‘Can machines think?’

Greg Detre

Thursday, 30 November, 2000

Mind VII

 

Our curiosity has focused on the problem of Mind since ancient Greek thought. The confusion has been expressed in many forms, revolving around the relation between the mind and body, physical and non-physical, or subjective and objective. The question can be turned on its head – ‘Are humans (merely) machines?’ Seen in this light, the fields of physics, physiology and computer science are key. Indeed, technological progress has provided a succession of new paradigms through which to try to see our brain-minds, and new mechanisms with which to try and create artificial minds.

Certainly, our knowledge of the workings of the brain has deepened, both at the neuronal level and at the macroscopic systems level (benefiting from greater temporal and spatial resolution in neuroimaging). At the same time, our understanding of the physical laws of our universe has broadened, and become quite removed from anything we would characterise as common-sensical or intuitive, especially at the quantum level. Lastly, in computer science, AI and connectionist researchers have felt able to make strong claims about the feasibility of building a machine that can think – half a century later, and their credibility has suffered considerably. The ambitions of both fields have since narrowed, along with the complexity of the phenomena they seek to model.

One of the problems with talking about mind is that it is, by definition, subjective. We cannot access others’ mental lives, except indirectly through language. We attempt to categorise our inner world as words – broad, shifting categories, never sure whether the other person is describing the same feeling when he uses a word, only that he uses it in an appropriate way. The thought experiment of the inverted spectrum could apply here. As a result, we have no means of identifying with certainty whether another being has a mental life like our own. Yet, we assume other humans do, because they behave similarly, they’re built similarly, and they talk about a mental life in such a convincing fashion that it seems impossible to imagine that they are zombies. However, the problem of other minds becomes even more crippling when we look at animals, machines or rocks, who cannot communicate on our terms, and whose behaviour isn’t so anthromorphic that we automatically impose an assumption of consciousness upon them.

Answering the question of whether machines can think comes down to two issues: what we mean by ‘thought’, and whether we can be described as machines.

The term, ‘machine’ has often been used to describe something as contrasted with a being with free will, an automaton. I am going to try and avoid discussion about determinism in physical processes here, and simply define machines as being ‘constructed and containing no living tissue’. We will return to the implications of such a definition below.

 

As Bill Bulko puts it, ‘Artificial intelligence is the art of making computers that behave like the ones in movies’. But as we will see, this may be asking too much. Certainly, ‘thinking’ is intelligent mental activity; but that doesn’t really help us. In order to know how hard a hurdle we are setting for machines, we need to place ‘thinking’ relative to the different standards of mentality that we seem to recognise. This is especially difficult because we need to remember that non-human forms of mentality will be very different to our own – we cannot simply disentangle the ‘human’ from ‘human mind’ on the basis of just our own case. I am going to try and set out a series of assertions which characterise thought, in increasing order of the strength of their claims.

1.     Newell and Simon[1] defined intelligence as:

first, specifying a goal; then

assessing a current situation to see how it differs from the goal; and

applying a set of operations that reduce the difference.

Thus, the power of one’s thinking is a measure of:

the accuracy and reach of one’s evaluations of the distance from current state to goal;

the flexibility of behaviour produced in response;

the range of different situations in which the system can operate and learn to operate in.

We regard ourselves as the yardstick of complex thought because humans display extremely flexible, long-term behaviour in a huge range of different situations, and so we regard ourselves as thinking considerably better than our fragile, insect-level robots.

2.     Hume’s model of mind was separated into impressions (i.e. sense-data) and ideas, our memories of impressions. Simple ideas relate directly to an impression, while complex ideas are comprised of other ideas. This explanation seems quite adequate when describing our idea of an angel (the conjunction of two simple ideas: a figure of a man plus wings), or heaven (various complex ideas - pearls plus gates, gold plus palaces etc.). Thinking is seen as the categorising and associating mechanism by which we assemble knowledge from experiencing.

If we take this idea of association a little further, we have the foundations of connectionism. Simply put, connectionism attempts to produce schematised models of the brain at the neuronal level. The belief is that the brain itself can be viewed as a machine, whose organisational complexity is such that it gives rise to high-level behaviour as an emergent phenomena. Each neuron is itself a biological machine of considerable intricacy, yet because the sub-neuronal activity is considered to be of less importance than the connectivity between the neurons, or ‘nodes’, their activity is modelled relatively simply, often as little more than binary logic gates or perhaps a non-linear function with a threshold. When even a small group of neurons are connected to each other with pre-set or self-organised ‘synaptic weights’, they can approximate practically any function, as well as robustly recognising and generalising patterns (i.e. learning). Connnectionists hope that by retaining the criterion of biological plausibility, they can learn the tricks that evolution has painstakingly developed, and eventually leapfrog the brain, given sufficient processing power.

Minsky claims that a move from association to meaning is not as large a step as it seems[2]. He uses the example of numbers as a case where we think we understand numbers, where they appear to have meaning for us – yet, we’re unable to pin it down. Russell and Whitehead’s somehat counter-intuitive attempts at defining ‘five’ as the ‘set of all possible sets with five members’ ran into paradox and inconsistency (though they were able to eventually overcome the fundamental circularity). According to Minsky, this is largely because their rigid definitions of number run contrary to the truth that meaning is relative – it derives from relations with other things that are known. To try and define a meaning in isolation, rather than as a piece of a  jigsaw of knowledge, is doomed. The more links a concept has, the more meaning – hence the futility of searching for the ‘real’ meaning of something, as though there it was only connected to one other thing, then it would scarcely ‘mean’ at all:

Rich, multiply-connected networks provide enough different ways to use knowledge that when one way doesn't work, you can try to figure out why. When there are many meanings in a network, you can turn things around in your mind and look at them from different perspectives; when you get stuck, you can try another view. That's what we mean by thinking!

3.     We usually describe thoughts as ‘intentional’, or as having propositional content, i.e. thoughts are always about something – they are directed at an object. In functional terms, a representation can be seen as something we can manipulate, a mental object that we can perform operations on in our mind before trying them out in our physical environment. Connectionism appears to demonstrate how a representation of the external world might be held in distributed form in the synaptic connections between neurons. We can imagine that many animals, especially higher primates, hold rudimentary inner representations of the physical world – this would be essential in tool use, for example. However, social animals, animals who live and cooperate in groups, need to work even harder to model their environment because they need to incorporate other agents. Such social cognition, having beliefs about others’ mental states, requires larger brains and more complex representations. When we turn our capacity to represent other minds onto our own mind, we are having higher order thoughts (HOTs)- thoughts about thoughts. This reflexiveness is often regarded as defining ‘self-awareness’.

4.     At this point, we start to have difficulty deciding which of our faculties underpin each other. We know that language allows us to express these higher order thoughts, to make them explicit so that they are communicable. But we are unable to ask a baby or mute adult whether they are having HOTs, so we cannot be sure whether we can have HOTs without language. Similarly, it seems unlikely that one could produce human language fully without HOTs. (However, chimps[3] and parrots[4] have demonstrated considerable facility at recognising, remembering and employing linguistic symbols meaningfully, like ‘food’, ‘chase’, ‘green’ and ‘more’, as well as almost full comprehension of spoken English. They can identify and refer to people, other animals and their own mood (I think), and spontaneously produce novel constructions, but even the most successful studies have shown practically no proper use of syntax.)

For this reason and others, Alan Turing chose language as the vehicle for discriminating and demonstrating intelligence in his imitation-game, the Turing test. Indeed, passing the Turing test surely demonstrates considerable and very human-like intelligence. Unfortunately, machines are a very long way from this goal, despite the misleading results of restricted public experiments like Tomorrow’s World and the Loebner Prize[5]. Part of the problem, and indeed the attraction, of the Turing test is that it requires so much of a successful participant. They need a vast meaning-coded lexicon, a knowledge of syntax that probably far exceeds the best contemporary understanding of linguistics and facility with all of English’s exceptions, nuances and clichés. Moreover, since the domain of conversation could wander from film to shopping to girlfriends, the machine would need to have accumulated a knowledge base of the mundane, useful and quintessentially human that takes us many years. Much of this knowledge is species-specific – what sense could a computer make of hunger, love or pain, if its physiology is so different to ours that it need not feel them? After all, though we regard them as truths, they are truths of the human condition – we feel the way we do because of the particular (arbitrary) way in which we have evolved. If we could live on water, then hunger would be meaningless, and possibly inconceivable. Almost all of our common sense and shared language is predicated on our shared physiology, and for similar reasons that Wittgenstein denies that we could have a private language, it seems that a public language understood by non-humans would be very different from English, by virtue of having to be more objective and so less expressive. The Turing test then could be described as a sufficient, but not necessary condition for intelligence. Any machine that passes will have to be intelligent, but probably considerably more intelligent than us in order to pass a test couched so firmly in our terms – after all, no human could pass a similar Turing test conducted in machine code or whalesong, say.

5.     Probably the ultimate requirement of a thinking machine is true phenomenological consciousness. Thought always has a phenomenological aspect for us – there is always something it is like for the human thinker to have a thought. We assume that a truly intelligent machine would, of necessity, know what it feels like to be itself.

 

The reason for spending so long debating what we mean by ‘thinking’ is that the answer to our question hinges on our criterion of thought. If we take the strongest claim for thought, that machines must be fluent English speakers, human-seeming, and (claim to) be fully conscious, then truly intelligent machines are a long way off. But, I think most people would be prepared to accept a less strong claim as ‘thinking’. I would like to set the boundary around about the third or fourth level – higher order thoughts, with some sort of means of expressing them. Testing for such criteria will be difficult – either we will have to learn this machine language, or machines would have to learn ours – which, as we have said, may be an even harder requirement.

Before continuing, there are a multitude of poor reasons that we should dismiss for being quite sure that machines can’t think.

There is a (Christian) religious argument against thinking machines, which could run on many lines. In pre-Darwinian times, the Christian Church would have dismissed the possibility of thinking machines out of hand. Humans were regarded as divinely created and distinct from the animal ‘bête machines’, elevated by our souls, language and free will. The Church had to substantially revise its assertions after the theory of evolution became scientifically sacrosanct (although there are still parts of Bible belt America where the congregation are held in Genesis’ thrall). Man and machine would now be considered divided by the divinely-instated mechanism of evolution. It remains to be seen whether the Church could credibly accept a machine consciousness into heaven. Any variation on this theme which holds humans to be without doubt the only intelligent life in the universe is equally unreasonable or would have to rely on some quasi-religious unscientific premise or leap of faith – it is remotely possible that humans are the only intelligent life in the universe, but there is no scientific reason to suppose that this is necessarily so. There is nothing special about humanity – we may be at the evolutionary pinnacle of our planet, but our view over the rest of the cosmos is sorely limited. The second main religious argument against intelligent machines is that it places us in a somewhat god-like role – by intimating that scientists can understand the workings of the highest orders of life, and even create new and ‘better’ life, reduces the divine miracle of life. This is unnacceptable for the same reason that evolution was considered abhorrent – what is so god-like about God if we can forge souls ourselves? Those with a more literary turn of mind might choose to see parallels between AI and the Tree of Knowledge of Good and Evil in the garden of Eden…

There is a much wider category of people, branded carbon-chauvinists, who deem machines wholly incapable of thought because they are not biologically alive. By this, I mean that machines need not be carbon-based, or indeed contain any of the chemicals in our bodies and their physiology is likely to be unrecognisably different. Furthermore, machines are artificial – we will have designed and built them. They would not exist ‘naturally’, i.e. had there not been other intelligent life to first create them. They are evolutionarily separate from us and the entire animal and plant kingdoms. In fact, to emphasise the differences, they could be altogether virtual. The machines we could be talking about might be lines of code amongst other lines of code, A-life programs running in a virtual environment on a completely mindless supercomputer. Although they may behave and look like simple life forms, virtual machines are even harder for carbon chauvinists to accept than steel and silicon.

Without wanting to digress from the topic of intelligence in machines, it might be worthwhile considering briefly how strong the case for classing certain machines in existence today as ‘alive’. Steen Rasmussen’s contentious pamphlet, “Aspects of Information, Life, Reality and Physics”, distributed at the second A-life conference, contained the following argument:

1.     A universal computer is indeed universal and can emulate any process (Turing)

2.     The essence of life is a process (von Neumann)

3.     There exist criteria by which we are able to distinguish living from non-living things

Accepting (1), (2) and (3) implies the possibility of life in a computer

4.     If somebody manages to develop life in a computer environment, which satisfied (3), it follows from (2) that those life-forms are just as alive as you and I

5.     Such an artificial organism must perceive a reality R2, which for itself is just as real as our ‘real’ reality R1 is for us.

6.     From (5) we conclude that R1 and R2 have the same ontological status. Although R2 in a material way is imbedded in R1, R2 is independent of R1.

If sound, this argument would expose the fallacy of dividing life into the terrestrially evolved and not. Clearly, the argument hinges on the first two premises, since we can devise criteria of our own to satisfy the third premise – after all, it may be that the notion of dividing the world into ‘alive’ and ‘inanimate’ is itself a human, artificial distinction. The first premise is reasonable, if we allow the machine an infinite amount of time. Although this sounds like quite a large requirement, it is merely a nod to the difference between easy and hard computational problems, i.e. those where we can solve them in finite time, and those where we don’t know if we can solve them at all. If nature can compute a process, then a computer (especially a massively parallel one) can too.  The second premise is the most important one – if we bar God from the proceedings, then we have to look to scientific criteria for defining life. We can easily imagine a machine that could fulfil the following criteria:

consume from its environment (by consuming and probably excreting)

respond to stimuli

operate autonomously (i.e. without control from another being)

reproduce (and evolve)

grow

even if it did not:

excrete

respire

have DNA

‘Excretion’ is only necessary if the consumption requires some transduction of energy. I have deliberately avoided it, since a virtual machine would ‘consume’ processor time and possibly virtual food, but need not ‘excrete’ in any real sense of the word. DNA and respiration are clearly the hallmarks of creatures which have evolved from common ancestry to live on this planet, and there is no reason to exclude all non-DNA non-respiring extra-terrestrials from being alive. After all, it seems likely that life in a Jovian atmosphere or a remote and remarkable planet would have adapted to become quite different from what we are used to. The five main criteria are all processes. So we can accept Rasmussen’s second premise, and potentially accept machines into our hallowed hall of Life.

Computers are often described as ‘mindless’ and slavishly mechanical – this is true, insofar as they obey the laws of physics to the letter, and that we are able to take advantage of our knowledge of those laws to carry out specific processes. However, to use the deterministic nature of computers today as evidence that they can never think creatively, spontaneously or have free will seems rather strange. After all, the human brain appears to obey the same laws of physics that machines do. If we accept the materialistic premise, then we can see that there are some machines (i.e. brains) which do have free will. The question of whether machines can exercise free will in the same way becomes an empirical one of ‘Can we discover whatever new physical understanding of matter is necessary to realise how things work in the brain, and then incorporate this into machines?’ If, as Penrose and Swinburne argue, quantum mechanics plays a vital role in psychophysical interaction, then we may need quantum computers before our machines can think properly. But having said all this, it may be that we can have intelligence without consciousness. Intelligence does not require free will in the same way, though often intelligence has the appearance of spontaneity.

The last reason for stating with certainty that machines could not be intelligent is processing power. Comparisons are difficult because it is hard to estimate the processing power of a single neuron relative to a computer. However, by comparing the processing power of retinal ganglion cells (where the number of neurons is known) with similar calculations on a computer, we can get an estimate of the brain’s processing power. Current estimates[6] place supercomputers about a decade behind the brain in processing power, but potentially ahead in storage. Although Moore’s Law (that computers of a given price will double in speed about every 18 months) has held for over three decades, Intel will soon have to consider quantum effects – however, it seems certain that the processing power available to researchers will ultimately grow to exceed that of the brain. After all, ‘if the automobile had followed the same development cycle as the computer, a Rolls-Royce would today cost $100, get a million miles per gallon, and explode once a year, killing everyone inside.’[7]

 

When asked ‘Can machines think?’, one prominent computer scientist replied, ‘Can a submarine swim?’ This would seem to indicate that the question is ill-posed somehow, or that the answer is a clear ‘no’. Yet, Minsky argues convincingly that it is unreasonable to claim that there is any basic difference between the minds of men and possible machines.

Indeed, as Danny Hillis notes, because machines are artificial rather than evolved, they can be built in such a way to allow them to re-wire, re-program and re-build themselves, and could well be in an even better situation than we are and lead richer, happier mental lives. This raises the ethical quandary of conscious machines. A machine demanding rights would probably be another sufficient sign of intelligence, raising the same issues as animal rights, but appearing even more alien and chauvinist.

There is a historical trend away from mysticism towards materialism. The mind remains as one of the very greatest mysteries. Modern materialism appears to shred this shroud with its claim – if you accept its premises, then fully intelligent machines should be very possible. If you don’t accept materialism, then the question comes down to whether something non-conscious can be ‘intelligent’, as opposed to merely intelligent-seeming – a strange distinction.



[1] Newell, A. & Simon H. A., General Problem Solver, a program that simulates human thought, 1963

[2] Marvin Minsky, MIT, Why people think that computers can’t, 1982

[3] E. Sue Savage-Rumbaugh & Roger Lewin, Kanzi: The ape at the brink of the human mind, 1994

[4] Pepperberg on ‘Grey Parrot Intelligence’, 1993, 1995

[5] Stuart M Shieber, Lessons from a Restricted Turing Test

[6] Daniel Crevier, The tumultuous history of the search for Artificial Intelligence

[7] Robert X. Cringely, InfoWorld