiently constrained for our current technology must so limit the task and domain as to render the test scientifically uninteresting. Adjusting the particulars of the Loebner competition rules will not help. By way of example, many years of effort have gone into the design of the tests of natural-language-processing systems used at the annual DARPA-sponsored Message Understanding Conferences. The trend among entrants over the last several conferences has been toward less and less sophisticated natural-language-processing techniques, concentrating instead on engineering tricks oriented to the exigencies of the restricted task - keyword-spotting, template-matching, and similar methods. In short, this is because such limited tests are better addressed in the near term by engineering (building bigger springs) than science (discovering the airfoil). Behavioral tests of intelligence are either too hard for a prize or too rewarding of incidentals.
At this stage, objective behavioral tests must give way to subjective
evaluative ones. A more appropriate way to reward novel, potentially
breakthrough-inducing efforts toward the eventual goal of mimicking
intelligent behavior would be to institute a prize for just such
efforts, on the model of the Nobel prizes, ACM's Turing award, and
similar subjectively determined awards. Rather than awarding lifelong
achievement or past accomplishments, however, the prize could be
awarded for particular discoveries, regardless of field, that the
committee determined were of sufficient originality, import, and
technical merit and that were deemed contributory to Turing's goal
(even though they may provide no incremental edge in a current-day
Turing test). To avoid rapt obeisance to AI conventional wisdom, the
awards committee would include eminent thinkers from a wide range of
related fields (much as the current Loebner prize committee does) but
to ensure technical fidelity, a nominating committee of researchers
from the pertinent technical fields should verify purported results
before passing them on for consideration. In order to prevent
degrading of the imprimatur of the reconstructed Loebner prize, it
would be awarded on an occasional basis, only when a sufficiently
deserving new result, idea, or development presented itself. I am not
ostentatious enough to provide examples that I believe would be
appropriate for such an award; I am sure that the reader can imagine
one or two.
As the years elapsed, and the speculations of this Loebner prize committee as documented in their past decisions began to prove perspicacious, the Loebner prize might grow in stature to that of the highly sought prizes of other scientific areas, and so provide a tremendous motivation for innovative ideas in the quest for an artificial intelligence.
The Second Annual Loebner Prize Competition was held at the Cambridge Center for Behavioral Studies on December 15, 1992. The number of computer entrants had decreased from six to three, with Joseph Weintraub's program, complete with the winning strategy from the previous year's competition, taking first prize once again, this time under the purported topic ``men vs. women''. Bigger springs had prevailed.
The research in this paper was supported in part by grant IRI-9157996 from the National Science Foundation and by matching funds from Xerox Corporation.
I am grateful to the many readers of earlier drafts of this paper: Ned Block, Noam Chomsky, Jacques Cohen, Daniel Dennett, Susan Cole Dranoff, Barbara Grosz, Harry Lewis, David Mumford, Fernando Pereira, Jeff Rosenschein, David Yarowsky, and two anonymous reviewers. I have incorporated many of their thoughtful comments into the paper, although the opinions presented here are my own, and should not be taken as necessarily representative of the previous readers' views.
Lessons from a
Restricted Turing Test
This document was generated using the LaTeX2HTML translator Version 0.6.4 (Tues Aug 30 1994) Copyright © 1993, 1994, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
The command line arguments were:
latex2html -split 0
loebner-rev-html.tex.
The translation was initiated by on Mon
Jun 2 18:37:14 EDT 1997 and modified by hand thereafter.