You probably noticed that there was a crescendo of difficulty in the sequence of previous articles, culminating appropriately with the mention of God. What could possibly be added after the article on Superintelligence and God? That article ended with a proposition about consciousness   being some kind of computational system, and a story about some more far-fetched consequences of that proposition. The story was based on the fact that since all computations can be simulated on other computers, consciousness could be simulated too.
In the opposite direction, we saw in the AI Versus Human Intelligence background article that this computational system does not seem to work like a classical computer at all (i.e., a transistor-based computer) and that research is being done into the possibility that it may be using quantum computations through neuronal microtubules. But other than those two references, consciousness was mainly left out on purpose.
Given that our subject is AI, why is it important that we do not leave consciousness out and focus on intelligence only? Because the relationship between AI and Artificial Consciousness (AC) is very consequential. As we will see below, some argue that consciousness and intelligence are orthogonal, and therefore AC would need a completely different approach than AI. If the AI that we develop now cannot have consciousness, would we be OK living in a future with far more numerous AI zombies (=extremely intelligent AI systems without consciousness) around us?
Our working assumption (although we will explore many other interesting possibilities), which is held by a majority of scientists and engineers working in AI development, is that consciousness and intelligence are not orthogonal and that in fact they are both emergent biological processes housed in the brain and moreover, that they can both still be thought as products of evolution. We will in a way adopt Gerald Edelman's position, which will be explained below. This is by no means the only way to view AC, and the reader will have to discriminate between the various theories outlined below.
In general, approaches to consciousness are either reductionist (for example: idealistic, meaning that everything reduces to mental processes and that nothing else is real; physicalistic, meaning that everything reduces to natural phenomena and nothing else is real), or dualistic, meaning that mental processes, including consciousness, have their own reality, distinct from the natural world. And finally we have the more radical idea of panpsychism, which we already saw that it has Hinduism/Buddhism origins, the idea that consciousness is a fundamental characteristic of nature itself, i.e., that rocks and atoms have some form of consciousness already built in. The difficulty of dealing with consciousness has been given a more definite formulation by David Chalmers.
Hopefully we will not be far off if we sum up Chalmers' view as proposing that while explanations of human behavior and human cognition may be easier to find (or easier to accept), the problem of why we have inner subjective experiences is of a different degree of difficulty altogether. Translated into a language more available to those of us working in AI, this view could be interpreted as follows. While we could design classical computational devices that behave and think like a human it would be far more difficult to think of consciousness as a classical computational device too. (Here classical means ordinary computers, i.e. not quantum computers)
Given the current level of our AI, this caution would certainly be justified. Although it is difficult to think of consciousness as a classical computation at this time, it is possible, given our evolutionary bias, to conceive of consciousness as having occurred through evolutionary biological changes of the brain, and therefore be still less mysterious and difficult.
Perhaps the beginning of our current efforts to understand consciousness can be traced all the way back to the work of Gerald Edelman, in the '70s. Edelman is one of the strongest proponents of strictly biological explanations of consciousness and of cognition as well; as strictly biological processes, consciousness and cognition are subject to evolutionary changes; Edelman does not subscribe to a CTM (Computational Theory of the Mind). Giulio Tononi, whose work we will look at in some detail soon, is very much influenced by Edelman, the two having published together a very influential work in 2001: "A Universe of Consciousness: How Matter Becomes Imagination".
Some things can be tied together now. There is another direction of thought, in addition to Edelman's, also supporting the argument that processes in the brain are more complex than classical computations. Sir Roger Penrose (whose work on consciousness we have already mentioned in the AI Versus Human Intelligence article) uses Gödel's Incompleteness to derive such thought. Namely, there must be something more to the human brain, because humans seem to understand certain things that conventional formal systems cannot understand, so these formal systems appear more limited than the human brains are.
We have seen in the Foundational Questions article that even though formal systems can be patched, for example you can add a patch to prove their consistency through a more powerful system, then the new system also suffers from the incompleteness issue, and so you end up with an unending tower of incomplete patches if you insist on having power and consistency. In other words, limitation is present in any formal system, which somehow the brain seems to find a way to overcome.
It follows that if the brain uses a different mechanism to make sense of the world, maybe consciousness is more than intelligence, because presumably intelligence can be all captured by classical computations. Penrose's argument is fascinating and frames our future discussions as to what may or may not be possible in our quest for AC.
Where are these microtubules located? To answer that, a short detour into the magnificently complicated world of the biological neuron is needed. You will not loose the flow of this article if you do not read about the neuron now, but sooner or later it would be helpful to at least pin down the terminology. Moreover, the biological neuron and the networks of such neurons are too important for us, because neural networks try to mimic their functionality. So we link to a short introduction, which you can consult now or later. Edelman, in the video above, already mentioned the neurons and their axons and dendrites.
So how can we bridge this extraordinary complexity and try to establish a path forward towards AC, or at least how can we attempt, even though we may be unsuccessful? First of all, maybe it would be good to start by establishing that the study of consciousness should be a scientific endeavor, not a spiritual one. Computation or not, it is a biological phenomenon, and can be studied as such. John Searle, now a professor emeritus at Berkeley, establishes some characteristics of consciousness: it is real, actually it is the most real think that we possess, it exists as you read this and think about it, it is irreducible and unified (these characteristics are set down more formally in a theory of consciousness which we present later). You press the Share button for this page because you like it so much that the thought becomes an action. It is the result of a sequence of firings of neurons in the cerebrum all the way through synapses to the neurons in the fingers and the fingers press the button!
There have been many attempts to build a theory of consciousness, and most of them do not belong here, because we choose to focus only on those that might help us to build better AI. So these theories should have predictive power and should somehow incorporate intelligence (and consciousness if possible) as a form of information processing in them; even biological information will do, we will look at synaptic activation/inhibition between neurons as a form of information transmission.
From our point of view with regards to the future of AI (a reductionist point of view), the two most promising such theories are the Integrated Information Theory (IIT), proposed by Giulio Tononi and the Global Workspace Theory(GWT), proposed by Bernard Baars. IIT is a top-down approach, while GWS works bottom-up. These two theories compete in many respects and discovering features that bind them together seems to be the right thing to do.
One concept they both leave out is energy and one may contend that information processing without considering the energy cost which this processing requires will not capture the whole story. (One can begin to think of the energy cost of information processing systems by appreciating the cost of giving birth to humans with increasingly large skulls housing such information processing systems. Which systems consume 20% of the entire energy the body needs, despite weighing only 2% of the total body.)
When I first saw the presentation of IIT two things caught my attention. First, the white board was full of mathematics, and secondly the word information was front and center. I instinctively felt that something must be there, that the approach was in the right direction. However, we will entertain a different interpretation of that theory and let's see why. There are two problems for us: first, the seemingly deep connection with panpsychism and second, the derived conclusion that no AI could ever be conscious. We will always assume embodied cognition, the idea that the brain needs a body to operate fully (not just that the body needs the brain to operate).
IIT avoids a philosophical stand (and dodging hard questions by asking simpler ones is a quality of scientific endeavor) and tries to produce a theory about the properties of consciousness. IIT then isolates the features that a physical system must have in order to support consciousness. Humans and AI and rocks and atoms are such physical systems and the question of whether they can support consciousness loses its speculative nature and becomes approachable. The idea that we can actually measure the degree of consciousness \(\phi \) in physical systems is a uniquely powerful one. And moreover, the theory has predictive power, as any respectable theory should; measuring \(\phi \) offers immediate practical benefits for the medical fields as \(\phi \) can distinguish between states of dreaming, anesthesia, coma, vegetative state, etc. How these measurements are done is interesting, but it is outside of the scope of our discussion.
As opposed to the top-down approach of IIT, the Global Workspace Theory (GWT) is a bottom-up approach to understanding consciousness. It is quite digestible by non specialists because it proposes an architecture for consciousness which can be easily translated into a computer and network architecture. The theory proposes an analogy with a theater, in which the stage represents consciousness, the actors shuffling in and out representing sensory information, and the viewing public representing the larger memory banks in the brain, especially the subconscious memory.
This can also be translated into computer speak as follows: at the center of the representation we have a high-speed memory (a cache in computer speak) which holds the temporary imprint of consciousness; the contents of this memory (the global workspace) is continuously swapped in and out of larger memory banks in the brain, such as the short-term and the long-term memories; the theory then posits that our momentary inner experiences are due to the quick neural connections between this global workspace and the sensory organs, especially vision. Bernard Baars is the author of this theory:
The Central Role of Information
Is it strange that consciousness can be looked at as a mechanism of integrating information in a "maximal" way? Not at all, it's almost a given that we should look for such explanations, given that we have now come to look at life itself as a mechanism of integrating information in a "maximal" way, where maximality is understood to occur as benefit for a purpose. Life's distinguishing feature is purpose. Rocks do not have an inner purpose, but organisms do: their purpose is to survive and reproduce. Biology gives us only one basic principle with which to analyze this purpose, Darwin's evolution, but it cannot explain how this purpose can come about from inanimate matter, which is governed by the laws of physics. (Can physics itself explain the emergence of purpose? That seems an untouchable question, although attempts have been made.)
The central concept that seems to make its way in these types of questions is information. So, can information also help explain the extraordinary jump from lifeless matter to purpose-driven life (and eventually to consciousness, as IIT proposes)? The second law of thermodynamics, one of the deepest laws of physics, tells us that within closed (i.e., isolated, with no exchanges with their environment) physical systems, disorder always increases; the precise term for this measurement of disorder is entropy, and even though we try to avoid using big technical words, this one cannot be avoided. Living organisms seem to defy the second law, because their inner entropy seems to decrease, not increase (more order, not less). How is that possible?
First of all, organisms are not close systems as they consume energy from their environment to maintain and decrease their entropy. Edwin Schrödinger wrote that living organisms feed on "negative entropy", by storing and using information obtained from their environment; this information would be called genetic information, which is nothing but a software code of how to survive and reproduce, i.e., how to have purpose. He did that in 1944 and in 1953, Crick and Watson discovered that this genetic information is indeed stored in the helix-like structure of the DNA molecule.
So it is this information that the organism has collected within its genes, from its environment, that it has allowed it to have purpose (survive and reproduce). So life is an information system built to optimize the storing and processing of information from its environment. We have gone full circle from looking at the brain as such a computational system to cells themselves being such computational systems. In that respect natural selection can be seen as the principle of minimizing the thermodynamic cost of information storage and processing. But how the first jump from no information integration to some information integration is made is still mysterious.
Many times this second law of thermodynamics is misused to derive theological conclusions, most notably to derive the existence of a Creator. Evolution, so it goes, cannot be true because it assumes a decrease in entropy for organisms, and entropy should always increase; so we must have a Creator. First of all, evolution is a biological principle stated for and bound to the Earth. As a thermodynamical system, the Earth is not closed, it is constantly consuming energy from the Sun and dissipating energy back to its environment.
Secondly, the second law refers to the system as a whole, it is an averaging statement over its parts. Parts of the system can decrease their entropy, while the system as a whole still has an increasing entropy. The same people who are using these kinds of arguments against Evolution seem to also be against the evidence of climate change caused by the activity of living organisms, especially humans, whereas one may even look at climate change as a corollary to the second law: the Earth biomass increases the Earth entropy.
But why is this relevant to our topic? As we saw in the article Graphs of Data, the most powerful AI systems consume vast amounts of energy. They are far from doing that consumption efficiently. So AI systems may continue to become more and more complex and powerful (low entropy) at the expense of increasing the Earth entropy, just like humans do. It is likely that AI systems, civilian or military, because of their energy hunger, may contribute to climate change in the future, even though some of these energy sources are renewable.
In 2017, the data centers of the world consumed 416 terawatts of electricity, about 3% of the electricity generated on the entire blue dot. These AI systems do not possess the natural mechanisms of efficient information integration. Not only is their (\(\phi = 0\)), their coefficient of information integration power relative to energy use is very low. Nature so far has been much better at creating predictive systems with high such coefficients.
We looked at some complex systems functioning within changing environments. If they must consume energy efficiently, they are likely to become prediction machines. It would be expensive to store all the history that the system has learned from its environment, so it must store knowledge efficiently, i.e., build a model of the environment in which it functions. And therefore it must be able to predict from this knowledge and the current environment state. Efficiency seems to require prediction, so nature itself is building predictive systems.
Why would nature evolve our brains to exhibit consciousness? As we said before, I cannot be sure but my senses tell me that you the reader (let's leave out the search engine crawlers) are a human. I cannot test your consciousness but I can assume that you have it. Being familiar with those capabilities that you have will allow me to predict better your actions. So one may also think of consciousness as nature's way of improving the capabilities of the prediction machine. It seems that this prediction machine is also being improved by the dreaming stages of sleep, during which consciousness is ON:
You have probably figured out by now that our baseline philosophical stand is reductionist, and more specifically it is that of metaphysical naturalism, i.e., that everything can be reduced to natural phenomena, including our minds (and consciousness as well). It may well be that the computations needed for consciousness are quantum computations, at a far more intimate level than what we have in traditional forms of computation, but nevertheless at the bottom of it all we still have natural processes.
The bit of experimental evidence we have for such a naturalistic stand comes from the realization that as the current AI systems gain in power, they begin to integrate information in ways that are beyond our current understanding. Neural networks in particular accumulate knowledge in some of their nodes that correspond to features of the system we are trying to model that we have not specifically asked for. This may look like an emergent phenomenon, namely that more data and more powerful neural architectures lead to the appearance of properties which cannot be easily explained.
So, is it possible that a sufficiently strong AI system begins to integrate information to such a degree that it becomes aware of its own abilities, and that an inner movie (call it subjective experience if you wish) begins to run through such a system? To pretend that our current experimental observations warrant a positive answer to this question would be a huge stretch, but nevertheless these observations point to the idea that "evolution" should not be discarded as an explanation for the emergence of consciousness. Let's pin this down as the emergence possibility, because we will refer to it and try to refine it later.
GWT and IIT have shed some light on the problem of consciousness, by treating it as an information processing system. Whether we adopt the IIT top-down approach or the GWT bottom-up approach, one question that would remain is what exactly are the brain structures and the connections between structures that support the formation of consciousness. Such study has been started and it is called the Neural Correlates of Consciousness (NCC), but before we look at those correlates, we'll see first that mapping of the brain parts and its connections, structurally and functionally, is also begun in a larger context, not just for the purpose of understanding consciousness:
We promised above that we will return to the Neural Correlates of Consciousness (NCC). The prevailing assumption, held by most researchers in the field, being that consciousness emerges in the brain, and cannot exist outside of the brain, the question is then what structures in the brain and what connections between those structures support consciousness. There is some progress. The claustrum is the most connected part of the brain, it has connections to almost all parts of the cortex, and so there are experiments which try to establish the claustrum as the director/conductor behind the stage of the GWT theater metaphor, who is synchronizing the fleeting parts of consciousness residing in various regions of the cortex. On the other hand, the example of the extraordinary case (mentioned in the following video) of a conscious patient without a cerebellum (where the subconscious resides) is evidence that the stage in the GWT theater metaphor can function without an audience!
We can now refine the emergence possibility we discussed above, but we need a new concept. A philosophical zombie (as opposed to traditional folk usages of the term zombie) is a thought experiment about a possible entity which looks like a human and behaves like a human, but has no consciousness, i.e., a human-like entity but with \(\phi = 0\). Whether such zombies are logically coherent and physically realizable is at the core of various philosophical views on consciousness. Some argue that zombies cannot exist at all, other argue that the concept is logically coherent but physically unrealizable and so on.
But since we focus on developing AI systems with rising levels of intelligence, non-existence and un-realizability look temporary. In fact, it is likely (based on current engineering) that many AI systems that are extraordinarily intelligent will in fact be such zombies and that they will be more numerous than us, conscious humans. Could consciousness arise inside such AI zombies? Could complexity in structure and a rising power in processing and integrating information (inputs from the environment) lead to the appearance of consciousness inside an AI zombie? That question is fundamental, and of course, we are far from having an answer.
If presented with a human-like creature, how would you know if it is a zombie or not? Let's assume that you can see it and touch it, that it is not a system behind a wall answering your questions, as in the Turing test setup. The system, having human like intelligence, if hid behind a wall, would pass the Turing test with no problem. How could we test whether that physical system is a human (\(\phi \ne 0\)) or a zombie (\(\phi = 0\))? Do you think it would be possible to size up its \(\phi\) from the outside, and without hooking up the system to an MRI machine or trying to measure biological responses with some other instruments?
Is there a set of predictive problems we could present to this system whose answers only a conscious human could find? After all, we want to figure out how good the system is at integrating information. If its power of integrating information seems extraordinary, far above the human level, it may still be a very capable zombie. If it passes the mirror test, recognizing itself in the mirror, it is either a human or it is again a very capable zombie, one that has learned to recognize its own image in pictures. Maybe such test of consciousness is not possible after all, as a sufficiently intelligent zombie will always be able to appear conscious to humans. This obviously has implications on how we proceed in developing AI.
What is reality?
First, we have seen that it is a useful simplification to treat both cognition and consciousness as computations; the mind would be the computation system performing these computations (i.e., the symbol manipulator, the Turing machine), and the neurons of the brain would be the physical entities supporting the computations; this is the so-called Computational Theory of the Mind (CTM). After we listened to Gerald Edelman's sobering views, we are also prepared to admit that useful a simplification as it may be, as an ultimate explanation of consciousness, it may be false. Understanding the complexity of the biological processes that would lead to consciousness in the human brain (and designing experiments which would prove that understanding) seems a bit too much at the present time.
But let's assume that a CTM can be somehow formulated in a widely accepted formalism, despite the warning from Edelman, and therefore both cognition and consciousness are computations. Then, they could be simulated, because all computations can be simulated. That opens up an entire slew of arguments, and we covered some of them already when we introduced the Simulation Argument in the Superintelligence and God article.
The Simulation Argument is a logically sound argument and it is very useful to keep in mind. But there is a strong temptation among younger workers in AI to entertain more spectacular scenarios of the type exposed in the Matrix movie. Some have proposed justifications that maybe the 6-day creation in the Bible is not that far off, since AI could spin a computer simulation of our civilization in 6 days easily. Follow-up deductions that the physical universe does not exist (and that everything is a simulated reality) are beginning to sound like the norm, not the exception. Because light is of such fundamental importance in the relativity theory of gravitation, others propose that the light of the pixels on a computer screen might be connected to this idea that we live in a simulation. If those Matrix-like scenarios seem wild to you, let's regain our ground and follow up with a fully mathematically justified approach:
We have seen in the Superintelligence and God article that Hinduism already contains the idea of panpsychism, namely that consciousness is a fundamental and universal property of our entire world, including the physical world (Buddhism pushes this idea even further, and Buddhism comes the closest to a world view somewhat compatible with scientific approaches). Panpsychism means that you can think of consciousness as an irreducible physical property of nature, like space and time (although even space and time might not be irreducible after all the punch of quantum physics!) . Or you may think of consciousness as a property present in all or some of the elementary particles, say a property on par with mass, charge and spin. Either way you will not be far from what panpsychism proposes.
Moreover, you can type "consciousness as ..." in the Google search box, and the auto-completion will show you the wide variety of possibilities that people entertain: matter, field, force, property of the Universe, a dimension, etc. This consciousness measure may be a natural number, a fractional number, an entire group of symmetries, or some other mathematical object, it does not matter since it is purely speculative at this time. So according to this view, our consciousness is nothing but the integration of all these consciousness "numbers" of all the particles that form our bodies. Weird and improbable as it may sound, panpsychism is appealing in its simplicity; one may argue that it is the ultimate expression of Occam's Razor, the most simplistic way to explain a world which has conscious beings in it.
Panpsychism is also appealing because it allows one to bridge all sorts of observable weirdness: weirdness of quantum mechanics, weirdness of AI and current neuroscience, and even the old weirdness of philosophy. Ultimately, one may choose to view this new consciousness property as being God's breath and the tour would be complete; according to this view, God would be eminently omnipresent. So if we are prepared to entertain these kinds of ideas, why not entertain some of their logical consequences? You would have to read the previous article on Superintelligence and God, where we try to establish a framework for dealing with questions about God.