Putting Ourselves Back in the Equation, page 10
3.2. INTEGRATED INFORMATION THEORY. Integrated information theory supposes that a neural network is conscious if its units are wired together tightly enough that they develop interesting group dynamics. The theory lays out a procedure to quantify their degree of integration, denoted by Φ. A disconnected, or feedforward, network, without any closed loops, has zero Φ. If some units are tightly integrated while others are feedforward, only the tightly integrated section is conscious. With enough wiring, Φ peaks. Beyond this point, additional wiring actually reduces Φ, since the network acts as a single block rather than as a true network with internal complexity.
Tononi equates Φ with the degree of consciousness of a neural network, be it an animal brain, an AI system, or really any network at all. This rubric appeals to our intuition that consciousness is not simply present or absent; it exists on a spectrum. You can be anywhere from comatose to wide awake and, even when awake, from zoning out to being fully present. Some achieve a heightened consciousness through meditation. We allow that other mammals and living things might be conscious, but assume, perhaps parochially, that they are conscious to a lesser degree than we are. IIT attributes these differences to the degree of brain integration. Placing all these conscious states on the same spectrum is probably oversimplifying the matter,42 but Tononi and his colleagues have some experimental justification for it: they have performed the split-and-randomize thought experiment on actual brains to measure, if not quite Φ, then something close to it. Put simply, they have built a consciousness meter.
“We wanted a tool to perturb the brain—knock on the brain and see whether it was responding,” Tononi told me. Their device combines two standard neurological instruments: a transcranial magnetic stimulation device, which is a handheld magnetic coil that zaps the brain to treat conditions such as depression;43 and an electroencephalogram (EEG) skullcap laced with electrodes to eavesdrop on the brain’s electrical signals. The coil delivers a magnetic pulse that acts like a clapper striking a bell, and then the EEG registers the resulting brain waves. An awake person’s whole brain reverberates, indicating a highly interconnected neural network. But in deeply sleeping or anesthetized people, or people in a persistent vegetative state, the response is localized and muted.44 The response indicates whether someone is conscious even if they’re unresponsive. (This is still an experimental procedure, and these people or their legal surrogates must give informed written consent.) They’ve also tested the system on rats and fruit flies.45
Having your brain zapped to probe your consciousness sounds trippy. Alas, having tried it myself, I can report that you don’t feel much of anything. When the doctor administering the test placed the coil against my head, the placement was a little bit off, and my finger twitched because the system was stimulating the motor areas of my brain. For a moment, I felt like a marionette. When properly situated, though, the system didn’t affect any brain functions. “It should prove the state of consciousness, without altering it.… There should be no induced hallucination or anything like that,” said Bjørn Erik Juel, a postdoc in Tononi’s lab. Afterward, I looked at the electrical traces and saw the waves of activity across my brain. They revealed nothing of what I was thinking, but did prove that different parts of my brain were communicating with one another and, if you accept IIT, that I was conscious.
Calculating Φ is just the start. Not only can you ascertain whether a network is conscious, you can also hazard a guess as to what it is conscious of. A thought, a feeling, a memory: these are configurations of the network. If you map its wiring in detail and work out how the neurons are switching on or off over time, you can read its mind. The brain-zapping procedure I experienced was too crude to look into my thoughts, but Tononi and his colleagues have done this for simple artificial networks. Each possible pattern of activity corresponds to an experience, and in some cases they can work out the structure of that experience—whether it has the features we associate with, say, color or spatial relations, as I will explore in chapter 8.46 But whether the experience is truly comparable to red or spatiality—or anything else we can identify with—is beyond mathematical analysis; only the network itself knows. The key point is that researchers have found a way to use mathematical and experimental methods to describe a correlate of experience.
THE MAN WHO IMPALED HIS BRAIN
Which part of your body is you—your conscious self? The brain has multiple parts and levels of organization, and for each, you can find researchers who think it’s the site of consciousness. Some put it in the cerebral cortex, some in the brain stem. Some think consciousness is not a property strictly of the brain, but involves the body, environment, and culture. Others go to the other extreme of scale and place it at the subcellular level, inside neurons.
IIT provides a clear path to answering the question: calculate Φ for the various candidate sites and look for the one with the highest value. That is hard to do in practice, but Tononi takes an educated guess. In humans, the cortex—especially its rear, or posterior, portion—is richly structured and interconnected, so he predicts that consciousness resides there. “Grid-like areas like the ones in the back [of the cerebral cortex], and even more so a pyramid of grids, are a fantastically good substrate for phi,” Tononi told me.
He and his colleagues cite neuroscience data in support of this supposition. Mélanie Boly, a neuroscientist who is also at Wisconsin, spoke in a lecture of a college student who, during the Spanish Civil War, escaped out a window, slipped, and impaled his head on an iron spike.47 The front of his brain was mangled; it’s hard to imagine a more devastating injury. Yet he survived and recovered. He got married, had two kids, and led a fairly normal life. The only indication that something was awry was that he was absent-minded: he had trouble sticking to a task and told the same jokes over and over. I know many people who never had an iron rod through their skull who are like that. But an injury to the rear of the brain leaves people much worse off, Boly said. They can lose entire categories of experience, such as color, or sink into a vegetative state from which they never emerge.
“Prefrontal cortex, which is a huge and highly evolved part of the brain, doesn’t seem to matter too much,” Tononi said. “But some parts in the back of the brain, for instance, in the back of the cortex—touch those, and you’re going to screw up something with consciousness.” As further evidence locating consciousness at the back of the brain, Jun Kitazono at the University of Tokyo and his colleagues analyzed the brain activity of two macaque monkeys, measured using implanted electrodes. When the monkeys were awake, the activity in the back of their brain was integrated; under anesthesia, it became fragmented.48 To be sure, skeptics have pushed back, and the locus of consciousness remains an open question.49
One aspect of this question is how even to define the various structures. There are no perfectly sharp boundaries in nature. Where does the brain give way to the rest of the nervous system? What are the limits of the body? “Physics doesn’t really tell us what are the objects, the entities, in the world,” Tononi said. “It’s all a giant field of things; it’s very complicated. But it doesn’t really put borders in any fundamental sense. So how do we know where things end and begin?”
IIT offers a principled way to carve nature at its joints because it analyzes any network at all. Typically that’s the brain or a part of it, but you could equally well consider the extended network formed by the brain, body, and outside environment, or the miniature network of proteins within cells. Within each network, the theory will identify which sections are the most internally coherent, and by comparing the values of Φ at different scales, you can pinpoint which scale is relevant. Therefore, although everything is connected to everything else and boundaries are fuzzy, you can still delineate the sentient portion of your brain or any other system.
* * *
BOTH PREDICTIVE CODING and IIT imply that minds are everywhere. Any time you come across a structure that sustains itself in a chaotic environment or has a high degree of integration, you are looking at a thing with the potential for inner experience. Both theories thus embrace a version of panpsychism.50 To be sure, these theories are very different from traditional panpsychism, because they place consciousness firmly within physical science, not beyond it, and do not claim that consciousness is a new fundamental property of nature.51
Both theories also pull back from saying that absolutely everything can be conscious.52 Without such limits, panpsychism can get out of hand. Consider the multiple levels of organization in the nervous system. Each is a thing in its own right, with some degree of internal integration. Does that mean that our posterior cortex and both brain hemispheres and the areas within them and the neural circuits that make them up and all our billions of neurons—all the pieces of us that do not strictly have a Φ of 0—are individually conscious? If your entire cerebrum has a certain value of Φ, your cerebrum minus one neuron also has some value of Φ, as does your cerebrum minus two neurons, and so on. That’s an awful lot of minds in one head.
The idea that our heads are stuffed full of conjoined minds not only is unsettling, but also threatens to make IIT unfalsifiable. All those minds would have different experiences, so the theory could never explain why you have the experience you do. Your brain might allow you to see in color, but a subset of its neurons might have no concept of color. It would just be the luck of the draw that your consciousness arises from one subset rather than the other. “It would mean that your experience could be anything,” Tononi said.
To eliminate this ambiguity, he argued that there is only one physical system—the brain—so it can’t support an infinite tower of minds. If you consider all those possible structures that might conceivably be identified with consciousness and calculate their Φ values, one of them will come out on top, and that’s where consciousness resides. Two minds can’t share the same neurons at once; the mind that is more tightly integrated prevents the other from existing at all. “Only one entity can exist,” Tononi told me. “There cannot be any overlap whatsoever.”
That said, several consciousnesses might arise within the brain as long as they don’t overlap with one another.53 Tononi has speculated that when your mind drifts on a long drive, for instance, it may just be your “main mind” that drifts. A temporary minimind may form elsewhere in your neural circuits, literally an autopilot that remains conscious of the road and keeps control of the car.54 Other researchers, too, have suggested there could be multiple “islands of awareness” in the brain, some of which might be disconnected from all sensory input and left to think in utter isolation.55
The no-overlap rule not only stops minds from proliferating endlessly, but also provides a way to meld them together, thus solving an old difficulty for panpsychism known as the combination problem.56 If two independent minds with a certain value of Φ link up to form a single mind with a higher value, they will cease to exist as individuals and assimilate into the collective. Indeed, this happens to us every morning when we wake up: individual regions within the brain, which might have some modicum of consciousness on their own, fuse together, and gradually the brain as a whole regains consciousness. Something similar might well happen in ant colonies.57 For groups of humans, though, it’s unlikely. Even our most tight-knit group is less integrated than a single brain, so human consciousness resides at the individual level. A corporation is not a conscious person deserving of the same rights as actual people, whatever the US Supreme Court may say.58
TO DO IS TO BE
In the second half of the twentieth century, the dominant theory of the mind was functionalism, the idea that mind is as mind does.59 In one popular version, the mind is the brain’s software, its main function is to process information, and the details of the hardware don’t really matter; any two neural configurations that perform the same functions will be equally conscious. IIT, on the other hand, focuses on the unity of both experience and the corresponding brain activity, so it identifies consciousness with the structure of a system rather than with its function. “It’s not what it does,” Tononi said. “It is what it is.”
Consider two networks. They do exactly the same thing—the same inputs yield the same outputs—and thus are indistinguishable from the outside. But when you open them up, they look very different. One is a neatly arrayed feedforward network, in which input produces output and stops there. The other is a feedback network in which signals can loop around. (These are like the top left and bottom left networks in figure 3.2.) The first network is not integrated. Its later stages of processing depend on the earlier ones, but the earlier ones are independent of the later ones, so there’s no potential for collective behavior. IIT deems it unconscious. It is what philosophers call a zombie—a body without a mind.60 The second is highly integrated. Each component depends on every other component. The network is capable of new and unexpected behavior that might surprise even its designer. IIT regards it as sentient.
Feedback is the mechanism that produces the unity that IIT identifies with consciousness. Its importance is a recurring theme in AI, neuroscience, and physics. Hermann von Helmholtz, John Hopfield, and other pioneers thought that feedback was essential to the creation of a mind. Many theories of consciousness other than IIT, not least predictive coding, also stress it. Not only is feedback more interesting than a feedforward architecture, but also it has practical advantages. It is usually more flexible and allows each unit to perform multiple roles. “In order [for a feedforward system] to be functionally equivalent to a system that has feedback, you need many more units and connections,” Tononi said. Those efficiencies matter a lot to an organism competing to survive in a world of scarce resources and may help to explain the origin of consciousness in the history of life on Earth.61 “It suggests a reason why consciousness, if it is integrated information, might have evolved,” he told me.
The distinction between the function and the structure of a network also bears on the question of what would happen if you simulated your brain on a computer. Suppose you built a chatbot from this simulation. The machine would make the same jokes you do, recommend the same music, and send the same racy texts to your lover—so everyone would think it’s really you. But would there be any feeling behind its utterances, or would they be delivered purely mechanically? Would the machine be conscious? Science-fiction shows such as Black Mirror and Westworld—which, at the rate technology is advancing, won’t be fiction for much longer—assume that it would, but cognitive scientists and philosophers are divided. Tononi sides with those who say: not so fast.62
He told me his colleagues once designed a tiny programmable computer. It had just sixty-six electronic parts performing basic logic functions, as opposed to the billions in a modern microprocessor—small enough to easily analyze, but large enough to run a bare-bones neural-network simulation. Using this physical computer, they could compare what IIT says about the simulated network versus the original neural network.
Tellingly, the computer worked in a very different way from a neural network. It was not a highly parallel system in which information was flowing every which way at once. Instead, the computer took one thing at a time. It considered one neuron, then another, then calculated what those neurons would do if they were interacting—but there was never actually any such interaction. Most of the computer’s operation was feedforward, implying that it was barely conscious at all. Its modicum of experience was nothing like that of the original network, and it was determined by the hardware, not by the simulation code. “It has nothing to do with what it is simulating,” Tononi said. “It could be simulating an avalanche, a hurricane, a brain—it doesn’t matter.”
Almost all computers today, from your laptop to a supercomputer, are like that, too. They have a degree of parallelism, such as multiple processing cores, but it is nothing compared to that of a neural network. If such a computer were running a full simulation of your brain, it would act like you, but have the experience of being a computer. It wouldn’t matter whether the computer was simulating you or a toad; internally it would feel the same. So those who seek digital immortality by cloning their brains in a computer might want to think twice. Conversely, if you have the experience of being a person, you probably are one—you’re not trapped in some dystopian science-fiction brain-uploading scenario. Karl Friston, working with the philosopher Wanja Wiese of the University of Mainz in Germany, has explored this issue from the perspective of predictive coding and reached a similar conclusion.63
Tononi worries that we face a bait-and-switch situation as computers become ever more lifelike. “The majority of people these days would still say, ‘Oh, no, no, it’s just a machine,’ but they have just the wrong notion of a machine,” he said. “They are still stuck with cold things sitting on the table or doing clunky things.” But advanced AIs such as ChatGPT and DALL-E are already able to do things that seem to be coming from deeply felt experience, including creating poetry and art.64 Able to write convincingly on any topic and solve problems in chess, math, and other domains, these systems are starting to demonstrate a generalized intelligence like that of humans. If we judge one of these systems to be conscious, we may accord it the rights of a person. “When that happens—and it shows emotion in a way that makes you cry and quotes poetry and this and that—I think there will be a gigantic switch. Everybody is going to say, ‘For God’s sake, how can we turn that thing off?’”
But, he added, “if IIT is right, that is tragically wrong.” Just because machines seem conscious doesn’t mean they are. “They may be really imposters,” he said. “There is nobody there.… You need a theory about what consciousness is to have a proper answer.” Late in 2022, David Chalmers reviewed what IIT and other theories have to say about ChatGPT and concluded that it is probably not conscious.65 The philosopher Susan Schneider of Florida Atlantic University agreed that we need to be careful: “Artificial general intelligence may not be conscious, and that will mean we’ll have a case of sapience without sentience.” She has speculated that complex, human-level consciousness may be a transitory phenomenon. Having arisen at some point in our evolutionary lineage, it may die out if and when AIs supplant our species: “Consciousness may be a blip, a momentary flowering of experience before the universe reverts to mindlessness.”66
