The trouble with artificial intelligence – making a machine that can
function as if it had human powers of reasoning – is that everyone understands
it. To many people it is clear that such a thing is impossible. Yet we are
surrounded by intelligent machines that are becoming more intelligent all
the time; and this is equally clear to others. Lots of people know that
machines can’t make inferences, but only follow rules blindly – while others
are sure that if you put enough rules into a computer it will be indistinguishable
from a person. Still others believe that the new neural network computers
are just like human brains and can function without rules. What is going
on – and what are intelligent machines?
In 1972 Hubert Dreyfus, a philosopher at the University of California
at Berkeley, wrote a book called What Computers Can’t Do. He argued that
the potential of intelligent machines was strictly limited. For many years
after its publication the book was violently attacked, and Dreyfus versus
the ‘artificial intelligentsia’ is still a hot topic and the frequent subject
of television debate.
Dreyfus made some predictions; some commentators say that the predictions
have failed, falsifying his argument, while others, including myself, believe
they have been a triumphant success. Dreyfus’s main line of argument was
about rules. Like the philosopher Ludwig Wittgenstein, Dreyfus states that
‘rules do not contain the rules of their own application’. That is, rules
cannot contain all the necessary information about the contexts in which
they should be applied, because that would require further rules, with more
rules to explain these rules and so on, ad infinitum. Consider, for example,
the following quotation from the Arizona Daily Star of 31 May 1986. ‘A rookie
bus driver, suspended for failing to follow correct emergency procedures
when a girl suffered a heart attack on his bus, was following overly strict
rules that prohibit drivers from leaving their routes without permission,
a union official said yesterday. ‘If the blame has to be put anywhere, put
it on the rules that those people have to follow’ (said the official). (A
spokesman from the bus company defended the rules: ‘You give them a little
leeway, and where does it end up?’ ‘
Advertisement
A computer programmed in a conventional way is like the bus driver:
it can follow rules but only in one way – it cannot recognise when the context
requires that the rules be broken, or followed in a different way.
Now, if you were determined to create an electronic bus driver, you
could put in a rule that says ‘Always stick to the route unless a passenger
suffers a heart attack, in which case drive to the nearest hospital’, and
that would take care of incidents like the one in Arizona. However, the
full range of contextual possibilities cannot be anticipated without an
infinite series of additions. For example, the ‘heart attack rule’ would
need to be modified with clauses about what happens if the nearest hospital
is closed because of an epidemic of legionnaire’s disease, or if an ambulance
with a full emergency team is driving by, or if the bus has been hijacked
and that is the cause of the passenger’s heart attack, and so on. Admittedly,
human beings such as the bus driver get these things wrong sometimes, but
there seems no way to program the computer to get them right in the way
that humans usually do. Let us call this problem ‘the rules critique’.
Dreyfus argued that computers would be successful in areas where rules
could cover every eventuality. For example, he said that there was no reason
why computers should not be very good at noughts and crosses and simple
translations based on dictionary definitions. But they would be bad at translating
languages; this requires sensitivity to context. They would also be bad
at chess, which, while open to exhaustive definition in principle, has so
many possible moves that a computer which encapsulated them would need more
memory location than there are particles in the universe. But Dreyfus was
wrong about chess: computers have turned out to be much better at the game
than he envisaged.
It is not at all clear, however, that Dreyfus’s ‘mistake’ works to his
disadvantage. The surprising thing about chess is that computers can play
it so well using ‘brute strength’ programming. Until recently most people
thought that the only path to success in computer chess was mimicking human
styles of play, including recognising general patterns and other ‘rules
of thumb’. It turns out that if the machine can look well ahead of a human
by simply working through many more possible moves, it can win even though
it cannot see to the end of the game. But the crucial point is that brute
strength success in chess says nothing about the abilities of computers
to manage other context-sensitive tasks such as translating languages, because
the successful chess computers are no more sensitive to context than the
earlier unsuccessful ones.
Computerised chess is misleading because the criterion of success is
winning the game, not mimicking human abilities. We are not tempted to make
the same mistake where physical abilities are concerned. There is no doubt
that a tractor can beat a human in a tug of war. But this proves nothing
about the tractor’s ability to have any of the other human competencies,
such as ceasing to pull when victory would mean dragging your opponents
onto a previously unnoticed rattlesnake. So, it seems that even if some
of Dreyfus’s predictions have been bypassed by ‘brute strength’ programming,
the rules critique still makes AI impossible.
Dreyfus’s classification of computable tasks and those that would prove
impossible has turned out to be remarkably prescient. However, if his own
rules critique is properly applied, it disallows even more types of intelligent
machines than he did. Therefore, the rules critique cannot be the whole
story.
For example, according to the rules critique even pocket calculators
are impossible. Imagine that I ask you to continue the series 2, 4, 6, 8
. . . This may look like a straightforward, fully definable application
of logic or arithmetic, but it is just as sensitive to context as the case
of the bus driver. For instance, in certain circumstances the correct continuation
of 2, 4, 6, 8 . . . is not ’10’, but ‘Who do we appreciate?’
Suppose I go back to trying to get round the series continuation problem
by giving a tighter set of instructions. Instead of saying ‘continue the
series’, I might say ‘add 2 to the last term’. But to follow that rule you
have to know how to ‘add 2′ in the way I intended it – you have to know
the context for that instruction. For example, one way of adding 2 to the
last term of the series is by writing ’28’. This is adding 2 – where there
was once an 8 there is now a 2 and an 8.
SEE WHAT I MEAN?
With a little ingenuity one discovers that this game of ‘awkward student’
can go on for ever. Getting someone to do what you want depends on their
complicity and their ability to see what you mean. The need for complicity
in understanding instructions is what is exploited when people create havoc
by ‘working to rule’.
The same applies throughout arithmetic. Take the sum ‘7 divided by 11
multiplied by 11’. What should the answer be, and why? It is not quite so
obvious as it seems at first. There are two choices. The obvious answer
is 7. But if you work it out in strict sequence, you will get 6.99999 with
a tail that goes on and on for as long as you have patience and ends in
a 6 or a 3. (You might decide to call it 6.9 recurring.) We usually choose
7 as the answer although there are occasions when one of the other answers
is more appropriate; it all depends on the context. For example, one of
the other choices is often more appropriate in the classroom, and certainly
more appropriate when thinking about computer rounding algorithms. If it
is true that simple arithmetic is context-dependent, and requires rules
upon rules upon rules to give the desired answer in the right context, then
Dreyfus’s argument outlaws not only language translators but arithmetic
machines too. This suggests there is something wrong with the argument.
I believe the correct answer to the problem of how machines mimic human
abilities is almost trivial. Machines must be able to mimic some of what
we do because we can mimic them. If I am (correctly) mimicking a machine
then it must be doing what I am doing. There is a class of human action
that is ideally suited to the mimicry of things. I call it ‘behaviour specific
action’. How we do most of the things we do depends on the circumstances
in which we do them. But there are tasks where only the physical movements,
not the context, are important.
In factory work, for example, a worker on a production line might be
told to ‘stand there and put wheels on the cars as they come past’. This
instruction will normally be carried out with many subtle variations which
are crucial to success. The worker has to compensate for variations in the
way the job is presented. Is there a piece of debris on that wheel? Is this
wheel a tight fit on the studs so that it needs a bit of ‘help’ to get it
on? The worker is also likely to vary the routine so as to relieve boredom.
On the other hand, there is an ideal of production line work represented
by ‘scientific management’ methods based on ‘time-and-motion’ studies, as
caricatured in Charlie Chaplin’s film Modern Times, in which the task is
carried out with identical physical movements every time. Under such a regime,
workers try to standardise ‘behaviour’.
It is this ideal that I call behaviour-specific action. It is quite
different from the way we normally repeat actions, with all those subtle
variations of behaviour to fit the context. It is worth noting that though
we normally associate repetitive behaviour with demeaning, mindless automatism,
it is just as frequently valued highly, as in the golf swing or competitive
high-board diving.
DOUBLE STANDARDS
The two ways of doing a task have their advantages and disadvantages.
A disadvantage of the standardised behaviour-specific action is that in
most cases it does not work very well. If enforced rigorously it has the
same stultifying effect on the workplace as a work to rule. Another disadvantage
is that it is hard for humans to standardise their behaviour. On production
lines it is mind-numbing, whereas on parade grounds, diving boards and golf
courses the reproduction of identical movements time after time takes a
lot of training.
A great advantage of behaviour-specific action, as opposed to normal
action, is that the human actor can be replaced by any machine that can
reproduce the behaviour. So we can build machines with a perfect golf swing
because the essence of the game is repetition of behaviour. But tennis is
different because there is an unpredictable opponent to beat. We can replace
workers on production lines with machines but only after we have organised
the factory so that the tasks can be accomplished through standardised behaviour,
rather than varying actions.
We can now explain those parts of arithmetic that a pocket calculator
can do. The mental routines of arithmetic are like routinised behaviour.
These routines actually become behaviour in, say, the standardised movements
required to do simple sums on a slide rule or abacus. Just because they
are sometimes performed with the head rather the hands does not make them
different in kind.
The theory of behaviour-specific action, then, gives us a way of explaining
the successes of existing machines. They are successful when they do what
we try to do with behaviour-specific routines. It is possible to extend
their capability by breaking down normal action into a series of behaviour-specific
acts, and this is what we do when we add more and more rules to cover the
exceptions in a computer program. ‘Cobbling together’ an ‘intelligent’ program
is the attempt to mimic action with lots of little pieces of behaviour.
Up to a point it will be successful, but it is not the same as reproducing
normal action. Thus the theory accounts for success while still explaining
failure.
The difference between this theory and the rules critique of Dreyfus
is that success is not accounted for by having spelt out all the rules for
an action but by our determination to try to execute certain acts by making
them identical to behaviour. We have not succeeded in completely defining
the context of an action; rather, we try to redefine certain acts as though
they were context independent. We do not always succeed, of course, but
in those cases we prefer the job that a machine could do to our own handiwork.
IMITATION INTELLIGENCE
The theory puts ‘intelligent’ machines on the same continuum as machines
designed primarily for physical work, and these are in turn continuous with
tools of various sorts and even with statues and natural objects. Can a
tree imitate a human? Yes, just to the extent that a human can imitate a
tree. Can the Chinese room imitate a native speaker? Yes, just to the extent
that the native speaker can imitate one who has learned the language from
dictionaries and primers (see Box 1 for an explanation of the Chinese room).
It is when we imitate machines with our heads rather than our bodies we
are tempted to think of them as intelligent.
The predictions that follow from the theory of behaviour-specific action
map quite closely onto Dreyfus’s – they need to because his have been so
successful – but there are some interesting differences. The first difference
has already been seen. Because the theory does not lead one to expect to
find whole domains in which the context can be completely pinned down by
a set of rules, behaviour-specific action draws attention to the bits of
arithmetic and so forth that calculators and other computers cannot do very
well. Try your calculator, or even your mainframe, on 7 divided by 11 multiplied
by 11; either that or some such sum will not work because computers do not
understand the context dependency of approximation. Designers of rounding
algorithms for computers are faced with a specific example of the general
problem of AI. The theory of behaviour-specific action says that only certain
acts are carried out in a behaviour-specific way and there is a lot of arithmetic
that cannot be done this way.
A second difference relates to our expectations about the future of
those intelligent machines that are not programmed with explicit rules.
The version of programming without rules that is currently attracting attention
is the neural network.
A neural network is an array of processors interlinked by connections
that can be strengthened or weakened; the concept takes its inspiration
from the interconnected neurons of the brain. A neural network is ‘trained’
by being given a series of examples of correct response and the connections
between its processors are strengthened or weakened according to its success
in reproducing what is wanted.
The neural network is never given an explicit body of rules to follow;
its ‘program’ is contained in the strengths and weaknesses of the different
links in the network. Because there are no specific rules, the rules critique
does not obviously apply. Could it be that the equivalent of the abilities
associated with normal human socialisation and sensitivity to context can
somehow be reproduced by training these brain-like computers? The Dreyfus
version of the rules critique seems to allow that they could. Dreyfus himself
is ambivalent on the issue.
The theory of behaviour-specific action, on the other hand, makes no
distinction between one kind of machine and another. Machines can only
reproduce tasks that can be broken down into series of behaviour-specific
acts. The only kind of machine that would be able to do more would be one
that could learn context-sensitivity in the same way that we do. But no
one knows how we do it, except that it has to do with being brought up in
social groups. We learn when to say ’10’ as the continuation of a series
and when to say ‘Who do we appreciate?’, not by being told, and not by
being given lots of examples that we can apply in the future, but by being
made members of a society. We learn it in just the same way as we learn
to speak English, or Chinese, or whatever is our natural language, and the
process is just as mysterious.
Harry Collins is professor of sociology and director of the Science
Studies Centre at the University of Bath. His latest book Artificial Experts:
Social Knowledge and Intelligent Machines is published by MIT Press.
* * *
1: Intelligence and the Chinese room
In 1980, John Searle, a philosopher at the University of California
at Berkeley, published his well-known ‘Chinese Room’ argument; he presented
his thesis again in the BBC’s 1984 ‘Reith Lectures’. The argument posits
a person who does not speak Chinese, concealed in a room with dictionaries
and a filing system. Written questions in Chinese are passed into the room.
The person uses the dictionaries and filing system to construct written
answers, also in Chinese. Searle argues that the answers coming from the
Room might be indistinguishable from those that would be produced by a native
Chinese speaker, even though the person does not understand what the questions
and answers mean. The questioner has had an intelligent written conversation
in Chinese with the Room. Searle says that this shows that producing apparently
intelligent behaviour cannot be taken to indicate true intelligence.
This line of thought has prompted a big debate. For example, one attempt
at an answer has been to claim that though the person in the room does not
understand Chinese, the Room as a whole does.
Though Searle’s argument may be philosophically sophisticated, to a
sociologist it seems misplaced. Suppose my neighbours do have heads full
of mechanisms that, like the Chinese Room, produce apparently intelligent
behaviour, while in reality they understand nothing. How would I ever know?
Barring postmortem investigation, behaviour that is truly indistinguishable
from that produced by an intelligent being might as well represent intelligence.
The more interesting questions concern, first, whether the person in
the Chinese Room could produce Chinese converesation indistinguishable
from that spoken by native Chinese and, secondly, what we might look for
to distinguish the two. For these, Dreyfus’s arguments and the theory of
behaviour-specific action are more relevant.
* * *
2: The Boston Turing test
Alan Turing, one of the founders of modern computing, suggested that
the way to investigate the intelligence of computers is to test them against
people. What he proposed is that an interrogator should conduct a typed
conversation with a computer and with a person and try to work out from
the answers which is which. If the interrogator’s judgements work out no
better than chance, then the machine ought to be thought of as intelligent.
Turing has now been taken at his word. On 8 November 1991 in Boston,
Massachusetts, a number of programs were interrogated by a panel of judges.
A program called PC Therapist III won a prize for being least distinguishable
from human controls. On 16 November The New York Times carried the headline
‘A computer mistaken for a human’. Does this mean that PC Therapist III
really is intelligent?
Searle’s Chinese room argument says that a machine can pass the Turing
test even if it does not understand. Nevertheless, it seems to me that the
Turing test, if conducted carefully, is a perfectly good test of what could
reasonably be called intelligence. Properly conducted, the test could be
used to detect socialisation and context sensitivity. To do this the interrogator
would have to ask questions that look for the computer’s ability to handle
language in context.
But PC Therapist III would not satisfy even this criterion, because
the Boston competition was set up in just such a way as to conceal context
sensitivity. The question sessions were restricted to a narrow domain for
each program, and the conversation had to be in normal English. In other
words, the competition was designed to test for the sort of English that
could be learned from a dictionary and a primer on English grammar. There
is no reason to think that computers should not do well in such circumstances
and it is no surprise that the competition was won by a relatively unsophisticated
program.
The theory of behaviour-specific action suggests that no currently conceivable
computer could even transcribe English speech that required context sensitivity
to be understood: ‘It is knot hard too right English of this kine on a weird
processor or other quay bored. (An yew can run yore spell chequer over that
won.)’ Of course, once such a sentence has been written down, it is easy
to see how to design a program that would cope with it, but that program
will not cope with the next strangely written sentence. In the Turing test,
conversation is previously unseen.
The theory of behaviour-specific action, then, gives us a simple test
which highlights most of the difficulties of making machines that mimic
human action. This test, which looks hard at acts of writing, is a subset
of the Turing test, but a different subset from that utilised in the Boston
competition.