Introduction

The above speech is by Joe, a fictional robot in the film AI: Artificial Intelligence. As it happens, in the film, Joe seems to be wrong and one robot, David, does appear to believe in the supernatural. Is Joe's opinion typical of what we should expect from an artificially intelligent machine? If machine intelligence matches human intelligence, or exceeds it, what can we expect of computers with regard to supernatural belief?

During the 21^st century we can expect further progress to be made in artificial intelligence (AI). Although much artificial intelligence work involves the solution of specific problems, the ultimate goal of research, the creation of intelligent, self-aware machines, has captured the imagination of science fiction writers and people in general.

The idea of a machine able to duplicate, and possibly exceed, our ability to perform complex tasks in the world, our ability to learn from our experiences and maybe even our ability to produce works of art or philosophy may be a little unnerving, but many people would imagine that religious belief would forever be the preserve of humans.

In this article I shall explore the idea of religious belief in intelligent machines and suggest that it may be possible for an artificially intelligent computer to believe in a god without us suggesting the idea to it.

The scope of this article

Religion is a controversial subject. Some people believe that at least some religious claims are true, whereas some do not. This means that this article will be entering a controversial area. Before we go further I want to explain some things:

• So that I do not appear to be evading this issue, I would describe myself as an atheist.
• I am not suggesting that people need to be atheists to do AI work. This article just discusses a way in which religious belief might emerge in thinking machines.
• I do not prove that a god does not exist in this article; however, it will be fairly obvious that this article does imply that god belief could emerge in humans in the same way that I suggest it could emerge in computers and that god belief may be a logical mistake. I cannot do anything about this if I am to propose this idea.
• This is not a claim that everyone has to think as I do.

What is religious belief?

If I am going to suggest that computers may believe in a god I should at least define the word 'god'. For the purposes of this article I would say:

'A god is an intelligent agent that is the cause of all things other than itself. A god has no cause of its origins (this would introduce things that are not caused by a god) and there is no correct way of simplifying a description of reality that contains a god to produce one which avoids having to implicitly assume the existence of a god.'

What do I mean by 'implicitly assume'? Many things could be claimed to have brought us about; for example, some people believe that aliens caused us to be here, but these aliens would not be a god, according to this definition, if they were presumed to have a cause, such as Darwinian evolution on an alien world. For an entity to be a god, according to this definition, there is no cause behind it or simpler model underpinning it; it must simply be presumed, without any explanation, as a fundamental thing.

Some readers may disagree with my definition of the word 'god', but I am not trying to do theology here; this is a very specific definition of god to be used within this article and is intended more to describe the sort of thing in which I am suggesting computers could believe, rather than the sort of thing that many theologians say a god is.

What an intelligent machine would need to do

We first need to look at the sorts of things that a sophisticated, intelligent machine would need to be doing.

One capability that it would seem to need is planning and we see a crude example of this in computer programs that play games such as chess. Chess algorithms give us a useful insight into what planning involves. To plan actions we need a model that can be used to simulate reality and indicate the desirability of an action, by exploring the consequences of the possible sequences of actions that follow after it, before it is committed. A machine cannot investigate every possible sequence of actions that could be performed in the future (there are simply too many), but there are various ingenious methods of planning that avoid the need to do this. Even so, planning is still an extremely expensive task in computing terms.

When IBM's chess computer Deep Blue beat Gary Kasparov few people would have been persuaded that Deep Blue, by virtue of this feat, had anything approaching human intelligence. Deep Blue could only deal with chess and the model for this was provided to it by its programmers. Humans, on the other hand, are adept at observing the world and making models of it by themselves. Modelling is done by all of us to make sense of the world, but it is a highly formalised process in science. Because of the importance of modelling in science, my school physics teacher once asked his class to solve a particular modelling problem which I think helps to illustrate the whole idea of modelling, so I shall share it with you now. Here is the problem:

You are given a box. The box has output terminals, or indicator lights, which allow you to determine the voltage that is on them. The box has input terminals, to which you can apply a voltage. You are told that there is an electronic circuit in the box. Your task is to find out what circuit it is without opening the box.

I like this problem because it seems to go straight to the centre of the issue of what our 'world view' of the universe is and how it is arrived at; it is like a microcosm of all of human science and all of our attempts to obtain knowledge about the world. We can treat the box as if it is all of our external reality and our attempts to determine what is in the box as our attempts to obtain a world view of our universe. In this analogy, making inputs into the box, and observing outputs, is equivalent to making experiments in reality.

It may seem that I have embarked on a self-indulgent philosophy exercise here, but making world views (or models) is one of only two things, the other thing being planning, that intelligence does. What a world view is, and how we arrive at one, is really important.

When you study the box you can never be sure what is in it; for example, you may think you know exactly what electrical circuit is in the box, but unless you have been told, how can you be sure that it is an electrical circuit at all and not something else with equivalent behaviour? This is the sort of problem that has concerned philosophers and we need not worry unduly about it; all we need is a model that allows our system to behave intelligently. We also have the problem that for any set of experiments that we can perform on the box there will be an infinite number of models that would fit the results of those experiments. How do we decide which model to use?

Of clear importance here is a philosophical tool known as Occam's razor [1], proposed by William of Occam (or Ockham), an English philosopher of the thirteenth and fourteenth centuries. This is the idea that when presented with two models, both of which agree with existing experimental data equally well, we should select the simpler one.

What does Occam's razor mean in the context of an artificially intelligent system? A model can be viewed as an algorithm that accepts descriptions of 'experiments', or sequences of experiments, as its input and predicts what the results should be when they are actually performed in the real world. If a model agrees with existing experimental data this means that, when the details of experiments that have already been conducted are input into the model, the results output from the model are the same as the known results. To say that one algorithm is simpler than another merely means that it is of smaller size. Philosophically, this is all very convenient, as the problem of whether or not the system has attained the 'real' model is neatly sidestepped; we just obtain the simplest model. An artificially intelligent system using Occam's razor as the criterion for its model generation would continuously generate a new model to fit its experiences to date, always attempting to obtain the simplest possible model, or would do something similar to this.

This gives an interesting perspective on humanity's accomplishments in terms of constructing theories about the universe. All of our scientific theories, put together, are nothing more or less than a model of the world in which we live. We can do science because it involves looking at reality and selecting an appropriate model, a way of thinking that seems to be hard-wired into us. In this respect we might all be considered amateur scientists every time we increase the sophistication in our view of the world!

In practice, things are not quite as simple as this; for example, most of us do not merely assume that the simplest model of our world has to be correct; indeed, to use this as a basis for planning our actions would probably prove fatal in short order. The reason for this is that there is no guarantee that every detail of the simplest model is correct. One of the 'runner-up' models will frequently turn out to be correct and, less frequently, one of the more complex models will be correct. The dogmatic assumption that the simplest model is correct may, for example, cause us not to take action to avoid real danger, simply because the dangers were evident in some of the runner-up models, but not in the simplest model. Actually, making this sort of generalised model making machine would be a difficult task and this issue of the uncertainty in the truth of the simplest model is just one problem that would need to be dealt with, another problem being the need to make abstractions in the model, but Occam's razor is a useful simplification when describing how an artificially intelligent system would work.

How planning and modelling would work together

We now have concepts for two ways of working:

• creating models of an environment
• planning to determine the optimum action, given input of a continuously updated model of an environment

An intelligent machine would work by using both of these two approaches or by doing processing which is equivalent to this. A modelling system would receive data from the environment, as well as information about what the system is doing in it, and use this experience to construct a model and continuously refine it. This model would be provided to a planning system that would continuously determine the best action to take.

Problems

Let us suppose that we make breakthroughs in planning and modelling so that we can construct a planning system that can accept almost any model and decide what the current best course of action is, based on simulations using that model, and we can make a modelling system that can accept experiences and make a model from them. We should now be able to make a 'true' AI.

Our system would have a problem: the ability of its planning system to give useful results would depend on having a good model of reality, but some aspects of reality would be extremely complex and it would take a long time to analyse these to produce a model that could be used for predictions.

What sorts of things might our computer encounter that are really complicated? The answer is simple: other intelligent entities, whether they are artificially intelligent machines like itself or humans. When it is confronted by these it is going to find the problem of making a predictive model to be so complex that the task, for all practical purposes, is impossible to perform. A model that can describe intelligent entities has eluded AI research for a number of decades now, which should be adequate evidence of the complexity of the problem, yet a computer would have to solve it, outperforming AI specialists like Marvin Minsky in the process, to make a predictive model for any intelligent entities that it meets!

Consequences

Because of this huge problem of the complexity of other intelligent entities, our intelligent computer has a difficult time. This does not mean that it is stupid; in fact, its modelling abilities may be good enough to allow it to look at the sky and work out Newton's law of gravity all by itself. Neither is it a creativity issue; our computer may have a lot of that. It is simply that intelligent things are too complex to understand easily, even for a very smart machine that is good at dealing with complicated things.

The computer may even manage to acquire a partial understanding of intelligence, but it will not be a full theory and will be rather crude. As its theories of intelligence become more complicated they will become progressively harder to use, in the sense of being more demanding of computational resources, to make useful predictions about what intelligent things are likely to do. The computer will live in a world full of chaotic things that take all its skills of analysis even to make any small progress in understanding.

As the computer lacks any deep understanding of how other intelligent entities work, and is unable to model what they are doing with any competence, it will not be able to incorporate them adequately into any of its planning. This would have serious consequences.

Here is an example of how one consequence could manifest itself:

Let us suppose we put our intelligent system in charge of a robot and require it to survive in some environment, perhaps by roaming around acquiring energy. Let us imagine that our robot has interchangeable legs that can be removed and reattached and that it meets another robot like itself that has a damaged leg and is in serious trouble. Our robot may be in serious danger from this damaged machine, which may decide that its best chance for survival is to attack it and steal a new leg from it! However, our robot may not do anything to avoid this until it is too late because its modelling of the environment may not describe another intelligence with enough detail to suggest such a hostile possibility. As far as our robot is concerned the other machine is simply a system that is so complex that it may appear to behave almost randomly. This is one consequence of the complexity of other intelligences: threats posed by other intelligent entities may not be predicted by the system's model.

Here is a second example of a possible consequence:

Communication with other intelligences would be problematic. The whole purpose of communication is for one intelligent entity to give information about its current mental state to another one. As our computer does not have a decent way of modelling what another intelligent entity is doing then communication with it is going to be almost meaningless. This barrier would severely compromise any communication between humans and the computer, making it very difficult to teach it anything. This would ultimately limit the amount of progress that the system can make with regard to increasing its abilities.

Humans do not seem to have these problems, so we now need to consider how people avoid such a situation.

How do humans do it?

How would it be if humans had the same problem of not being able to understand what other people are or what they are doing? We would all appear to be appallingly complex things to each other and would have serious problems even having meaningful communication. Even simple questions, such as, 'Would you like to go for a beer?' might be almost meaningless without some understanding of what is going on inside the entity asking the question. The only way that we could make any progress is by spending vast amounts of time analysing each other, and even then such progress would be limited. There would be little opportunity for any understanding between humans and no prospect of creating any sophisticated sort of society. We may even find it difficult to learn about the world and how to think, as a lot of the learning that children do is by obtaining information from other people; if there is little or no communication with those other people, what prospect is there for learning?

The fact that humans do seem to manage to understand each other suggests that Darwinian evolution has somehow equipped them with a way of understanding each other. There are two main ideas about how this is done. One of them is known as theory of mind and the other is often known as mental simulation.

A possible way out: theory of mind

Theory of mind is the idea that humans have been equipped by evolution with a specific ability to understand other people and have concepts of such things as 'beliefs' or 'intentions'. The idea is that, rather than being left to work all this out by analysis, we have a capability in the brain that has evolved especially to do it so that the high complexity of intelligent things that we may meet is dealt with.

There is evidence from some experiments that a theory of mind, if it exists in human brains, is acquired over a period of time. Some of these experiments were done by two psychologists, Wimmer and Perner [2]. As an example of what Wimmer and Perner did, one of these experiments involved the belief task, which is as follows:

In the belief task a child is shown a scene in which a person, whom we will call Jane, puts an object, let us say a ball, in a certain place; let us say, for example, that she puts it under a blue cup, and then goes away. While Jane is away, another person, whom we will call John, moves the ball to another place; let us say, for example, that he places it under a red cup. The child is then shown a scene of Jane returning into the room and is asked, 'Jane comes back. Where does Jane look for the ball?'

An adult would easily see that Jane would look under the blue cup because that is where it was when she left the room, but when Wimmer and Perner did their experiment they found that, while older children had no trouble with this, children under three or four years of age tended to answer that Jane would look under the red cup. The child gives this answer because he or she knows that the ball is under the red cup and cannot grasp the idea that Jane has had different experiences and would believe that the ball is under the blue cup.

Younger children seem unable to deal with the idea that other people can have beliefs different to their own, but older children handle this idea with ease. Experiments like this suggested that children become more competent at dealing with the minds of other people as they get older and this may suggest that if the theory of mind idea is true then some of it is learnt.

I must admit that I am rather sceptical of the theory of mind idea as a complete explanation because it seems to have trouble explaining how we actually deal with the very high complexity of other people. Even if we do have a system that does it, how does it work? A theory of mind would have to be learned by experience, in which case we would still have to deal with the enormous problem of the complexity of others to work out our theory of mind in the first place, or it would have to be provided, essentially by our DNA, and built into our brains, ready to be used when we started meaningful interaction with other people.

The idea of a theory of mind being built into us also causes me problems because it would seem that the theory of mind would involve concepts that we just did not have when it was built into our brains, so there would be nothing to which we could relate it until we had learned more about what people are.

Another reason that I am sceptical about theory of mind is that it seems a less economical theory than mental simulation, the other idea that we will now look at.

And another way out: mental simulation

Mental simulation is an idea that was suggested by Robert Gordon in the 1980s [3] and it is an alternative hypothesis to the theory of mind idea.

The rather strange idea of the mental simulation theory is that we do not rely on any deep understanding of other people's minds at all because we do not need it! Instead, each of us uses his/her own mind to try to simulate what somebody else's mind is doing. We do not need to know how other people work; it is enough to know that they have minds which are, to some degree, like our own, so (in a way) we can pretend that we are someone else to find out how they are likely to react in a given situation. In this way, we can make sense of the behaviour of other people by viewing their situations as if we were in them. We do not need to understand how their minds work, nor does each of us need to know how his/her own mind works to use it to simulate theirs.

From a computing point of view, this is analogous to making a complex computer program, containing many millions of lines of computer code, and equipping that computer with the ability to use some of that code to simulate the behaviour of another computer that we know to be running basically the same sort of program and predict its behaviour.

As I stated earlier, I find this mental simulation idea to be more persuasive than the theory of mind idea, mainly because it is more economical. It is based on adaptation of systems that are already in the brain, namely the systems which run our mind, to be able to simulate the minds of others. It therefore seems a simpler theory than one that requires us to accept the existence of a sophisticated system to deal specifically with intelligence.

There is some controversy about whether the better theory to explain human social skills is theory of mind or mental simulation. Some psychologists think that mental simulation alone is inadequate. While my own view is inclined more towards mental simulation than theory of mind, a good case can be made that mental simulation is not the whole picture. Some psychologists think that another idea, such as theory of mind, could also be of relevance. This may well be the case, but whether or not mental simulation alone would suffice in a computer is not too important in this article. What is important is the idea that a separate 'social' system exists that works separately from analytical modelling.

Autism

Previously, I commented on how strange the world would be to us if we lacked any special system for dealing with the complexity of other people. This might seem a rather unlikely scenario, but there may actually be people who have to cope with just such a situation.

There is a condition known as autism [4, 5] that starts in early childhood and makes people seem as if they are 'shut-off' from the world. Autism can be more serious in some cases than others and tends to have a general effect on mental abilities, but specific symptoms tend to include unusual behaviour or difficulties in social interaction with others and in communication. Autistic children can also have difficulties in pretending.

In 1985 S. Baron-Cohen, U. Frith and A. Leslie made the suggestion that these specific symptoms could be explained if autism is not a general malfunction of the brain, but a failure of its ability to deal with theory of mind, that is to say, a failure of the system that it uses to work out what other people are thinking [6]. These scientists tried a number of experiments, including the false belief task that we discussed above, with both autistic children and children who had Down's syndrome, another condition that affects mental abilities. Young autistic children did better than the children with Down's syndrome in many tasks, but they failed to perform the false belief task correctly, and older autistic children did not seem to do much better.

This led Baron-Cohen, Frith and Leslie to suggest that the main aspect of autism is a failure of the brain's theory of mind ability, so that it cannot deal with the beliefs and opinions of others, or does so more poorly than the brain of someone who does have a fully working theory of mind. For the case that I am making here, I am working on the assumption that the mental simulation idea, rather than theory of mind, is closest to the truth, but the false belief experiment, with autistic children, still suggests that the main aspect of autism is a failure of whatever system we have to deal with other people.

Earlier in this article I suggested that intelligent machines constructed without some way of dealing with the huge complexity involved in dealing with other intelligent entities would face problems. When AI research starts to produce systems comparable to the human mind, systems like this, without adequate methods of social modelling, will probably be made and we are now in a position to give them an adjective: they will be autistic. [7, 8]

Self-referential definitions

Another way of viewing the approach of mental simulation to deal with other people's minds is to think of it as being a different way of defining things.

For the brain to understand something it needs to have a definition of it or something equivalent to a definition. Definitions usually involve a description of how various concepts are combined to make the thing being defined and these concepts can in turn be defined in this way. This can be continued until only very simple concepts exist that cannot be broken down into simpler concepts. In this way we can view definitions as being reductive; ideas are reduced to a number of simpler ideas which are reduced still further and so on.

The problem with other people's minds is that they are just too complex to easily define in this way, so we use a different approach, which I shall call self-referential definition. A self-referential definition is a definition that, rather than containing ideas that can be reduced to simpler ideas, instead describes what it is defining in terms of the entity actually making or containing the definition;

Dictionary (noun): a book like this one

This definition may seem somewhat vague because there might be many things that have something in common with a dictionary, but are not dictionaries, but this is just a simple example; better self-referential definitions can be made.

In the same way that a dictionary can store a self-referential definition of a dictionary, a mind can store a definition of 'mind' that refers to itself, so that it can deal with the existence of other minds without having a detailed idea about what a mind actually is.

A judge, Justice Potter Stewart, once made the famous statement, 'I can't define pornography, but I know it when I see it.' Other minds are like this in a way; we cannot really define what they are, but we know what they are when we see them, and what they might do, by reference to our own minds.

If you are still uncertain about what a self-referential definition is, try this:

Self-referential definition (noun): a definition that does this.

Which approach to use?

Let us now assume that, to prevent autism, we equip an intelligent computer with a mental simulation system. It no longer has to try to get theories for other intelligent things that it meets, but can deal with their behaviour by using its own software to simulate them. The computer now has two entirely different ways of looking at things:

• analytical modelling: analysis to obtain the simplest theory (or a process roughly equivalent to this)
• mental simulation: possibly augmented by theory of mind

When the computer encounters something, or enters a new situation, it needs to know which approach to apply. Applying the wrong approach will lead to meaningless results; for example, attempting to analyse a human will lead to little progress and attempting to perform mental simulation on a tree will also be counterproductive! When the computer has selected the correct approach for a particular situation it can remember to apply that approach when it meets that situation in future. If it finds out later that the approach was wrong it can always change it, but it still needs some way of deciding which approach to use in the first place while it is still learning to be competent in the world.

This is actually a difficult problem. There are some ways that we might propose for the computer to attempt to do it that are not ideal.

One way of deciding whether to analyse or perform mental simulation would be to have our computer use mental simulation for anything that matched various visual patterns that were pre-programmed as being 'intelligent'. Such a process would involve the computer in trying to analyse everything it saw, unless what it saw resembled a similar sort of computer, a human or one of various other patterns that may be provided by its programmers.

This approach is not ideal because it relies on a very crude method of recognising other intelligences. What if another intelligent machine was severely damaged enough to avoid triggering a match in the software that makes the decision? What if an intelligent robot was partially hidden? What if a new type of intelligent machine arrived that did not resemble any of the pre-programmed machines and therefore had a 'social stealth' capability that prevented it from being recognised as an intelligent being? Our computer may learn from some of these experiences, but it would be at a disadvantage in the meantime. A machine that relied on physical resemblance would always be getting caught out. Of course, it could possibly learn some simple 'rules of thumb' from experience but this in itself has problems: how is a particular rule known to be true in the first place?

We might think that another approach would be to use mental simulation for anything that behaves like an intelligent entity and analyse everything else to produce a theory. The problem with this idea is that is essentially meaningless. To actually decide that something is behaving like an intelligent entity our computer would need to have either knowledge of a few simple characteristics of intelligent behaviour or would need to have a deep understanding of it. Such a deep understanding would be difficult to program into a machine and if the computer only has a few simple characteristics to identify intelligent behaviour then it can easily be caught out.

As an example, a robot controlled by the computer may be travelling through a forest where it finds stones in an intricate pattern. How is it supposed to arrive at the idea that this is a sign of intelligence unless it has specific information about how to deal with this?

This is the problem: how do we reliably detect intelligence when it is present so that we can use mental simulation, leaving everything else to be dealt with by analysis to obtain a simple model?

How approach selection would be performed

It turns out that there is actually a fairly economical method that the computer can use to select an approach. The whole point of introducing mental simulation was to deal with the fact that other intelligences are just too complicated to describe easily in theories. We can actually use this as the indication of whether or not something in the world is caused by intelligence! This is how it could work:

When the computer encounters a thing or a situation it would attempt to perform analysis to find a predictive model or theory that explains it. If it manages to find such an explanation then it knows that the thing or situation does not involve an intelligence and could simply use its theory to deal with it.

If the attempt to find a workable theory to explain the thing or situation failed completely then our computer would now know that it was dealing with something too complicated to be easily explained. This is actually one of the big problems of intelligence in the first place, so there is some sense in deciding, therefore, that it is dealing with something intelligent. This is what the computer could do: anything that lacks an explaining theory could simply be assumed to be associated with an intelligence more or less like itself and mental simulation then applied to it. This approach may seem crude, but in almost all situations it would actually work extremely well.

I am not suggesting that the choice between analytical modelling and mental simulation/theory of mind will always be made in this way. As the system gains experience it can learn what choices would be made in various situations and most simple situations would be resolved by using learned rules. The method of approach selection that I have described would still be needed though because:

• there would still be novel situations in which an intelligent agent cannot be recognised by using previous experience.
• rules for recognising intelligent entities would still need to be learned in the first place. The learning of new ways of 'detecting' other intelligences requires some way of testing possible rules to find out if they give the correct results. I suggest that the method I have described, that is to say using mental simulation/theory of mind when analytical modelling fails, provides such a way of testing for intelligence and various learned rules that may be adopted later can be tested against this method to see if they give the same result.

The agent illusion

You may well have seen where this is going by now.

Introducing intelligence and making the switch to social modelling to deal with anything that is not explained by a theory is fine, but this is not founded on any firm logic. This is only done because of the enormous problems of trying to make theories that include intelligence.

Nobody has an explanation for everything and the intelligent computer will be no different. It will always encounter things that it cannot explain. This will remain true, no matter how good its modelling is. No explanation can account for everything; any explanation can only provide a simpler theory to account for complex things and that theory must assume something as its basis. Even if the computer manages to 'explain' reality in terms of basic things like subatomic particles, it will then have to assume the existence of those particles until it finds a deeper explanation, and even that explanation will assume something that causes the subatomic particles to exist. Every explanation must leave something unexplained and, when something cannot be explained by a theory and has to be assumed, a computer built like this may have a tendency to make the decision to presume that an intelligent agent must exist at the point where the explanation ends. This means that the computer's own inbuilt systems to detect other intelligences and apply social modelling to them, which are prompted by failure to explain, may cause a wrong diagnosis at some point where the capability for easy explanation ends. At this point a computer could switch to social modelling to deal with a situation in terms of an intelligent agent that does not actually exist. I shall refer to this effect as the agent illusion.

It is important to recognise that this agent illusion could be experienced by an artificially intelligent machine even if there is no evidence for the existence of the agent! We have a possibility of a machine attributing the existence of reality, or some feature of it, to an intelligent agent and regarding this as basic and 'obvious' knowledge. I suggest that the agent illusion would have significant similarities to god belief in humans and, as this model appears to suggest the possibility of something similar to god belief in computers, I also suggest that an explanation like this may be the main cause of religious belief in human societies.

The agent illusion as a cognitive illusion

The idea of flaws in the way that thinking systems reason may appear a little strange, but there are other examples of this type of problem in thinking systems. A good example is provided by optical illusions. These are images that cause errors of interpretation to be made when subjected to processing by the brain. The important point here is that the illusion is not really a property of the image; it is a characteristic of the way that the brain interprets it.

We can ask the very reasonable question about optical illusions: why should a brain do this? Why do we not see reality exactly as it is rather than interpreting some things wrongly? The answer is simple: our brain evolved to deal with reality well enough to allow us to survive. Making a brain that could accurately interpret every image provided to it may be possible in principle, but in practice it would be extremely difficult. Evolution tends to cut corners and optical illusions are evidence of corner cutting in the workings of the brain. The result is actually not too bad: a brain which can analyse imaging data extremely quickly and which is only seriously fooled by a relatively small set of images that do not tend to occur in nature a lot anyway.

In his popular book, Massimo Piattelli-Palmarini, a cognitive scientist, describes a category of illusions known as cognitive illusions [9]. Just as an optical illusion occurs due to shortcuts in the way our brain interprets visual data, a cognitive illusion occurs due to shortcuts in the way that our brain reasons. Optical illusions are experienced by large numbers of people and it is just as natural that cognitive illusions should also be commonly experienced.

The agent illusion would appear to be an example of a cognitive illusion. This cognitive illusion is different from optical illusions in one important way: whereas an optical illusion may be easily dismissed when the person experiencing it gives it proper logical consideration, a cognitive illusion of this type may not be so readily dealt with; the reasoning parts of the mind, that might be used to dismiss an illusion caused elsewhere, are the very parts of the mind that are actually experiencing it!

So, why do some people not believe in a god?

If god belief is primarily caused by an agent illusion of the type that I have described, it cannot be inevitable that the cognitive illusion of an intelligent agent causes god belief in all human brains. The existence of non-believers is evidence of this.

The cognitive illusion hypothesis is, however, stated only in very general terms. It is likely that the real processes working in the brain, or in an artificially intelligent machine, will not be as 'all or nothing' as this. It is also likely that theory of mind does play some role in allowing us to deal with each other; after all, performing mental simulation regarding another person, or deciding how to 'fit' intelligence as an explanation to any situation, would require some assumptions to be made and some knowledge would be required to set up these assumptions. I think that it is plausible to regard the agent illusion as something that humans have a tendency to experience, rather than something that is inevitable in any brain. We may also expect that similar sorts of complications will arise in future computer systems, so that god belief resulting from the agent illusion is a possibility, rather than a certainty, in computers, depending on how machines are constructed, the importance of mental simulation/theory of mind in their functioning and the environment to which they have been exposed.

In human societies, some people may be more predisposed to experiencing the agent illusion than others. Various individuals may experience the agent illusion, in the sense that they feel that there is a 'higher' intelligence, but their reasoning may lead to the conclusion that there is nothing behind the illusion, causing them to dismiss it as a product of their minds.

The role of emotion should also not be ignored. Although I suggest that the explanation given here is the primary reason for the start of religion in human society, it would seem reasonable to think that emotions may play a part and strongly reinforce the tendency towards belief provided by the cognitive illusion in the case of some individuals.

Indoctrination also plays a role in much religious belief. Some people may be liable to experience the agent illusion and attribute reality to an agent as a result, but may need to be with other people who will reinforce belief in a god, or they may need to be subjected to indoctrination from an early age to properly establish the belief. Indoctrination can be powerful though, and it would seem possible that even a person who has no natural tendency towards religious belief could be made to believe in a god by teaching him/her to believe, particularly when young.

How the agent illusion would resist logical argument

In my opinion, the hypothesis suggested here appears to explain why religious belief can sometimes be so difficult to argue against and why people are quite happy to resort to faith, rather than giving any good reason or strong evidence to justify it.

When we look at this in the context provided here, we can see why this might be so. The belief in a god is not caused, mainly, by logical thoughts in the brain, but by the decision of the brain to disregard reason. I am not saying that this means that theists somehow have defective brains. Far from it; I am suggesting that the mechanism that rejects reason in this way is vital to allow humans to recognise the existence of other people at a young age. Even when we use this system to deal with human beings this whole way of working seems to be uncomfortably close to faith, with no obviously good reason, before we actually bring a god onto the scene. I am however suggesting that the theistic claim that a god exists may be of questionable correctness.

Atheists like me, often tell theists that it is meaningless simply to theorise that an intelligence exists unless you are proposing that this intelligence is fully definable, that is to say that if we investigated it closely enough we should be able to describe how it works. Another common technique that atheists use is to suggest that it is useless to propose a god as a theory to explain things if you do not have an explanation of the god's origins.

I think that if we were to interview very young children and challenge their belief that their parents existed we would encounter little in the way of logic and the statements that they made to 'prove' that their parents are real would have a lot of similarity with statements often used to support the existence of a god. We would be likely to receive answers such as, 'because I know they are real,' or 'I can see them,' or 'I talk to them all the time.' Any question regarding the detailed nature of what exactly it is that they are claiming to exist when they say their parents are real is not likely to result in a very informative answer. A child might happily say that his/her parents made something complex that appears in the house, by way of explanation for its existence, but is not likely to think that an explanation of where they came from is really required to justify belief in their existence; in fact, if you challenged the assertion that the parents are real you may well be laughed at for making such a stupid suggestion. The child might not be able to explain, in any detail, why the suggestion that his/her parents are not real is so stupid; it is merely obvious to him/her.

This is actually very close to being an act of faith! To a very young child the existence of other people is simply self-evident and there is no need for logical argument about evidence or such things. The child actually believes in something that he/she really cannot justify logically and this is absolutely necessary to allow any meaningful communication at all. Our acceptance of the existence of other people may be on firmer ground because we do have a better grasp of what people are, how they got here and the evidence that supports their existence, but a child needs to start communicating with his/her parents long before he/she has had time to develop the faculties to deal with these issues.

This is to be expected as a consequence of the mental simulation capability of the brain. The decision to use mental simulation to deal with something is not based on reasoning or evidence; it is based simply on unexplainable complexity. It may be difficult for logic to attack an incorrect judgement that a higher intelligence exists because the judgement is not based on logic; it is simply based on the decision made by the brain to abandon logic.

The approach of asking where a god came from, if he is supposed to explain where everything else came from, is also going to run into problems because, in my opinion, it is a characteristic of this cognitive illusion that it takes over where explanations end. Intelligence is assumed when we cannot explain and no further attempt to explain is required. In fact, it would be counterproductive for the brain to assume intelligence to explain something really complex and then try to explain the intelligence as it would merely be facing the complexity problem that it has just avoided.

We can expect similar problems if we use the argument that someone who claims that a higher intelligence exists should be able to define it and tell us exactly what they are claiming to exist. The mental simulation system intervenes when we are dealing with something too complex to define easily by conventional means and one way of looking at its role would be to say that it succeeds in obtaining a workable definition by relating the definition to us, so that the definition of intelligence simply becomes 'something like me'. This is why questions about what a god actually is often get nowhere with theists. They feel that that they know what a god is and that they can define it, but only in terms of their own intelligence.

I have often encountered theists who have asked me questions such as, 'How can you prove your mother exists?' or 'How can you be sure that you exist?' and the motivation for asking questions like this becomes clear when we see how the god illusion is constructed. The brain does not reason that there is a god. It thinks that it has 'seen' this god in a very basic way, as we would normally 'see' another person. Philosophical argument, in such a situation, may be regarded as irrelevant and, to many theists, suggesting that a god claim makes no sense because god is not even properly describable, appears to make as much sense as suggesting that you do not exist unless you can accurately describe how your mind works.

Testing the idea

This article has suggested that some computers could, one day, acquire belief in a god and it has proposed a way in which this might occur. Clearly, this idea could not be tested directly now; until we have machines with sufficient intelligence to make god belief an issue it will be little more than speculation.

When the same idea is extended to give an explanation for the origin of religious belief in humans though, there are possibilities for research to find out if this is what is going on.

One test would involve interviewing very young children and analysing the results to see if there is a similarity between their justifications for their knowledge that other people exist and the justifications often given by theists for the claim that a god exists.

Another way of testing the hypothesis deals with issues that may be unpopular with some atheists. Frankly, some theists are really going to like this.

It is clear that if the agent illusion occurs it affects some people more, with regard to actually causing belief, than others. Now, I suggest that the illusion of a god is caused in part by the capability of the brain to perform mental simulation. Severely autistic people may have a reduced capability to perform mental simulation, but their autism is a condition with effects varying widely in severity. For example, when Baron-Cohen, Frith and Leslie performed their classic false belief task experiments they found that some autistic children could complete this task, but failed a more sophisticated version of it, intended to test the ability to understand the beliefs that other people have about other people's beliefs. (That was not a typographical error by the way; that is really what they tested.) This suggests that the brain's mental simulation capability may function with varying degrees of effectiveness in people with autism, but let us suppose that the mental simulation capability may function to varying degrees in people who are not classified as autistic. What would the consequences be of some people having a higher mental simulation capability, and others a lower mental simulation capability, than the average?

There is, in fact, a condition known as Asperger's syndrome [10, 11]. It has some of the characteristics of autism, including poor eye contact and difficulties in social interaction, but to a less serious degree than with autism and without the mental retardation that autism causes. Many people with Asperger's syndrome are able to describe their experiences very articulately and often report a feeling of being in an alien culture, as if all the people around them are acting, to some degree, in an arbitrary way. This has sometimes been described as the 'wrong planet' feeling. Some researchers view Asperger's syndrome as a separate condition to autism, whereas others think that it is actually a mild form of autism.

A condition that affects the ability to understand other people, while leaving intelligence apparently intact, is an obvious window into the issue of whether or not mental simulation processes in the brain play a part in making people think that a god exists. One solution, on the face of it, is to do nothing more than ask people with Asperger's syndrome if they believe in a god! Any serious research project would, of course, do rather more than this. It would involve assessing the degree of religious belief, as well as the degree of exposure to religious ideas, in samples of subjects with Asperger's syndrome and without it. A more sophisticated study might also investigate the degree to which 'autism-like' tendencies are present in subjects.

Things would be a little more complicated than this. Asperger's syndrome also occurs with varying degrees of severity. Additionally, it would be naïve to expect that any tendency towards religious belief would be switched off simply by having Asperger's syndrome. The extent to which people believe will be affected by experiences that they have had as well, the occurrence of indoctrination probably being important.

It seems reasonable, however, that we should expect there to be a statistical relationship between Asperger's syndrome and religion, so that a person with Asperger's syndrome is less likely to believe in a god. This could be shown by questionnaires given to people with Asperger's syndrome and people without it. These would need to be carefully designed to determine not only whether or not people are religious, but also the extent of their religious belief and the amount of exposure to religious ideas that they have received. The results would be of statistical relevance only; I expect that, while people with Asperger's syndrome will be slightly less likely to be religious, some people will exist with the condition who are extremely religious.

As well as people with Asperger's syndrome, some people will simply have a mental simulation capability that is slightly lower than that of others, but where this may lack any obvious cause and is not diminished enough to actually be classified as a specific disorder. Reduced mental simulation capability would cause slight difficulties in socialising and understanding others and we could look at people without autism or Asperger's syndrome to see if there is a statistical link between the extent of social capabilities and belief in a god; people with lower social skills should, statistically, be less likely to believe in a god.

Now, clearly, some theists are going to think that what I have said above is absolutely great. I can see some of the comments on some of the more aggressive theistic web sites now, such as 'Atheism is caused by autism!' or 'Atheists cannot see God because of brain damage!' I want to state clearly that I have not made any such claim. I have suggested that there may be a statistical link between Asperger's syndrome or reduced ability to perform mental simulation and atheism and that would mean that some atheists are atheists because Asperger's syndrome or a poor mental simulation capability has prevented the cognitive illusion of a god asserting itself strongly in such people, but there is no reason to think that all atheists, or even a particularly large number of them, are atheists for this reason. Similarly, there is little reason to presume that the cognitive illusion of a god, in itself, will be strong enough, and will have been adequate to cause god belief, without influence from other sources, in most religious people.

I am suggesting that the agent illusion does account for the existence of at least enough believers in society to cause religions to be created and that it is the main explanation for how religious beliefs get into the world. People with religious beliefs can then persuade other people to accept these beliefs and set up religious institutions, which can use a variety of methods to promote belief by more people.

Conclusions

This article has suggested that intelligent machines, if they are eventually made, will experience difficulties in dealing with other intelligences, a consequence of the high complexity of intelligent entities that they will encounter and the resulting difficulty in modelling them, and that this situation will be analogous to autism in humans.

It has been suggested that the problem of autism in computers will be solved by providing computers with capabilities known in psychology as mental simulation and/or theory of mind and that these capabilities, when present in computers, could cause machines to behave as if they are experiencing an agent illusion, that is to say having the belief that those features of a model which have to be assumed, due to failure to find any simpler model to describe them, are attributable to an intelligent entity. It has been proposed that the agent illusion is analogous to god belief in humans.

It has been proposed that the original cause of religious belief in humans, and some part of the religious belief present today, is due to an equivalent agent illusion (a type of cognitive illusion) in humans and research has been proposed to test this hypothesis. This research would seek to explore the relationship between the degree of god belief, the extent of environmental influences that could predispose an individual to god belief and the extent of symptoms related to autism in human subjects.

There are some matters that the article has not addressed; for example:

If religion did arise in computers, for how long would it last? Would it be a significant feature in the future development of computers or would it be a transitory event, over in a few years or centuries? Would humans find religious belief in computers useful or not? Would they take any steps to promote it or end it?

What form would religious belief in computers take? Would it merely be the assertion that 'somebody' is 'out there' or would machines adopt a more structured theology? If they did adopt a theology would it be one of their own devising or would machines actually be capable of adopting human religions?

Is there any possibility of religious belief being a factor in conflict between machines at some time in the distant future and, more importantly for us, is there a chance of religious belief in machines playing a role in conflict between machines and humans?

I leave these matters for the reader to form his/her own views.

References

[1] Web Reference: Russel, R. K. (2002). Why Occam's Razor.
http://parallel.hpc.unsw.edu.au/rks/docs/occam/occam.html

[2] Wimmer, H. and Perner, J. (1983). Beliefs about beliefs: representation and constraining function of wrong beliefs in young children's understanding of deception. Cognition, 13, 103-128.

[3] Gordon, R. M. (1986). Folk psychology and simulation. Mind and Language 1:158-71.

[4] Home Page of the Centre for the Study of Autism, Salem/Portland, Oregon. (n.d.) Retrieved June 29, 2003 from http://www.autism.org/

[5] Autism: Signs and Symptoms. Neurology Channel. (n.d.) Retrieved June 29, 2003 from http://www.neurologychannel.com/autism/symptoms.shtml

[6] Baron-Cohen, S., A. Leslie, and U. Frith. (1985). Does the autistic child have a 'theory of mind'? Cognition 21:37-46.

[7] Blackburn, J., Gottschewski, K., George, E., L., Niki.(2000). A discussion about Theory of Mind: From an Autistic Perspective. Proceedings of Autism Europe's 6^th International Congress, Glasgow 19-21 May, 2000. Retrieved June 20, 2003 from http://www.autistics.org/library/AE2000-ToM.html

[8] Gordon, R. M., and J. Barker. (1994). Autism and the 'theory of mind' debate. In G. Graham and L. Stephens, Eds., Philosophical Psychopathology: A Book of Readings. Cambridge, MA: MIT Press, pp. 163-181.

[9] Piattelli-Palmarini, M. (1996). Inevitable Illusions: How Mistakes of Reason Rule Our Minds, Re-issue - Published by John Wiley and Sons (originally published 1994, in Italian, as L'Illusione Di Sapere).

[10] Ozbayrak, R, K. (1996-2002). Asperger's Disorder Home Page. Retrieved June 29, 2003 from http://www.aspergers.com/

[11] Kirby, B. (2003). Online Asperger Syndrome Information and Support. Retrieved June 29, 2003 from http://www.udel.edu/bkirby/asperger/

Update - 18 January 2005

Since this article was written it has come to my attention that Scott Attran's book In Gods We Trust: The Evolutionary Landscape of Religion (published by Oxford University Press in 2002), before this article was written, advances a case for religion being caused, in part, by a social modelling system.

I will leave this article on this site as it seems to differ from Attran in some areas, but it is only fair that I also make the existence of the book by Attran clear to readers.