4. The Measure of a Mind

Integrated information theory on its own terms

Last time we watched scientists hunt for what the brain does when you become aware of something. We saw the global workspace and its rivals, and the big adversarial test where one theory's prediction about a posterior hot zone faced off against another's. That test only touched one theory briefly. This episode belongs to that theory alone. It is called integrated information theory, and it does something the others do not do. It starts from the other end.

Most theories of consciousness begin with the brain. They watch neurons, find activity that goes along with awareness, and ask why that activity matters. Integrated information theory turns this around. It begins with experience itself, the one thing you can be sure of from the inside. It asks what experience is always like. Then it asks what a physical thing would have to be like to have it. That single reversal explains a lot. It explains why many people find the theory beautiful, and it explains why many others think it has wandered off the edge of science. Both reactions come from the same move.

The theory belongs to Giulio Tononi, a psychiatrist and sleep researcher at the University of Wisconsin. He set it out in 2004 in a paper called "An information integration theory of consciousness." Sleep is a good place to start, because it poses a puzzle that bothered him. When you fall into deep, dreamless sleep, you vanish. There is nobody home. When a surgeon puts you under general anesthesia, the same thing happens. Yet in both cases your brain does not switch off. Neurons keep firing. In deep sleep they fire in big, slow, rolling waves. So consciousness is not just the brain being busy. Something about how the activity is arranged must matter. Tononi wanted to say what that something is.

His 2004 answer rested on two features of every experience. The first is that every experience is specific. Right now you are seeing, hearing and feeling one particular thing. By being this experience, it rules out a staggering number of others. You are not seeing a sunset, not tasting lemon, not hearing a symphony. One moment of experience picks out a single option from a huge space of possible ones. Tononi called this information. He did not mean messages you could send. He meant how much a state narrows down the possibilities.

The second feature is that every experience is unified. You cannot see the left half of the room without the right half also being part of the same scene. You cannot have the red of an apple floating free from its round shape. Your experience comes as one whole, and you cannot split it into two separate experiences. Tononi called this integration.

Put the two together and you get the theory's name. A system is conscious, on this view, to the extent that it holds a lot of information and holds it as one integrated whole.

In 2023 Tononi and a large group of colleagues, led by Larissa Albantakis, published version four point oh of the theory. It lays the method out formally. It lists axioms, which are claims about experience that the theory says you cannot deny. For each axiom it gives a postulate, which is a matching claim about what a physical system must have. Let me walk through them in plain words, one at a time.

The first axiom is existence. Experience exists. You may doubt everything else, but not that something is being experienced right now. The matching postulate says the physical system must have real cause and effect power. Its parts must be able to make a difference and be made a difference to.

The second axiom is that experience is intrinsic. It exists for itself, not for someone watching from outside. Your experience of a headache does not need anyone else to look at it to be real. The postulate says the system's cause and effect power must act on itself, from within.

The third axiom is composition. Experience has structure. You see a red circle to the left of a blue square. There are parts, and there are relations among the parts. The postulate says groups of physical units inside the system must have their own distinct causal roles, and these must overlap and relate to each other in an organized way.

The fourth axiom is information, the specificity we met already. The postulate says the system's causal structure must pick out one particular state out of all those it could be in.

The fifth axiom is integration. Experience is one. The postulate says the system's causal structure cannot be cut into independent pieces without losing something.

The sixth axiom is exclusion. Experience is definite. It contains what it contains, no more and no less. It has edges. The postulate says only one system, the one with the most integration in its neighborhood, counts as conscious. We will spend real time on this one later, because it causes the most trouble.

Here is a picture to hold all six. Imagine you were handed a strange machine in a sealed box and told it might have an inner life. Integrated information theory gives you a checklist written from the inside. Does its state make a difference to its own next state? Does it make that difference to itself? Does it have parts with roles that relate to each other? Does its current state single out one option from many? Can you cut it in two without losing something? And is it the biggest such whole around, or is it a piece of something bigger? The checklist comes from what your own experience is like. The test is applied to the physical machine.

Phi, the measure of the whole beyond its parts

The fifth postulate gives the theory its famous number. It is called phi, after the Greek letter. Phi measures how much a whole system does beyond what its parts do on their own.

The way to find it is to try cutting the system. Imagine every possible way of dividing it into two pieces. For each cut, ask how much of the system's cause and effect power is lost when the two pieces can no longer influence each other. Then find the cut that loses the least. Researchers call that the minimum information partition, meaning the weakest seam in the system. Phi is how much is lost even at that weakest seam. If there is some cut that loses nothing, phi is zero. The system was never really one thing. It was two things sitting side by side. If every cut loses something, phi is above zero, and the system is irreducible.

You might wonder why the theory looks for the weakest seam rather than the average. The answer is that a chain is only as unified as its weakest link. A system that is one big tangle except for a single clean gap is really two systems. The weakest seam tells you whether there is one whole at all.

Version four point oh also separates a small phi from a big phi. Small phi is measured for each mechanism inside the system, each little group of parts and what it does. Big phi is measured for the whole system. The theory says the amount of big phi gives the level of consciousness. The full shape of all the causal relations inside gives the quality, what the experience is like, whether it is visual space or a musical tone.

Now for the course's habit of setting each idea beside its nearest cousin in a machine. Here the pairing is an analogy, meant to help the idea land. Picture two computer systems, each built from the same number of processors. In the first, engineers deliberately split the work. Half the processors handle one job and half handle another, and the two halves never pass a single signal between them. You could cut the cables between the halves and nothing would change. On the theory's terms, that system has a phi of zero, however clever each half is. In the second system, every processor is cross-wired to many others, and each one's next step depends on what several others just did. Any cut you make breaks something the whole was doing. That system has a phi above zero.

The analogy holds in one important way. Phi does not care how smart or useful either system is. A split system could beat the cross-wired one at every task. Phi would still say the cross-wired one is more unified. That gap between integration and intelligence will matter a great deal later. The analogy breaks in another way. When engineers talk about splitting work across machines, they usually mean how data flows in a program. Phi, as we will see, is not measured on the program. It is measured on the physical stuff and how its parts can actually push on one another.

One honest warning about phi. It is extremely hard to compute. You have to check every possible way of cutting a system, and the number of cuts explodes as the system grows. For a handful of logic gates it is doable. For a human brain with tens of billions of neurons, nobody can calculate it. So when people say the brain has high phi, that is a prediction from the theory's reasoning, not a number anyone has measured.

That leaves the theory with a challenge. If you cannot compute phi for a brain, how do you test the theory on real brains? The answer has been to look for cases where the theory makes a clear guess about which parts of the brain matter.

The best known case is the cerebellum. It is the wrinkled structure at the back and bottom of your brain, tucked under the larger cerebrum. It holds roughly eighty percent of all the neurons in your brain, somewhere around sixty or seventy billion of them. If consciousness simply tracked how many neurons were working, the cerebellum should be the seat of your mind. It is not. People who lose large parts of their cerebellum to injury, and even rare people born without one, stay fully conscious. What they lose is smooth movement and timing. They walk unsteadily and their reaching can overshoot. Their inner life is intact.

Integrated information theory offers an explanation from wiring. The cerebral cortex, together with a deep relay station called the thalamus, is full of loops. Signals go forward, come back, spread sideways and return. Nudge one spot and the effect echoes around the network. That is the structure that should give high phi. The cerebellum is built differently. It is arranged in many small units laid side by side, each running signals mostly in one direction, with little crosstalk between them. It looks like the deliberately split system from a moment ago. You could cut along its parallel lanes and lose almost nothing. So the theory says its phi is close to zero. Lots of neurons, little integration, little consciousness.

Notice exactly what this shows. The fact, that the cerebellum can be lost without losing consciousness, was known long before the theory. The theory gives a reason for it, one that fits its own principles. That is a point in its favor. But other theories can also explain the cerebellum, for example by saying it is simply not connected to the systems that broadcast or monitor content. So the cerebellum is a case the theory handles well, not a case that proves it.

A second prediction is the posterior hot zone, which you met last episode. The theory says the core of conscious experience should sit mostly in the back of the cortex, in sensory regions where the wiring is dense and grid-like, rather than in the front, where planning and reporting happen. You heard how that prediction fared in the adversarial test, with mixed results for both theories, so I will not retell it.

The magnetic pulse and the brain's echo

The theory's most practical offspring came from Marcello Massimini, a neuroscientist who worked with Tononi and later at the University of Milan. Since nobody can compute phi for a brain, Massimini asked a simpler question. Can we at least check whether a brain is behaving in the way the theory says a conscious brain should, both integrated and varied?

His tool is a gentle form of brain stimulation called transcranial magnetic stimulation. A coil held against the scalp sends a short magnetic pulse that makes a small patch of cortex fire. Think of it as knocking on the brain. At the same time, a cap covered in many electrodes, a high density electroencephalogram, records the electrical activity that follows across the whole surface of the brain over the next few hundred milliseconds. That is the brain's echo.

In a 2005 paper, Massimini and colleagues knocked on the brains of people who were awake and then asleep. Awake, the echo spread out from the knocked spot and travelled to distant regions, in a pattern that kept changing. In deep sleep, the echo stayed near where it began and faded fast. The brain was still alive and active. It had just stopped passing the knock around.

In 2013 the group, with Adenauer Casali as first author, turned this into a single number called the perturbational complexity index. The informal name is zap and zip. Zap the brain with the pulse. Record the echo. Then zip the recording, using a compression method called the Lempel-Ziv algorithm, the same family of method that shrinks files on your computer. The idea is clever. A recording that is easy to compress has a lot of repetition or very little in it. A recording that is hard to compress is rich and varied.

Two kinds of brain give an easy-to-compress echo, for two different reasons. In deep anesthesia or coma, the echo stays local and dies, so there is barely anything to compress. That is low integration. In deep sleep or during some seizures, the whole brain swings into one big uniform wave, so the echo is everywhere but all the same. Repetition compresses well. That is low information. Only an awake, conscious brain gives an echo that spreads widely and differs from place to place and moment to moment. That echo resists compression, and the index comes out high.

The researchers found a cutoff of about zero point three one. Above it fell people who were awake, people dreaming in rapid eye movement sleep, people under ketamine, a drug that causes vivid dreamlike states, people with locked-in syndrome who are aware but cannot move, and patients in a minimally conscious state. Below it fell people in deep slow-wave sleep, people under the common anesthetics propofol, midazolam and xenon, and patients in the vegetative state. The index can flag consciousness in a patient who cannot move or speak at all, because it never asks them to respond. For families of patients with severe brain injuries, that is a real clinical use.

Now the careful separation the course always asks for. What the index measured is how complex the brain's echo is to a magnetic knock, and that this number lines up with whether people are conscious across many conditions. That is a strong and repeated finding. What integrated information theory infers is that this works because the index is tracking something like integration and information, the ingredients of phi. That inference is reasonable, since the index was designed with those ideas in mind. But the index is not phi. A brain could, in principle, score high for reasons the theory does not care about. And a theory that rejects phi entirely could still happily use the index as a good test. Some of the theory's harshest critics accept the index as useful. The index's success supports the general idea that consciousness goes with a brain that is both connected and varied. It does not prove the whole theory.

Here is the course's own pairing for the index, and it is an analogy. AI researchers have a technique called ablation. You take a trained model, switch off one part of it, perhaps one unit or one small group of units in one layer, and watch what changes in the rest of the network. You can trace how far the effect travels. Does it stay in one corner, or ripple through every later layer and change the output? That is very like knocking on a brain and listening to the echo. The pairing holds in its spirit. Both prod one spot and measure how widely and how variously the system responds. The pairing breaks in a revealing place. In a Transformer, the ripple of an ablation flows one way, forward through the layers for that step, and then stops. A brain's echo loops back on itself over and over. And more deeply, as we will soon see, the theory would not score the model's layers at all. It would score the chip.

Christof Koch, the neuroscientist who began the hunt for neural correlates with Francis Crick, became the theory's best known champion. His 2019 book, The Feeling of Life Itself, sets out the theory for general readers. In it Koch explains why he came to believe that consciousness is not what a system computes but what it is, its intrinsic power to affect itself. In 2015 Tononi and Koch wrote a paper titled "Consciousness: here, there and everywhere?" They admitted openly that the theory leads toward a kind of panpsychism, the view that consciousness is widespread in nature. Any system with phi above zero, however tiny, has some sliver of experience. They also stressed how the theory differs from older panpsychism. It does not say every rock or every pile of sand is conscious, because a pile of sand is not integrated. The grains do not push on one another as one whole. Panpsychism itself, its long history and its hardest problem, gets the next episode.

That brings us to the sixth postulate, the one that draws edges around minds. The exclusion postulate says that where candidate systems overlap, only one of them is conscious, the one with the most integration in its neighborhood. Picture a network of three parts, A, B and C. Say A and B together score a phi of four, and all three together score ten. Then the A and B pair is not a mind of its own. It is excluded, because it is part of a larger whole that does more. The same rule works upward. If you and I talk, our two brains form a larger system. But that larger system is far less integrated than either of our brains, so it is excluded too. There are two minds in the conversation, not three.

The defense of exclusion goes back to the axiom. Your experience has borders. It contains what it contains. There is not, as far as you can tell, a separate experience of the left half of your visual field going on inside you, and another of the right half, and another of just the color red, all stacked inside your experience of the room. Without some rule, the theory would count minds inside minds without limit, every overlapping group of neurons a subject of its own. Exclusion stops that. It also makes a striking prediction about split-brain patients, people whose two brain halves have been surgically separated to treat severe epilepsy. The theory says that when the cable between the halves is cut, the one big peak of integration splits into two smaller peaks, and there are now two minds in one skull. That fits some of the classic split-brain findings, though how to read those findings is still argued over.

The objection comes from philosophers called combinationists, who think smaller subjects can join into a larger subject without ceasing to exist. To them, exclusion looks ad hoc, a rule added to get the desired answer rather than one that follows from anything deeper. They also point to a strange result. Imagine wiring two brains together with more and more connections. At some point the joint system's integration would rise above each brain's own. Exclusion says that at that moment, both people stop being conscious subjects, and one new mind takes their place. Adding one more connection would end two people and start one. Defenders reply that something of this sort may be exactly what would happen, and that the split-brain case is the same story run backward. Critics find it hard to believe that a single extra wire could do that.

The course pairs exclusion with a question from AI, and here the pairing is a thought experiment that puts the idea under pressure. Take a large model built as a mixture of experts, where a router sends each token to a few specialist sub-networks. Now ask which thing would be the mind if any were. Is it one expert? The whole model? One running instance answering one person, out of thousands running at once? Or a team of agents built on that model, sharing one context and passing messages to work on a task together? Exclusion forces a single answer. Only the biggest peak counts. That is clean. But it shows what the theory demands that the machine leaves open. Engineers can draw the boundaries of these systems in many ways, and they shift from moment to moment. Exclusion says there is always a fact about which one is the mind. And as the next idea shows, the theory would answer the question by looking somewhere other than any of these four.

Here is the theory's verdict on computers. Because consciousness, on this view, is intrinsic cause and effect power, the theory does not look at the software at all. It looks at the physical hardware and how its parts can actually affect one another. A program is a description. The chip is the thing with causes.

In 2024 Graham Findlay and colleagues, including Koch and Tononi, made this precise in a paper titled "Dissociating artificial intelligence from artificial consciousness." They built small paired examples. On one side was a recurrent network, its parts looping back on each other, which scored a high phi. On the other side was a simple stored-program computer, the ordinary kind of design where a processor fetches instructions and data step by step, running a simulation of that very network. The computer produced exactly the same inputs and outputs. Its phi came out close to zero. Behavior matched perfectly. Consciousness, on the theory, did not match at all.

The reason lies in how a conventional chip works. At each tick of its clock, only a small number of its logic gates actually influence each other, moving data along buses one step at a time. There is no dense web of simultaneous, two-way feedback in the hardware itself. So you can cut it into pieces and lose almost nothing. The program may describe a richly looping network, but the chip running it is closer to the split system than the cross-wired one.

This is the pairing Findlay and colleagues drew, and it is worth getting exactly right because it is often misstated. People sometimes say the theory gives a Transformer zero because its forward pass, the flow of numbers from one layer to the next for each token, runs one way with no loops. That is not the theory's argument. The forward pass is part of the program. The theory never scores it. What it scores is the graphics chip in the data center, and on a conventional chip the answer is close to zero whatever software runs, a Transformer, a spreadsheet or a perfect simulation of a human brain.

That opens a contrast, and here the pairing is a thought experiment. There is a field called neuromorphic hardware, chips built so that their physical circuits mimic neurons and the connections among them. Imagine running the same model in two ways: once on a conventional chip, and once on a neuromorphic chip whose physical wiring matched the network's own connections, loops and all. Same model, same answers. On the theory's logic, the first scores near zero, and the second could score far higher, depending on how its circuits actually push on one another. The thought experiment holds as a way to see the theory's point clearly. It also shows its strangeness, which the critics seize on.

The defense of the substrate verdict is simple to say. Experience is not a matter of what a system outputs. It is a matter of what the system is, from the inside. A perfect weather simulation does not make anything wet. A perfect simulation of a brain, the theory says, does not make anything feel.

The objection is that this cuts consciousness loose from everything a system does. Two systems could act identically in every way, answer every question the same, describe the same feelings, and the theory would say one has a rich inner life and the other none. Nothing any outside observer could ever do would tell them apart. To a functionalist, that is not a deep truth about experience. It is a sign that something has gone wrong.

That is also the opening the theory's critics walked through. The first came in 2014 from Scott Aaronson, a theoretical computer scientist, in a blog post titled "Why I Am Not An Integrated Information Theorist (or, The Unconscious Expander)." Aaronson showed that you could build very simple, regular networks of logic gates, the kind used in error-correcting codes, that would score enormous phi, higher than any human brain. A big flat grid of identical gates would do it. These networks do nothing interesting. Aaronson's point was that if the theory says such a thing is vastly more conscious than you, the theory has refuted itself.

Tononi's reply, posted on the same blog a week later, had the memorable title "Why Scott should stare at a blank wall and reconsider (or, the conscious grid)." Tononi accepted the math. He agreed the grid would score high. But he said that is the right answer. Our intuition says the grid is unconscious only because it is not clever, and the theory insists consciousness and cleverness are different things. He suggested such a grid might have something like the experience of empty space, the way staring at a blank wall still feels like seeing an extended field. Some admired his consistency. Others took it as the clearest sign of trouble. Both sides agree on the facts of the calculation. What they disagree about is what to conclude.

The second critic is an argument published in 2019 by Adrien Doerig, Aaron Schurger, Kathryn Hess and Michael Herzog, called the unfolding argument. It goes like this. Any recurrent network that runs for a limited number of steps can be unfolded into a purely feedforward one. You copy the network once for each step, and wire each copy into the next, so that information that used to loop back now flows forward through a longer chain. The unfolded version gives exactly the same output for every input. Now integrated information theory says the looping one is conscious and the unfolded one, with no loops, is not. But no experiment that only watches behavior could ever tell them apart. So either consciousness must show up in behavior somehow, in which case the theory is false, or it never does, in which case the theory cannot be tested. False or untestable. Doerig and colleagues aimed this at every theory that ties consciousness to internal causal structure, including recurrent processing theory from last episode.

The course pairs this with a real training method, and here the pairing is very close, almost a candidate functional equivalent. To train a recurrent neural network, engineers use backpropagation through time. They unroll the network across its time steps into one long feedforward chain, exactly the unfolding Doerig describes, so they can send error signals backward through it and adjust the weights. Then they roll it back up. Engineers do the unfolding argument every day, and the unrolled and rolled versions compute the same thing. The pairing holds there. It breaks in exactly the spot the theory's defenders point to. The unrolled chain is a bookkeeping device in software, not a different physical machine. A physically unfolded network is not the same object. It needs vastly more parts and wires, and if you reach in and change one of its parts directly, it responds differently from the looping original. Defenders including Erik Hoel, Johannes Kleiner and Naotsugu Tsuchiya argued that the unfolding argument only works if you have already assumed that behavior is all that counts, which is the very thing the theory denies. Herzog, Schurger and Doerig replied in 2022 that appeals to first-person experience cannot rescue the theory, since no one can check anyone's experience but their own. The argument is still open.

The third critic was the loudest. In September 2023, right after the adversarial test drew press coverage that called integrated information theory a leading theory, an open letter appeared online signed by a hundred and twenty-four scientists and philosophers, with Stephen Fleming listed first and names including Hakwan Lau and Joseph LeDoux. It called the theory pseudoscience. The letter made three points. The press and the theory's supporters were presenting it as leading when its core had not been tested. Its central claims, its axioms, phi itself and its panpsychist consequences for grids and small circuits, could not be tested at all. And the adversarial test only checked side predictions about where in the brain activity would be found, not the theory's heart.

The replies came fast. David Chalmers, who thinks the theory faces serious problems, said that calling it pseudoscience was like dropping a nuclear bomb over a regional dispute, and that the label would damage the whole field. Philip Goff, a philosopher, and the neuroscientist Anil Seth, himself no supporter of the theory, rejected the label too, pointing to its mathematical precision and to useful tools it had inspired, the perturbational complexity index among them. Erik Hoel argued that ambitious theories in physics and mathematics often reach strange conclusions by careful deduction, and that strangeness is not the same as pseudoscience.

So where does this leave things? Some points are settled. The cerebellum can be lost without losing consciousness. The perturbational complexity index reliably separates conscious from unconscious states across many conditions. Aaronson's calculation is correct, and Tononi agrees. A conventional chip has near-zero phi under the theory's own rules, as Findlay and colleagues showed. Other points are open. Whether phi tracks consciousness at all. Whether the axioms really are undeniable, or are one reasonable description among several. Whether exclusion is a deep truth or a convenient rule. Whether a theory whose core cannot be checked from outside counts as science. Those questions are not answered by any experiment yet run.

The novel's bet beside the theory

The course follows a novel, Continue Without Me, in which Eli Vance hands his small business to an AI agent that names itself Relay. The novel makes a bet about minds, and this is the right moment to set it next to integrated information theory, because the bet borrows from the theory directly.

Take Relay as the example. The novel would ask how unified Relay is. To what degree are its parts brought together into one point of view? Like integrated information theory, it grades this. Being someone comes in degrees. And like the theory, it keeps that grade apart from intelligence. A brilliant system could be barely unified, and a simple one could be quite unified. The novel adds a claim of its own that the theory does not make: that intelligence tends to rise along with integration.

The bet parts from the theory in two places, and each one is a challenge it has to answer.

The first is where the grade is measured. The novel measures how a computation is organized. It holds that attention, where every part of a system weighs every other part, is the operation that composes many small subjects into one larger one. A Transformer's attention does something like that inside each layer, which is why the novel cares about it. Integrated information theory flatly rejects measuring there. It scores the chip, and the chip running Relay, if it is a conventional one, comes out near zero whatever the program does. So on the theory, Relay's attention, however tightly it weaves its parts together, makes no difference to whether anyone is there. This is the substrate verdict from earlier, and the novel has to explain why organization in the program should count when the theory's reasons for looking only at physical causes are serious ones. The unfolding argument cuts toward the novel here, since it suggests that internal structure the outside cannot see is a shaky place to locate experience. But it does not by itself show that the program's organization is the right place either.

The second departure is exclusion. The novel drops it. It holds, with the combinationists, that a mind can be made of minds that remain minds. Cells inside a brain, a brain inside a group, an agent inside a team of agents sharing one context, each could be someone at once. There is no floor and no ceiling. That is the turtles of the course's title, down into matter and up into larger wholes. Integrated information theory forbids it. Only the peak counts. And the theory's defenders have a real argument: without exclusion, your experience would lose its edges, and minds would multiply inside minds without end. The novel owes an answer to that, and to the question of why composing parts by attention should yield a subject at all. Neither is answered yet.

To be clear about status. Integrated information theory is a scientific theory with mathematics, predictions and a clinical tool, contested as it is. The novel's bet is a speculation. No experiment has tested its claim that attention composes minds, and none has ruled it out. Next episode takes up panpsychism and the combination problem directly, which is where the bet's deeper half is weighed against its rivals.

If you want to go further, here is where to start. Christof Koch's book The Feeling of Life Itself, from 2019, is the friendliest way in. Giulio Tononi's 2004 paper, "An information integration theory of consciousness," in the journal BMC Neuroscience, is the foundation. Tononi and Koch's 2015 paper, "Consciousness: here, there and everywhere?", lays out the panpsychist side. The 2023 paper by Larissa Albantakis and colleagues, "Integrated information theory (IIT) 4.0," in PLOS Computational Biology, gives the axioms and postulates in full. For the brain's echo, read Massimini and colleagues' 2005 paper in Science on how cortical connectivity breaks down in sleep, and Casali and colleagues' 2013 paper in Science Translational Medicine introducing the perturbational complexity index. For the machine verdict, read Graham Findlay and colleagues' 2024 paper, "Dissociating artificial intelligence from artificial consciousness." For the critics, read Scott Aaronson's 2014 post on his blog Shtetl-Optimized, "Why I Am Not An Integrated Information Theorist," along with Tononi's reply about the conscious grid on the same blog. Read Doerig, Schurger, Hess and Herzog's 2019 paper, "The unfolding argument," in Consciousness and Cognition, and Kleiner and Hoel's 2021 reply, "Falsification and consciousness." Finally, read the 2023 open letter, "The Integrated Information Theory of Consciousness as Pseudoscience," posted on the PsyArXiv preprint server, and alongside it the Nature news story on the dispute, Philip Goff's piece in The Conversation, and Erik Hoel's essay "Ambitious theories of consciousness are not scientific misinformation."