The idea of artificial intelligence is not a new one. Thomas Hobbes, in his introduction to Leviathan (1651), wrote:
... why may we not say that all automata (engines that move themselves by springs and wheels as doth a watch) have an artificial life? For what is the heart, but a spring; and the nerves, but so many strings; and the joints, but so many wheels, giving motion to the whole body?
Hobbes’ point here is that it is possible to understand a human being in mechanical terms (he was a materialist) and that the analogy between human and machine may help us to understand our own nature. In the age of the digital computer, this claim has become more prevalent and has even led to a computational theory of mind. But some philosophers maintain that there is something unique about human intelligence, something that prevents machines from ever being fully conscious. It is therefore interesting to ask, what has to be true of a system to say that it matches our mental capacities? Or, to put it more simply, can a machine think?
Substance dualists will deny the possibility of artificial intelligence as they claim the mind is not a physical system. Identity theorists also tend to reject it, as they believe that mental states are identical to biological states in the brain – although the concept of multiple realisability allows for machine-based states. It is the theory of functionalism, therefore, that is the starting-point for any discussion of artificial intelligence.
Functionalism defines mental processes in terms of their functions, and their relation to other states. Put simply, the brain is a device that receives symbolic inputs through perception and sensation, compares those inputs to existing states, and sends outputs to the motor system. So an increase in temperature in a hot room will lead to sweating and removal of clothes; the same increase in a cold room will reduce shivering; and so on. In computational terms, the brain is the hardware and the mind is the software, and just as software can run on different machines and platforms, so minds can in principle be realised in both human brains and artificial systems.
This is sometimes referred to as a ‘black box’ view of the mind: we do not ask what is inside the box, merely measure its function through various inputs and outputs. By separating mind from any description of its physical basis, it becomes theoretically possible to reproduce mental experience on a machine. In the world of AI, this claim was optimistically made in the famous proposal for the Dartmouth College Research Project in 1956:
Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.
Note the use of the term ‘simulate’ here. A distinction is usually drawn between two types of artificial intelligence. Weak AI allows only that machines can simulate particular examples of intelligent behaviour accurately. Strong AI asserts that such activity is real intelligence and constitutes a mind. The claims of strong AI are now usually referred to as AGI – artificial general intelligence – to make clear the larger scope of the project. While the position of weak AI is relatively uncontroversial, interesting philosophical questions arise in consideration of the claims of strong AI – indeed, several philosophers reject its claims outright.
The agenda for discussion of strong AI was set in 1950 by Alan Turing in his seminal essay Computing Machinery and Intelligence. Turing posed the question, can machines think? His approach was to redefine the question in terms of what he called ‘the imitation game’. Imagine having a texting conversation with two interlocutors, a human and a computer program. To pass Turing’s test, the computer must be able to fool you into thinking that it is the human. Turing himself expected computers to be able to achieve this by the year 2000:
I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.
In a certain sense, Turing’s prediction has come true [1]. Traditional AI research has tackled particular problems of intelligence with the aim of resolving them through abstract computation – and has often succeeded. In 1997 IBM’s chess computer Deep Blue famously beat the world champion Garry Kasparov under standard tournament rules. There is also a great deal of low-level evidence that machines display intelligence: we have missiles that can seek out heat sources, and thermostats that can adjust themselves to ambient temperatures; we talk of ‘smart’ phones and other devices, and often refer to our computers as if they were active, intelligent beings; we have programs that can learn from experience and partially adapt themselves; and so on. In computational tasks such as calculation or theorem proving, machines have outstripped us. However, in other areas the development of AI has been disappointing. Natural language processing has made slow progress, and it has proven difficult to build a robot with apparently simple capabilities like object recognition, spatial awareness, and so on. This perhaps suggests that the strong AI project will not be realised any time soon, if at all.
Even if a computer program did pass the Turing Test, we can still enquire whether the machine was really ‘thinking’ or merely simulating intelligent behaviour. Turing himself rejected this distinction, and went on to identify and address several common objections in his essay.
He refers to the dualist response as The Theological Objection: since mind is not a property of matter but a separate substance, it cannot be possessed by a machine. In other words, a machine does not have a soul. Turing dismisses this objection, wondering just why God would not see fit to endow a machine with a soul. The Argument from Consciousness is the claim that machines can have no conscious awareness of their processes: unlike humans, they have no feelings or emotions to affect them. There may be no evidence of such consciousness, but Turing argues that we have no more reason to attribute consciousness to other people. Despite this, based on the sophistication of their behaviour and responses, we do infer that other people have minds – and a program which passed the Turing Test would earn the same inference. Anyway, perhaps machines are simply dispassionate thinkers, and thought does not require such feelings to operate (although some philosophers insist that human minds are exactly this combination of reason and feeling). Arguments from Various Disabilities point to things that humans do that computers cannot. Certainly, not all mental activity feels like computation: how does one program creativity, poetry, imagination, insight, humour, love and so on? Turing suggests that such objections stem (ironically) from a failure of the imagination. We cannot yet conceive of such abilities in a machine only because we have never observed them before. He also argues that intentionality, a serious problem for proponents of strong AI, is possible in a machine. When a computer is resolving an equation, it can be said to be ‘thinking about’ that equation. If this is correct, then in principle a computer could ‘think about’ itself, thus achieving something like self-awareness. Lady Lovelace’s Objection [2] is that machines can only do what we program them to do and cannot therefore innovate like humans. We are autonomous, while computers are automata. The distinction here seems to revolve around freedom of the will, but problems understanding human freedom – and indeed the very processes by which we come to innovate – make the distinction unclear. The Argument from Informality of Behaviour draws attention to our capacity to improvise: it is not possible to create a system of rules to deal with every eventuality, so we use our experience and judgement to guide our action, as much as rules of conduct (think of negotiating a business deal, or cooking a complex dinner). Turing acknowledges the truth of this, but insists that it does not follow that we are not machines. The laws of behaviour that guide us could be programmed into a sufficiently large computer.
The most famous objection to functionalism and the project of strong AI has come from John Searle and his Chinese Room argument. Searle asks us to imagine a person in a locked room, receiving posted instructions in Chinese and posting appropriate responses back. Crucially, this person only speaks English and therefore must refer to a set of instructions for translation. A computer which did this successfully could pass the Turing Test, because it might convince a Chinese interlocutor that it is human. But, argues Searle, it is doing no more than the man in the locked room: he does not understand Chinese, but merely simulates understanding. There is a vital distinction between semantics and syntax. A computer using language in this purely syntactic manner lacks the characteristic understanding displayed by human intelligence.
Many philosophers feel that there is something wrong with Searle’s analogy. In general, successful language use only occurs when syntax and semantics coincide. The Chinese Room thought experiment separates these two elements artificially, and therefore the analogy is flawed. Others have suggested that while the man in the locked room cannot be said to understand Chinese, the system as a whole (which includes the instructions for translation and the mental skill needed to match the symbols) does indeed display understanding. Some philosophers have noted just how difficult it would be for a person in the locked room to convince their interlocutor: it would require thousands of years and astronomically many instructions, suggesting that a translation machine would be far more complex than Searle’s analogy allows.
Searle’s argument is a reaction to the functionalist account of mind which omits the subjective experience of mental states – the ‘qualia’ that characterise our consciousness. Imagine two subjects who experience colours differently: when A sees a red object, A has the same experience that B has when B sees a green object (and vice versa). Both call red objects ‘red’ and green ‘green’, even though their internal experiences differ. If asked to sort out a pile of red and green buttons, each would sort by colour the same way – given the same input, each would produce the same output. According to functionalism, this means they have the same mental state. But although their behaviour is identical, their experiences are subjectively different, and this internal difference is essential to a proper characterisation of mind. A Turing machine lacks this subjective quality and cannot therefore be described as properly intelligent.
Ultimately, Searle believes that human consciousness can only exist as a property of biological human brains (a version of Identity Theory). It is perhaps plausible that not all physical substances can generate consciousness, just as not all can conduct electricity. A similar position is held by Hubert Dreyfus who insists that our intelligence is intimately tied up with a range of other faculties: our senses, our emotions, our language, our subjective representation of the world. Without similar apparatus, similar minds will not exist: artificial minds will not be like our minds.
Dreyfus rejects a notion at the heart of AI – that all knowledge can be represented by a system of symbols and rules. Some higher-level behaviour depends instead upon intuition, unconscious instincts which cannot be formalised as computational rules. Consider driving a car: being a good driver is more than simply following rules. The ability to drive well and safely is open-ended, and good drivers are able to deal with unimagined and unpredictable situations because of an unconscious background of commonsense knowledge. Taking his cue from Heidegger, Dreyfus draws a distinction between ‘knowing-that’ and ‘knowing-how’, and argues that only our immersion in the world produces the latter. When we recognise a face, for instance, we do not consciously use symbolic reasoning the way a computer program might match photographs. This background of commonsense knowledge is impossible to represent formally and therefore impossible to program. Strong AI is simply not possible.
There is a curious twist in the history of AI. In 1966 Joseph Weizenbaum published an early chatbot program called ELIZA, which used open questions (‘Can you elaborate on that?’) and repetition of key words (‘I’m depressed’—‘I’m sorry to hear you are depressed’) to engage its interlocutors. Weizenbaum was so appalled by the degree to which people were fooled by, and emotionally involved with, ELIZA that he later became one of AI’s fiercest critics. He came to believe that computers should not play important roles in therapy, medicine, the law or even customer service, because they lack qualities like compassion and the capacity for wise judgement. Artificial intelligence used in this way, Weizenbaum claimed, would represent a threat to human dignity.