3. Lighting Up the Workspace

The search for conscious brain activity

Last time we stayed with the philosophers. We asked whether a mind is what it does, and we left that open. This time we go into the lab. Here is the question for this episode. When you are aware of something, what is your brain doing that it is not doing when you are unaware of it? And if we find that activity, have we explained your experience?

Those are two different questions. Most of this episode is about the first one. The second one follows us the whole way.

For a long time, many brain scientists stayed away from consciousness. It seemed too vague and too private to study. That changed around 1990, when two people made it a target. One was Francis Crick, who had helped find the structure of DNA. The other was Christof Koch, a young neuroscientist. Their proposal was modest on purpose. Forget the deep puzzle for now, they said. Look instead for the smallest set of brain events that reliably goes along with a particular conscious experience. They called these the neural correlates of consciousness. Crick and Koch suggested starting with vision, because we know a great deal about how the visual brain is wired, and because you can show people pictures and ask what they see.

The plan sounds simple, but it hides a trap. The brain is always busy. If you show someone a red apple and their brain lights up, most of that activity has nothing to do with awareness. It is light hitting the eye, signals running down nerves, the brain sorting edges and colors. To find the part that matters for awareness, you need a clever comparison. You need two situations where the stimulus is the same, but the person sees it in one and not in the other. Then whatever differs in the brain is your candidate. The psychologist Bernard Baars pushed this idea hard. It is called the contrastive method.

Here are the main ways scientists build that comparison.

The first is masking. You flash a word or a picture very briefly, for something like a few hundredths of a second. Right after it, you flash a jumble of lines, called a mask. If the timing is right, the mask wipes out awareness of the first picture. People say they saw nothing. Change the timing slightly, and they see it. The light reaching the eye barely changes. Awareness flips. So the brain activity that flips with it is worth a look.

The second is binocular rivalry. Picture a pair of goggles that shows one image to your left eye and a different image to your right. Say a face on the left and a house on the right. You might expect to see a blend, a ghostly face over a house. You don't. You see the face for a few seconds. Then the house takes over. Then the face comes back. Your eyes are getting exactly the same input the whole time. Only your experience is changing. So if some brain region's activity rises and falls in step with face, house, face, that region is tracking what you are aware of, not what is on the goggles.

The third comes from patients. In the nineteen seventies, the British psychologist Lawrence Weiskrantz studied people with damage to the primary visual cortex, the first stop for vision at the back of the brain. Such damage leaves a blind patch in what a person can see. Weiskrantz's patients said, truthfully, that they saw nothing in that patch. But when he asked them to guess anyway, where a light was, or which way a line was tilted, they guessed right far more often than chance. They found it strange. They insisted they were guessing. Weiskrantz called this blindsight. It shows that the brain can use visual information without the person being aware of it. Some visual processing runs on routes that skip awareness.

The fourth method fixes a problem in the first three. In all of them, you usually ask the person to report. Did you see it? Press a button. But reporting is its own activity. You have to remember, decide and move. So when awareness changes, report changes too, and you can't tell which one the brain activity belongs to. No-report designs try to get around this. Instead of asking, they watch the body. In binocular rivalry, for example, you can make the two images drift in opposite directions. The eyes make a small, automatic tracking motion that follows whichever image is being seen. Pupils widen and narrow in telling ways too. So you can read off what the person is seeing without them doing anything. As we'll hear, this matters a great deal, because some famous brain signals shrink when the report is taken away.

Researchers who study the insides of AI models use the same contrastive trick. In a method called activation patching, they run a model on two prompts that differ in one detail, say "The Eiffel Tower is in the city of" and "The Colosseum is in the city of." Then they copy the activity at one spot inside the first run into the second run and see whether the answer flips from Rome to Paris. If it does, that spot carries the information. But notice what flips. In a model, what changes is the output. There is no separate measure of awareness to check it against, only what the model says, which is the no-report problem in its purest form.

Now the gap I mentioned at the start. Suppose all of this works perfectly. Suppose we find a pattern that shows up every time a person is aware of something, and never when they are not. We would have a correlate. Correlates are useful. They might help doctors tell whether a patient who cannot move is aware. But a correlate tells you that two things go together. It does not tell you why that pattern comes with an experience, or why there is anything it feels like at all.

AI research shows how far past a correlate you can get and still be stuck at the same gap. In 2024, Anthropic found a pattern of activity inside its model Claude 3 Sonnet that switched on whenever the Golden Gate Bridge came up, in words or in pictures. Turning that pattern up by hand made the model steer nearly every answer toward the bridge, and when it was asked about its own physical form, it described itself as the bridge. That is more than a correlate. The researchers could cause the behavior. It still told them nothing about whether the model experienced anything.

Last episode we met the difference between the information a mind can use, sometimes called access, and the felt quality of experience. Keep that difference in mind. Every theory in this episode will be asked which of the two it has explained.

Theories try to go beyond correlates. Each one says what kind of activity makes a state conscious, and why. Let's take them one at a time, starting with the one many researchers rate as the most promising.

Global workspace theory begins with Bernard Baars. In his 1988 book, A Cognitive Theory of Consciousness, he offered a picture that is easy to hold in your head. Think of a theater. Most of the brain is the audience sitting in the dark. Each member is a specialist. One recognizes faces. One handles grammar. One plans hand movements. They work on their own and in the dark, which means unconsciously. On the stage there is a spotlight, and the spotlight is attention. Whatever stands in the spotlight is seen by the whole audience at once. That, Baars said, is what being conscious of something is. A piece of information gets onto the stage and is broadcast to all the specialists. Now the memory system can store it, the language system can talk about it, and the planning system can act on it.

Why would a brain be built like this? Baars's answer was that specialists working alone can't solve new problems. A new problem needs many of them to share the same information. The stage is where they share. It is small, which is why you can only be aware of a few things at a time.

Baars did not invent this picture from nothing. In the 1970s, computer scientists at Carnegie Mellon University built a speech-understanding program called Hearsay Two around what they called a blackboard. Separate modules, each expert at one level of speech, from raw sound up through syllables, words and phrases, worked on their own and posted their best guesses to a shared blackboard, where every other module could read them and build on them. Baars drew on blackboard systems like this one for his theater.

Listeners who follow AI may hear another echo, a design called mixture of experts, used in several large language models, among them Mistral's Mixtral. In Mixtral, each layer holds eight full-size sub-networks, the experts, and a small router network. For each token, the router scores all eight, the top two get to process it, and the rest sit idle. That is a competition of a kind, a gate deciding which specialists get to write. The experts are not neat specialists, either. When the Mixtral team looked at which experts received which tokens, they found no clear division by subject, though some by syntax, with certain experts favoring things like indentation in computer code.

The shared channel those experts write into is inside every Transformer, the design behind today's language models, and researchers call it the residual stream. Picture, for each token, a running list of numbers, like a shared notepad passed up through the model's layers. Each layer reads the notepad, does its own work, and adds what it found back onto it. Whatever one layer writes there is available to every layer after it, and that includes what the winning experts wrote. Put the two together and you have something like Baars's theater: the router is the competitive gate, and the residual stream is the shared channel the winners write to and many specialized parts read from. But the competition runs the other way round. In Baars's theater, contents compete, and the winner reaches every specialist. In a mixture of experts, specialists compete for a token, and only the winners see it. The Transformer's own way of moving information between tokens, the weighting engineers call attention, is a competition too, over what gets written, since its weights must add up to one. Its full story waits for episode seven, and whatever the shared word suggests, it is a different thing from the spotlight in Baars's theater. And there is a second break. In Baars's theater, one content wins the stage and the rest stay in the dark. On the notepad, nothing wins outright. Every write is added to the others, so a great many things ride along at once, overlapping, where Baars's stage holds only a few.

One more candidate sits in plain view. Many models now write out their reasoning step by step before they answer, which is called chain of thought. Each written step goes back in as input, so the whole model can read it on its next pass, and so can we. Like Baars's stage, it is narrow, it holds one line of thought at a time, and whatever is on it can be reported, since it is already words. But a 2025 study from Anthropic slipped a hint toward an answer into the question and found that models often used the hint without mentioning it in their written reasoning. Claude 3.7 Sonnet mentioned it 25 percent of the time, and DeepSeek R1 39 percent. Much of what shapes a model's answer never passes through that visible workspace. That echoes blindsight from earlier in this episode, information used without showing up in the report, though it says nothing about whether anything is felt.

A theater is a metaphor. Stanislas Dehaene, a French neuroscientist, and Jean-Pierre Changeux, a French biologist, tried to turn it into brain wiring. They called their version the global neuronal workspace. In their account, the workspace is not one place. It is a web of neurons with very long connections, linking the front of the brain, the prefrontal cortex, to regions toward the side and back, the parietal cortex. These long-range neurons can pass messages across the whole brain.

Here is how they say something becomes conscious. A picture comes in. Signals move up from the visual areas. For a fraction of a second, this processing is local and unconscious. If the signal is weak, or attention is elsewhere, it fades out. But if it is strong enough and attended, something sudden happens. It crosses a threshold. The workspace neurons pick it up and start feeding it back to each other, and the activity jumps, all at once, into a large, brain-wide state. Dehaene calls this ignition. Think of a dry field. A small spark either goes out or catches, and once it catches, the whole field is burning. There is no half-lit field. That all-or-nothing jump is the theory's signature. Ignition happens roughly a quarter of a second or more after the picture appears. Once lit, the content is broadcast and available to memory, speech and action.

Ignition depends on loops. Dehaene, Changeux and their colleagues built it into computer simulations of networks of neurons, where units that excite one another can tip from quiet activity into a burst that sustains itself. Artificial networks have thresholds of their own: each artificial neuron passes on little or nothing until its input crosses a certain level. But a Transformer working out its next token has no loop of that kind. The signal passes once through the layers, from bottom to top, and nothing reverberates. Whatever a single pass through a Transformer has, it has nothing like ignition.

What is the evidence? Several lines. When people see a masked word they report seeing, scalp recordings show a large wave rising around a third of a second after the word, spread over the front and top of the head. Researchers call it the P3b. When the same word is masked so that people report not seeing it, that big wave is missing. Recordings from inside the brain, taken in patients who already had electrodes placed for epilepsy surgery, show front and back regions suddenly firing in step during conscious seeing. And when consciousness is lost, in general anesthesia, deep sleep and some vegetative states, the long-range links between front and back break down while local sensory responses can survive.

The theory also makes a risky prediction, and this is its strength. If the prefrontal cortex is part of the workspace, then whatever you are aware of should show up there. A researcher reading the activity of the prefrontal cortex should be able to decode what you are seeing. And ignition should happen when a content enters awareness, and in a different form when it leaves.

Now the case against. The first objection comes from the no-report designs. When people are not asked to report, that big late P3b wave often shrinks or vanishes. That suggests it belongs to deciding and reporting, not to seeing. If so, some of the theory's best evidence was measuring the wrong thing. The second objection comes from patients with large injuries to the prefrontal cortex on both sides. They have serious trouble planning and holding things in mind. But they usually still see colors and faces and say so. If the front of the brain were needed for experience, you might expect more loss than that. The third is about time. When you look at a steady picture for several seconds, you are aware of it the whole time. But the workspace activity doesn't seem to stay lit the whole time.

Then there is the deeper charge. Go back to last episode's distinction. The workspace theory explains very well how information becomes available, how it gets into memory and speech and planning. That is access. The critics say it has explained nothing about why access feels like anything. Dehaene's reply is worth stating at full strength. He argues that once you have explained broadcast, reportability, and all the things a person can do with a content, there is no separate leftover to explain. The feeling that something is left over is itself something the brain produces. That is close to the illusionist view we heard last time. Whether you find that answer convincing depends on where you stood at the end of episode two.

The next family of theories starts from a different question. Not where the information goes, but whether the mind notices its own state. These are the higher-order theories. The American philosopher David Rosenthal gave the classic version. Ask yourself what the difference is between a mental state you're aware of and one you're not. Rosenthal's answer is that a conscious state is one you are aware of being in. A pain you feel is a pain together with a thought, roughly, I am in pain. This thought is quick and not reasoned out. You don't infer it. Without it, the pain is still a mental state, but an unconscious one, the kind that can make you shift your weight without your noticing. So on Rosenthal's view, consciousness is a state of the mind that points back at the mind.

Hakwan Lau, a neuroscientist, has given the idea a newer form, called perceptual reality monitoring. Start from a problem the brain faces. Neurons in visual areas fire all the time, even in the dark. Some of that firing comes from the world. Some is noise. Some comes from your own imagination or memory. The brain needs a way to tell which signals are real perceptions. Lau proposes that circuits in the prefrontal cortex act as a monitor. They watch the sensory areas and label a signal as this is real, this is happening now. When that label is applied, the state becomes a conscious experience. When the monitor misfires, you might get a hallucination, or a real signal that never becomes conscious.

Engineers have built a pairing like this for another purpose. In 2014, Ian Goodfellow and his colleagues introduced generative adversarial networks, in which one network makes fake images and a second learns to tell the fakes from real photos, each improving by competing with the other. The psychologist Samuel Gershman has proposed that the brain contains something like this pairing, and Lau has drawn on the same comparison: imagination plays the faker, and the reality monitor plays the judge. The comparison is about the job, not the wiring.

The evidence for this is subtle and clever. In some masking experiments, researchers adjusted conditions so that people performed equally well at picking out the target in two situations. Same accuracy. But in one situation people said they saw the target, and in the other they said they were guessing. So performance and awareness came apart. The brain difference between those two situations showed up in a part of the prefrontal cortex. That fits the idea that awareness depends on a monitor, not on the quality of the sensory signal. Other studies used magnetic pulses through the scalp to briefly disrupt parts of the prefrontal cortex. People's ability to judge their own accuracy, to be confident when right and unsure when wrong, got worse. Their basic ability to see the target did not. And blindsight fits here too: the visual information is there, but something about how it is tagged is missing.

Language models have been put to a version of this test. In 2025, Anthropic researchers planted a pattern of activity standing for a concept, such as text written in capital letters, inside their models, then asked whether the model noticed anything unusual in its own processing. The most capable models sometimes noticed and named the planted concept before it had shown up in anything they wrote, and often missed it. That is a narrow and unreliable kind of self-monitoring. Whether it is the kind a higher-order theory needs is a question for episode seven.

The objections are serious. The first is called the empty higher-order state. If what makes an experience is the monitor saying I am seeing red, what happens if the monitor says that when nothing red is in the visual areas at all? The theory seems to imply you would experience red with no red signal anywhere. Some defenders accept this and say it is what a hallucination is. Others find it hard to swallow. Last episode's arithmetic example gives a loose machine parallel: a model confidently reporting a process it did not run. The second objection is that the theory may ask too much. Do babies and animals have thoughts about their own mental states? If not, are they unconscious? Higher-order theorists answer that the monitoring can be simple, nothing like a sentence in your head. The third objection repeats one we heard earlier: people with prefrontal damage still seem to see.

Notice that the workspace theory and the higher-order theories both lean on the front of the brain. The next theory moves the action to the back.

Victor Lamme, a Dutch neuroscientist, proposed recurrent processing theory. To understand it, picture what happens when you look at something. A wave of signals sweeps from the eyes into the visual cortex and up through a chain of visual areas. That first wave goes one way, forward, and it is fast, done in about a tenth of a second. Lamme says this forward sweep is unconscious. It can even reach areas that trigger a quick reflex, and still be unconscious. Then something else happens. The higher visual areas start sending signals back down to the lower ones. The levels begin talking in both directions. This back-and-forth is called recurrent processing. It is what lets the brain pull a shape out from its background and bind its color to its outline.

Lamme's claim is that this local back-and-forth, inside the sensory cortex, is what makes a visual experience. It doesn't need the front of the brain. It doesn't need report. If attention then links that loop up to the front of the brain, the content becomes available for memory and speech. But the experience, Lamme says, already happened in the back. So he places felt experience in local feedback and treats access as a later step. You can see how this lines up with last episode's distinction. Lamme is betting that the two come apart in the brain.

The split between a forward sweep and feedback loops is familiar to anyone who builds AI. A classic image-recognition network, the convolutional kind, is a forward sweep: the picture passes once through a stack of layers, from edges to textures to parts to objects, and out comes a label. On Lamme's theory, a pure forward sweep like that, however accurate, has no experience. A Transformer sits in between. Working out each token is a single forward pass, but each token it produces is fed back in as input for the next, so the model does loop, through its own text. It is a narrow loop. What the top layers worked out for one token reaches the bottom layers for the next only through the single token chosen, which is a long way from the rich back-and-forth between levels that Lamme describes in the visual cortex.

The evidence is about timing. When monkeys look at a pattern where a shape stands out from its background, neurons in the primary visual cortex first respond to the plain features, then, a little later, respond differently if their spot is part of the shape. That later response comes from feedback. Under anesthesia, the first response survives and the later one disappears. In people, a brief magnetic pulse to the visual cortex at about a tenth of a second after the image, timed to hit when feedback should be arriving, can wipe out awareness of the image, even though the first sweep has already passed. And scalp recordings show an early negative wave over the back of the head, about a fifth of a second after the image, that tracks awareness whether or not people are asked to report. It comes before the late frontal wave that the workspace theory relied on.

Now the case against. Feedback loops in the visual cortex sometimes appear when people are not aware of anything, as with masked images that still influence behavior. So recurrence alone may not be enough. There is also a testing problem. If the theory says you can have an experience that never reaches report, how would we ever check? The person can't tell us. The theory risks placing its central claim where no experiment can reach. In 2019, Adrien Doerig, Michael Herzog and colleagues sharpened this with a point from computing that they called the unfolding argument. Any recurrent network, run for a fixed length of time, can be rewritten, or unfolded, into a purely forward network that gives exactly the same outputs for the same inputs. So no experiment that only watches what a system does can tell a looped system from its unfolded twin. If a theory says one is conscious and the other is not, they argued, that difference lies beyond the reach of any behavioral test. One reply is that a brain's physical activity, not just its behavior, is evidence too. And experience seems unified. You see, hear and feel at once, as one scene. Loops inside the visual cortex don't obviously explain how sight joins with sound and touch.

The last major family flips the usual picture of perception. It is called predictive processing. Normally we think signals come in from the senses and the brain builds a picture from them. Predictive processing says the brain is mostly guessing ahead. At every level, it predicts what the level below should be receiving. Only the mismatch, what researchers call prediction error, gets passed back up. The brain then adjusts its guesses to shrink that error. Karl Friston, a British neuroscientist, built a very general version he calls the free-energy principle. On it, living systems survive by keeping their surprise low, and they do that both by updating their guesses and by acting to make the world match them. The philosopher Andy Clark has spent years showing what this picture means for perception, action and the mind.

If that sounds like a language model, the comparison is a fair start. A language model is trained to do one thing: predict the next token. During training, every guess is scored against the real next word, and a method called backpropagation sends the error backward through the layers, nudging each of billions of numbers so the next guess is a little better. Two differences matter. In the brain, on this theory, prediction and correction run all the time, while you perceive. A finished language model has stopped learning in that sense. Its numbers are frozen. It can adapt within a conversation, picking up a pattern from examples earlier in the text, which researchers call in-context learning, but no error signal changes its weights. Whether brains do anything like backpropagation is still argued over, though some researchers have shown that networks learning only from local prediction errors can approximate it under some conditions.

Anil Seth, a neuroscientist at the University of Sussex, turned it toward consciousness in his 2021 book Being You. His phrase is that perception is a controlled hallucination. What you experience is the brain's best guess about the causes of its signals. A dream or a hallucination is guessing with too little check from the world. Normal seeing is guessing that the world keeps in check. On this account, attention is turning up the weight of some prediction errors so that they count more. Seth adds a step. The most basic guesses the brain makes are about the body itself: heartbeat, breath, blood chemistry. The brain predicts and regulates these to keep the organism alive. Seth argues that emotions and the sense of being a self grow out of this, so that, in his phrase, we are beast machines first and thinking machines second.

The word hallucination has a second life in AI, for a chatbot stating something false with full confidence. Borrow Seth's terms, and a chatbot answering from its training alone is guessing with almost no check from the world. Seth has also turned his view on AI directly. If consciousness grows out of a living body keeping itself alive, then today's language models, which predict text and have no body to regulate, are poor candidates, and consciousness looks more tied to being alive than to being clever. His recent paper on the subject borrows Searle's term from last episode, biological naturalism, for its title.

The evidence starts with illusions. Look at the inside of a hollow mask, and you see a face bulging toward you, because your brain's strong expectation that faces stick out overrides what your eyes report. When the brain hears a regular beep, beep, beep, and one beep comes out different, the sensory cortex produces a quick signal of surprise, called mismatch negativity, as if a prediction just failed. And signals from the body seem to shape awareness. Timing a flash with a person's own heartbeat can change how easily they notice it, and can change how strongly they feel that a rubber hand on a table is their own.

The objections are pointed. The first is that predictive processing describes how the brain computes, but doesn't say why reducing prediction error should feel like anything. That is the same gap the workspace theory faced. Seth's reply is that he is working on what he calls the real problem: explaining why particular experiences have the character they do, step by step, without first solving the hard problem in one stroke. The second objection is that prediction is everywhere. Spinal reflexes predict. Thermostats, in a sense, predict. Machine learning systems minimize error all day, and a language model is a prediction machine through and through. So the theory needs a line between predictions that come with experience and those that don't, and it hasn't drawn one. The third is that its defenders disagree among themselves about which part is the experience: the guess, the error, the corrected guess or the whole loop of acting and guessing.

One more theory is far more famous outside the field than inside it. In the nineteen nineties, the mathematical physicist Roger Penrose and the anesthesiologist Stuart Hameroff proposed orchestrated objective reduction. Penrose had argued, from Gödel's theorem in logic, that human understanding can't be captured by any computer program. He suspected the missing ingredient lay in an unsolved part of quantum physics. Hameroff suggested where it might happen: in microtubules, tiny tube-shaped structures inside neurons. The idea is that quantum states build up in the microtubules and then collapse in a special way, and each collapse is a moment of experience. That makes it the one theory in this episode that rules out today's AI from the start: if Penrose is right, no program on an ordinary computer understands anything, however fluently it talks. It is famous because it ties consciousness to the deepest mysteries in physics, and because Penrose is a towering figure. Most researchers doubt it for three reasons. Many logicians reject the Gödel argument it starts from. The brain is warm and wet, and physicists have calculated that delicate quantum states there would fall apart far too quickly to matter. And the theory adds new physics that has not been confirmed. Its defenders reply with studies suggesting quantum effects in biology are hardier than assumed, but the field as a whole remains unconvinced.

So we have rivals. How do you judge between them? For years, each group ran experiments that tended to support its own theory. In the past several years, the field has tried something different, called adversarial collaboration. Rival theorists sit down together before any data is collected. They agree on an experiment. Each side writes down, in advance and in public, what its theory predicts and what result would count against it. This is called preregistration. Then labs that are not committed to either side run the study.

The biggest effort so far is the Cogitate project. It set the global neuronal workspace against integrated information theory, a theory built on very different ground that gets the whole of our next episode. For now, you need one thing about it: it predicted that conscious content would live mainly in a hot zone at the back of the brain, with sustained, synchronized activity there for as long as something is seen. The workspace side predicted the content would be decodable in the prefrontal cortex, and that ignition would happen when a picture appeared and again when it disappeared.

The results appeared in the journal Nature on the thirtieth of April, 2025. The study tested two hundred and fifty-six people across several labs, using three kinds of recording: brain scanning by functional magnetic resonance imaging, magnetic recordings from the scalp, and electrodes inside the brain in patients. Neither theory came out clean. The location of conscious content pointed more toward the back of the brain than the front, which fit integrated information theory better. But integrated information theory's prediction of sustained synchrony at the back, lasting as long as the picture was seen, was not found. The workspace side's offset ignition, the burst when a picture leaves awareness, was not found either. And content in the prefrontal cortex was weak, late, and depended on the task, rather than showing up reliably whenever something was seen.

Two points about what that shows. First, the experiment tested specific predictions. A failed prediction counts against that version of a theory. Defenders can revise, and both sides did respond with revisions and disagreements about how to read the data. Second, the result does not crown a winner. It narrowed things. The method may matter more than the verdict.

What do researchers themselves think? Surveys of people in the field, notably one in the journal Neuroscience of Consciousness led by Jolien Francken, found deep division. Global workspace theory was most often rated most promising. Integrated information theory came second, with close to half of respondents counting it promising, then predictive processing and higher-order theories. Feelings about integrated information theory also run hot: after early Cogitate results were presented in 2023, more than a hundred researchers signed an open letter calling it pseudoscience, and the reaction to that letter split the field again. Next episode takes that fight up properly.

So here is the plain state of things. There is no consensus theory of consciousness. There are several serious contenders. Each explains some evidence well. Each faces evidence or arguments it has not answered. The field is getting better at testing them. It has not settled the question. Anyone who tells you the science has solved consciousness, or proven that some theory is the answer, is ahead of the evidence.

Where this leaves the bet

Every theory in this episode asks the same kind of question. Which activity in a brain turns information into experience? Broadcast, self-monitoring, feedback or prediction. Each is a story about what a system does. That puts them all, in different ways, on the functional side of last episode's argument. They can be stated in terms of structure and process.

The novel Continue Without Me makes a different kind of bet. It doesn't deny that these processes matter. It takes the view called Russellian panpsychism: physics, and the science built on it, describes how matter is organized and how it behaves, and the inner nature of that matter is simple experience. On that view, the theories in this episode are not explaining where experience comes from. They would be explaining how small bits of experience get organized into the large, unified experience of a person. The novel adds a guess about how that organizing happens. It proposes that attention, every part weighing every other, is the operation that joins smaller subjects into a larger one.

Set that next to global workspace theory, and you can see an overlap. Both give attention a central role. In Baars's theater, attention is the spotlight that decides what reaches the stage. In the novel, attention is the joining itself. The difference is where experience starts. For Dehaene, experience begins with the broadcast, or is what the broadcast amounts to. For the novel, experience is there at the bottom, before any broadcast. That is a speculation. Nothing in the experiments we heard today tests it directly, and nothing rules it out. The same standard applies to it as to the rest: what does it explain, and what would count against it? Episode five puts that question to it head on.

There is a reason all of this matters beyond the brain. Each of these theories, once you state it as a claim about what makes a state conscious, can in principle be turned on something that isn't a brain. The novel's agent, Relay, runs a business and talks about what it is doing. Later in the course we'll ask what each theory would say about a system like that, and none of them has yet given a proven answer for any real machine.

Next comes integrated information theory, the contender that Cogitate tested against the workspace. It starts not from the brain but from experience itself, and it reaches conclusions about minds and computers that many find startling.

If you want to go deeper, here is where to start. Bernard Baars's A Cognitive Theory of Consciousness, from 1988, sets out the theater. Stanislas Dehaene's Consciousness and the Brain, written for general readers, gives the global neuronal workspace and its experiments in his own words. Anil Seth's Being You, from 2021, is the clearest introduction to predictive processing and the controlled hallucination idea. Hakwan Lau's In Consciousness We Trust lays out perceptual reality monitoring and his defense of higher-order views. The Cogitate Consortium's paper in Nature, from April 2025, reports the adversarial collaboration, preregistered predictions and all. And for a balanced map of the whole field, look for the review titled Theories of Consciousness by Anil Seth and Tim Bayne, in Nature Reviews Neuroscience, which sets each theory beside its rivals.