The growth in the study of Artificial Intelligence – computers performing human-like tasks such as playing chess and making medical diagnoses – over the last few decades has led people to ask if these computers are actually, or could ever be, conscious. Consciousness is a difficult concept to define. We know what consciousness feels like from the inside, but how are we to make any judgement about its existence in computers? Even a definition is difficult, but probably there is reasonable agreement that consciousness involves awareness, the precise nature of which is a matter for debate (and some of the forms of which will be discussed in this article).
Suppose we recorded an event on a video camera. Can we then say that the video camera is aware of what is going on? Most would answer “No”. The camera may register the image, but it has no understanding of what the image means. The notion of understanding is thus very important to the story. Over twenty years ago, Terry Winograd, using a computer considerably less powerful than the one I am typing this upon, put together SHRDLU, a computer program that could read in English words, interpret them, and act upon its “understanding” of the sentences by manipulating a world of coloured blocks, using a robot arm. Although this system only worked for a limited world – that of the coloured blocks – it was capable of responding accurately to complex instructions. SHRDLU certainly gave the outward appearance of understanding, but is this enough? To some, yes, but John Searle would argue differently.
In 1980, Searle put forward a thought experiment, known as The Chinese Room, and for a long time this, and later refinements, have been a major cause for debate. Searle does not dispute that programs such as SHRDLU can simulate understanding (what he terms “Weak Artificial Intelligence”, or “Weak AI”), but he does dispute “Strong AI”, which essentially claims that the computer can actually understand. His objection is not just based on lack of complexity; he argues that no digital computer program will ever be able to produce something that can actually understand.
The Chinese Room experiment asks you to imagine that you can speak only English, and are locked in a room with a whole set of Chinese symbols, and a set of rules written in English. This set of rules, which Searle likens to a program, instruct you to pass certain symbols out of the room whenever specific symbols are passed in. The argument is that provided you can follow the instructions accurately, to someone outside the room, you appear to understand Chinese, but in reality, you have no understanding at all – you are merely following the “program”. This basically parodies the famous Turing test for understanding – whereby understanding is judged by a human conversing with a computer and another human. If the first human cannot tell the difference, then we say the computer really does understand.
So how does the Chinese Room set up differ from us, as humans, understanding a foreign language? When I, as a non-native, try to speak German (my knowledge of Chinese is not sufficient!), I follow certain rules, mapping between German words and English ones, adding endings to words under specific circumstances, and moving the verb to the end of the sentence under the appropriate condition. While a trifle more complex than the Chinese Room scenario, this too is very much following a set of rules, and could be described as a program. It might be argued that the German only becomes more than just meaningless symbols because it is mapped to English, where it has meaning.
So what precisely is meaning? Marvin Minsky, an AI pioneer, argues that meaning is a whole load of interconnected pieces of information, stored in what he calls ‘webs of meaning’. For example, the concept “five” is not just stored as “one more than four”, but can crop up in all sorts of senses – five aardvarks, five pencils, five points on a star. The brain has the capability to pick out the presence of ‘five’, even if the situation is unfamiliar. This proves very useful – the brain isn’t just limited to one narrow definition or interpretation, but can call on a whole host of similar situations from the past, and then try and apply that knowledge to the new situation.
So, if we were to encode such a network of information within a computer, would it then be able to understand? Those in the doubters’ camp would refer back again to the Chinese Room scenario – and argue that the system is still only blindly following rules (or a program) to process the symbols, which are completely meaningless to it.
So how does the brain achieve understanding and consciousness, and give the appearance of a mind? Assuming that they are not beamed in from outside, then they must either be part of the brain or, to put it another way, products created by the working of the brain – just as all the component pieces of metal etc. come together to make up the “car-ness” of a car. Some might doubt this, but in doing so inevitably run into difficulty specifying how the non-physical mind interacts with the physical brain.
The human brain contains a particular type of nerve cell, called a neuron. These have a spidery shape that allows them to connect, and thus communicate with one another. The chemical make up of these neurons means that they can generate and convey tiny pulses of electricity along the membranes that link them. New links can form and strengthen. There are an estimated 100 billion neurons in a typical brain, and each of these may link to thousands of other neurons, or in certain cases to parts of the body such as muscles. Leaving aside the possibility of anything being “beamed in from outside”, or some as yet undiscovered process going on (Penrose hypothesises the existence of quantum effects), then all our thoughts, feelings etc. are the culmination of the activity of the neurons and other bodily cells. All our complex thoughts, logical deductions etc. are based upon tiny little cells, which receive electrical signals from cells they are connected to, and respond by sending electrical signals to other cells. This seems like a very mechanical process – the neurons are acting purely according to the rules of chemistry – they are not processing the electrical signals with any intelligence, they are just following the program dictated by the laws of chemistry – so why does this process classify as understanding and intelligence – doesn’t the Chinese Room argument apply here also?
Ultimately, the Chinese Room is a thought experiment, and what you think of it is not a purely logical conclusion – it will be based around your intuition as to what constitutes understanding. Even if you agree with the conclusion that the original Chinese Room does not exhibit true understanding, does that necessarily mean that any purely algorithmic process cannot exhibit understanding? If it could, then a complex enough system could understand many things. It would thus have awareness, and hence consciousness.
Even if you doubt that true understanding is displayed by more traditional computer approaches, can we achieve it by getting closer to the mechanism in the brain? Neural networks are an attempt to do just this, by mimicking the processes going on it the brain. Unlike a typical computer, which has one CPU (Central Processing Unit) – the part of the computer that follows the actions specified by the program and manipulates data – neural networks effectively have more. They are conceptually made up of a whole array of processing units, working in parallel, each of which is connected to many others, just like the neurons that make up the brain. Each receives signals from other processors linked to it, and responds by sending electrical signals to various processors it is linked to. The precise details of how it decides which signals to pass on and to whom is decided by a program running on each of the processors, and can be adjusted by changing various weightings within the instructions. The process is exactly similar to that taking place in neurons, in that the inputs are mapped to the outputs. In practice they can be simulated on ‘conventional’ computers, but over recent years special computers have been built (e.g. Hillis’s Connection Machine) which are ideally suited for use for a neural network.
Neural networks work rather like brains, in that they first need to be trained, so that they can adjust the weightings used in the processors to produce the desired result. Once trained, they are particularly good at just the sort of thing that machines such as those described earlier are bad at, but we as humans do as second nature. More traditional computers are particularly bad at recognising characters. They can be programmed to check if a character is a letter ‘a’ in a specific font, but if the font is changed, the letter is rotated, the letter is not printed well, or if the character is hand written, the program will not succeed. The neural network (e.g. WISARD) will. While we are not aware of neurons firing when we think, the process taking place in the brain is close to that going on in a neural net. It also ties in with Minsky’s notion that understanding is based around a network of interlinked concepts – a web of meaning.
The position that everything comes from the interaction of neurons is supported by those such as Churchland, but opposed by people such as Thomas Nagel, Searle, and Frank Jackson. Johnson puts forward the thought experiment of the Black and White Room. Here, a scientist called Mary has lived in a totally black and white room all her life, and has had no contact with the outside world other than via a black and white television. She studies black and white books and gradually comes to know all the physical facts pertinent to everyday colours and colour vision. Suddenly, she ventures outside the room, and sees a red rose. Jackson proposes that no amount of understanding about the physical properties of colour can prepare Mary for the ‘raw feeling’, or qualia, that she experiences when she actually sees the colour red for the first time. This is intended to support the claim that there are processes other than physical ones that cause what we experience. Others would argue that whatever the thing is we would experience, it is still coming – somehow – from the neurons. While traditional computers would not get this experience – recognising the wavelength of light that is red is not the same as experiencing it, theoretically there seems no reason why a design like neural networks could not be connected in such a way as to cause certain situations to have an emotional type effect – by adjusting the weightings in the processors we can adjust the ‘thinking’. This seems entirely natural – when we get frightened, our bodies produce certain chemicals, which make us think in ‘fight or flight’ mode. Our whole thinking alters when we are in such a state, and we might well respond differently to a given situation because of this. Although we are aware of other “higher level” thoughts which we may use to justify it, underneath it seems likely that it is just the neurons responding differently to the inputs (because of different chemicals being present) and hence giving different outputs. A number of researchers are currently working on such systems – building “emotion” into computers.
Searle argues against the neural network approach by adapting his Chinese Room to be a Chinese gym, with a team of people passing the symbols around, each knowing their own little bit of the rules, but not having understanding. Therefore, he argues, the whole group cannot understand. Two responses seem possible – firstly, the whole notion of a gym of people passing stuff around is so bizarre as to provoke an intuitive rejection of the possibility that understanding is occurring, and secondly, how different is this to what is going on with neurons? Could not the same criticisms be levelled at the brain? In other words, provided the computer system works in much the same way as the brain, whatever conclusion you reach about the understanding and consciousness of one must also be applied to the other.
While we are still at the very basic stages at present, in theory there seems no particular reason why as our understanding of the brain grows, we could not build a hugely complicated interconnected web of processors to mimic precisely the behaviour of the brain. Even if current neural network understanding is not complete, the neurons behaviour – receiving input, and producing certain output dependent upon it, is very close to what goes on in a computer. The low level architecture would be similar, and so if put together right, the outward effects should be too. If we attribute to a “machine made of meat” the qualities of understanding and consciousness, why not one made of silicon.