Machines may be adept at going through the processes, but what would it take for them to reach understanding?

Artificial intelligence is one of humankind's greatest and oldest ambitions. The quest for non-human intelligence has captivated magicians, astrologers and mystics for as long as such professions have existed, but it took Aristotle to kick things off properly. He was the first to start organising laws of thought and the way they interact with the real world -- the basic concepts behind AI. That was in the third century BC, and 2,300 years later we still haven't cracked the problem.

Part of the trouble is that nobody knows what AI is. In fact, nobody even knows what I is. While consciousness is something we all have and enjoy every second of our waking lives, nobody knows how it all really works. A thought is as difficult to isolate from our mental experiences as a single breath of wind is from the weather.

A new sister discipline to AI, cognitive science, has started up to try and track down the mechanisms of mind: while it can say that linguistics, philosophy, neurochemistry, anthropology and so on are all part of the mix, it can't say how these combine to make up our selves.

All this hasn't stopped research into machines that think. Since the 1950s, when Alan Turing famously predicted that by the year 2000 machines would be able to pass as human in conversation, the field has attracted high hopes, brilliant minds and heartbreaking failure in equal measure. Because 50 years of failure eventually starts to affect funding, even in academia, the AI field has diversified and experts have established themselves in other areas where they can be said to have had some success.

Strong AI vs weak AI

There are now two sorts of AI: strong AI, which is the business of making computers think, and weak AI, which has computers modelling some aspects of human behaviour. Marketing people love weak AI -- if you see a product described as having artificial intelligence or being 'smart', the chances are that it's got some aspect of weak AI in it.

Weak AI has the enormous benefit that it can be described. There are a load of approaches, techniques and tools that have evolved. Knowledge engineering -- where information about something is codified and put into a database -- is of great commercial value, even if it has moved well away from its cognitive beginnings.

Fuzzy logic, visual recognition, natural language processing and other ways of dealing with real-world data all have their roots in AI while paying back some of the research investment in other fields -- the paperclip in Microsoft Office uses a strand of weak AI called Bayesian belief networks, while the grammar checker comes from AI language research. Yet while all these model part of what we know about our capabilities as sentient beings, none seem close to providing true sentience. Whatever that is.

Some philosophers think that mind is impossible to model: John Searle of the Philosophy Department at Berkeley has proposed this with his Chinese Room analogy.

Take a room with two slots in the wall, an English-speaking man inside and a rulebook. The rulebook tells him how to deal with Chinese sentences that are pushed through the slot -- how to choose characters with which to reply, and what order to send them back out through the second slot. The responses may be perfect Chinese, but it does not logically follow on that the man is actually understanding the language as a native speaker would, rather than merely processing it.

Thus, says Searle, any machine with a programmed set of responses cannot be considered intelligent. Unsurprisingly, this is not a popular position among cognitive scientists, but the fact that such a consideration still has currency shows that the AI as a field is a long way from being as established as physics.

Less drastically, many researchers, such as James Hendler of the University of Maryland, say that whatever machine consciousness is created is unlikely to be like ours. "Is your cat aware?" asks Hendler, pointing out that it undoubtedly is in some ways, but not in others.

A lot of AI research is dogged by the feeling that if something is intelligent, it should be in some way like us. Alan Turing's test for conversational skills in a computer has been the subject of some high profile prizes and much research but consistently produces results that, even when carefully limited according to subject matter and style, would make a toddler chuckle.

On the other hand, a program like Eliza, which makes no claims towards intelligence but merely reiterates what the user types, can easily fool some people for some time into thinking they're talking to a human. Nevertheless, there is no alternative starting point for AI other than ourselves.

Recreating a sense of self is an infinitely confusing goal

Our minds are many things, but they are not pure products of computation. What we consider our sense of self is as much mediated by our personal history and experiences and our interaction with others as it is by the raw grey matter of the brain. And mind is not what it seems -- we think we live in the present, where we perceive events and have thoughts in real time.

We don't -- we live about half a second in the past. Brain scans show that is how long it takes for a perception to become fully integrated with our awareness. Yet we can catch a cricket ball, drive a car and communicate with each other much more quickly than that -- all aspects of being that are in some way removed from our immediate awareness while seemingly part of it.

And the human brain, despite having many similarities to a computer, is unthinkably complex. Each of the neurons in the brain can have up to a thousand connections. When a certain set of conditions at those connections is met, the neuron fires -- but those conditions can vary from moment to moment, depending on what happened last, what state the neuron's in, what the conditions around it are. That's a very large number of variables to take into account, which the neuron does up to a thousand times a second -- and there are 100 billion neurons in the human brain. That's up to 100 trillion synapses, each of which would have to be modelled in a replica system.

Enormous numbers, even if researchers such as Hans Moravec of Carnegie Mellon University think that we'll achieve the necessary 100 million MIPS/100 million MB machines in twenty to thirty years' time. He points out that the human brain has many more neurons than it does bits of DNA -- in other words, the neurons are put together according to a coded scheme, rather than individually, and that much of the creation of mind probably happens after birth. Our machines, should we build them, may programme themselves as babies do.

Learning remains a mystery

As AI researcher Oliver Selfridge of MIT asks: "Can you think of a chore or duty that a human being doesn't do better the second time, or a chore or duty that a computer does that it does do better the second time?"

He points out that an impressively profound form of learning is called common sense, to which increasing attention is being paid. AI should above all be able to learn from its mistakes and its environment, something that neural networks and genetic or evolutionary engineering include in their basic technology.

Yet another strand of thought says that, as 50 percent of the brain is devoted to processing perceptions, more thought should be given to the mechanics of working out what's happening in the real world and combining these experiences into a cohesive map of the world outside the machine.

After fifty years, there's not much coherence to AI and cognitive science. It's a field where, in the absence of strong direction and much controversy, alternative theories abound. One of the more intriguing is that of quantum consciousness -- that thought emerges from the interactions of components at a sub-atomic scale. The idea's been around since the 80s, but has suffered somewhat from an almost total lack of ideas about how it might work -- and plenty of reasons why it might not.

Large-scale organised quantum events have only ever been observed in laboratory conditions where thermal and other chaotic effects are carefully removed, which cannot be said of the warm wet gloop in the brain.

However, anaesthesiologist Stuart Hameroff has recently proposed that microtubules, tiny structures within cells, may be the site of quantum consciousness. He's been joined by mathematician and consciousness theorist Roger Penrose, who also proposed quantum mechanisms for thought in his book The Emperor's New Mind.

Microtubules are cylindrical molecules around 25 nanometres across that were fairly recently discovered -- embarrassingly, the solvent once used to make specimens for electron microscopy dissolves their proteins, and it wasn't until the solvent was changed in the 70s that their existence was first spotted. The microtubules fit together like Sticklebricks to form an internal skeleton within a cell, and this is normally thought to be their major function.

However, they also possess the ability to switch between states in a nanosecond -- one of the fastest biological processes known -- and may well cooperate to process information in single-cell animals. Unfortunately, there's no known mechanism that would allow them to cooperate with other microtubules across cell walls. Which is where quantum effects come in, says Hameroff.

Exactly how or what is still open to conjecture, and whether there's more to it than the joke philosopher David Chalmers made -- "Consciousness is a mystery, quantum mechanics is a mystery. When you have two mysteries, well maybe there is really only one. Perhaps they are the same thing" -- remains to be seen.

So AI is a goal that nobody knows how to achieve, or whether it's achievable at all, or what it'll be when we get there. Not a strong candidate for hope. However, the spinoffs of research to date continue to be important and as computers get more powerful they'll be able to perform more apparently intelligent functions.

Whether this will eventually include true sentience like HAL9000 -- and whether HAL was truly intelligent -- is as debatable as the number of angels that can dance on the head of a pin. But if we interact with computers as if they were human and abdicate decisions to them as soon as they seem capable of coping, one day we may wake up and find that we've created an intelligent world by accident.

Our current study of strong AI is to the real thing what alchemy was to chemistry. Lots of people are searching for the philosopher's stone, but through sympathetic magic and stabs in the dark rather than through a comprehensive understanding of the problem. Like chemistry, real AI may turn out to be completely different to the thing for which we search in our ignorance. Twenty more years of study and twenty more years of technology, and we might just start knowing where to go.

In ZDNet's Artificial Intelligence Special, ZDNet charts the road to sentience, examines the technologies that will take us from sci-fi to sci-fact, and asks if machines should have rights.

Copyright © 2009 CBS Interactive, a CBS Company. All Rights Reserved.
ZDNET is a registered service mark of CBS Interactive. ZDNET Logo is a service mark of CBS Interactive.