“Computing Machinery and Intelligence.”
1 It is frequently claimed that,
in that paper, Turing proposed a test
for machine intelligence.
Those who believe that Turing proposed a test for machine intelligence
should read that paper. Turing understood that science requires agreement
on how to measure the properties being discussed. Turing rejected “Can
machines think?” as an unscientific
question because there was no measurement-based definition of “think.”
That question is not one that a scientist
should try to answer.
Turing wrote: “If the meaning of the
words ‘machine’ and ‘think’ are to be
found by examining how they are commonly used it is difficult to escape the conclusion that the meaning and the answer
to the question, “Can machines think?” is
to be sought in a statistical survey such as
a Gallup poll. But this is absurd. Instead of
attempting such a definition I shall replace
the question by another, which is closely
related to it and is expressed in relatively
Turing’s proposed replacement
question was defined by an experiment.
He described a game (the imitation
game) in which a human and a machine
would answer questions and observers
would attempt to use those answers to
identify the machine. If questioners
could not reliably identify the machine,
that machine passed the test.
Turing never represented his replacement question as equivalent to
“Can machines think?” He wrote, “The
original question, ‘Can machines think?’
I believe to be too meaningless to deserve
discussion.” A meaningless question
cannot be equivalent to a scientific one.
Most of Turing’s paper was not
about either machine intelligence
or thinking; it discussed how to test
whether or not a machine had some
well-specified property. He also speculated about when we might have a machine that would pass his test and demolished many arguments that might
be used to assert that no machine
could ever pass his test. He did not try
to design a machine that would pass
his test; there is no indication that he
thought that would be useful.
Joseph Weizenbaum’s Eliza
Anyone interested in the Turing Test
should study the work of the late MIT
Writing game-playing programs is
harmless and builds capabilities. How-
ever, I am very concerned by the pro-
posal that practical products should ap-
ply human methods. Imitating humans
is rarely the best way for a computer to
perform a task. Imitating humans may
result in programs that are untrust-
worthy and dangerous.
To explain my reservations about AI,
this column discusses incidents from
the early days of AI research. Though
the stories are old, the lessons they
teach us remain relevant today.
AI researchers sometimes describe their
approach as “heuristic programming.”
An early CMU Ph.D. thesis defined a
heuristic program as one that “does not
always get the right answer.” Heuristic
programs are based on “rules of thumb,”
that is, rules based on experience but not
supported by theory.c
“Heuristic” is not a desirable attri-
bute of software. People can use rules
of thumb safely because, when rules
suggest doing something stupid, most
people won’t do it. Computers execute
their programs unquestioningly; they
should be controlled by programs that
can be demonstrated to behave cor-
rectly in any situation that might arise.
The domain of applicability of a pro-
gram should be clearly documented.
Truly trustworthy programs warn their
users whenever they are applied out-
side that domain.
Heuristics can be safely used in a
˲ The specification allows several acceptable solutions and the heuristic is
used either to select one of them or to
determine the presentation order.
˲ The heuristic is intended to speed
up a program that conducts a search
that will either find a solution or establish that there is no solution.
In other situations, heuristic programming is untrustworthy programming.
What Alan Turing Really Said
Alan Turing is sometimes called the
“Father of AI” because of a 1950 paper,
c Those who write heuristic programs rarely
characterize the set of conditions under which
the program would produce an incorrect result. See the section “An AI System for Constructing Parsers.”
ACM’s Interactions magazine
explores critical relationships
between people and
emerging innovations and
industry leaders from around
the world across important
applications of design thinking
and the broadening ;eld of
Our readers represent a growing
community of practice that is
of increasing and vital global
To learn more about us,
visit our award-winning website
Follow us on
Facebook and Twitter
IX_XRDS_ThirdVertical_V01.indd 1 3/18/15 3:35PM