168 | Anil Seth on Emergence, Information, and Consciousness
What this covers
Anil Seth and Sean Carroll explore how consciousness can be understood as a physical phenomenon without invoking magic or mystery. The conversation centers on Seth's framework for treating consciousness as weakly emergent—a property of complex systems that cannot be reduced to its parts in simple ways, yet remains fully compatible with the laws of physics. Rather than resort to strong emergence, which would require new physical laws to appear at large scales, Seth proposes using information-theoretic tools like Granger causality and transfer entropy to operationalize what emergence actually means and measure it in data. The underlying claim is that consciousness is fundamentally a Bayesian, prediction-driven process anchored in the body's need to regulate itself—not a passive reception of sensory information but an active inference about the world.
The discussion moves across several distinct territories. Seth and Carroll clarify what dynamical independence is and how it lets us discover emergent structures empirically rather than assume them philosophically. They examine perception through the lens of predictive processing, where sensory input functions as prediction error rather than raw material for conscious experience. A substantial thread questions substrate independence—the assumption that consciousness could run on any physical medium—by noting that biological systems blur the line between substrate and process in ways computers do not. The conversation also addresses live debates: whether weakly emergent properties actually vanish during unconsciousness, how memory's reconstructive nature bears on the reliability of experience, and why panpsychism and idealism sidestep rather than solve the puzzle of consciousness. Throughout, Seth and Carroll maintain that exhausting what complicated neural systems can accomplish under rich interpretation is methodologically prior to concluding that consciousness transcends physics.
Seth argues that consciousness should be approached as a weakly emergent phenomenon fully compatible with physicalism—measurable via information-theoretic tools like dynamical independence—and that conscious experience is fundamentally a Bayesian, body-grounded prediction process rather than a passive readout of sensory data.
- Weak emergence (whole more than parts, but causally closed) can be operationalized quantitatively via Granger causality / transfer entropy, dissolving the need for 'magic' strong emergence.
- Perception is the brain's best-guess Bayesian inference on the causes of sensory signals, with sensory input acting only as prediction error.
- The brain's predictive machinery is grounded in regulating bodily physiology, making consciousness deeply tied to embodiment and casting doubt on substrate independence.
Perception is the brain's top-down predictive inference about sensory causes, not bottom-up feature reading.
- Perception is the result of unconscious Bayesian inference: sensory signals arrive without labels and the brain, in the dark and quiet of the skull, continually infers the most likely causes of those signals; the content of perceptual experience is the brain's best guess (the posterior), combining sensory data with prior expectations derived from evolution, development, and recent experience.
“everything that we can perceive is constrained by the generative models that are encoded in our brains. But this generative models can change and develop over time, and we can therefore learn to perceive new things through experience.”
- What we perceive is the content of the brain's top-down perceptual predictions, not a bottom-up readout of sensory features; sensory signals flowing into the brain are really prediction errors—the difference between what the brain expects and what it gets at every level of processing—that serve only to calibrate and rein in the top-down predictions, a process that approximates Bayesian inference.
“what we actually perceive as the stuff going in the other direction, the top-down predictions that are being reined in by the sensory prediction errors from the world. And that process approximates Bayesian inference.”
- We can perceive novel things despite living in a world of the expected because perception is deeply hierarchical: high-level perceptual contents (faces, objects) are built from lower-level features (brightness variations, lines, shapes) shared with familiar things, so encountering something new requires only a small leap at the top of the hierarchy, sometimes accompanied by a felt sense of surprise.
“even if you see something that you haven't seen before... it's still going to share a lot of the same lower-level features with other things that your perceptual system is very used to making best guesses about... it's only at the last bit that you have to make a little leap, and see something new.”
- The brain encodes a generative model of the causes of sensory signals that supplies the predictions compared against prediction errors, so everything we can perceive is constrained by these models; crucially, generative models develop with experience, which is how we learn to perceive new distinctions—e.g., learning to taste differences in red wine that were previously indistinguishable.
“everything that we can perceive is constrained by the generative models that are encoded in our brains. But this generative models can change and develop over time, and we can therefore learn to perceive new things through experience.”
Weak emergence preserves the intuition that wholes exceed their parts while remaining fully physically explicable.
- Nominal emergence is a property that by definition can only apply to a whole and never to a part (e.g., a circle is nominally emergent from its constituent points); combining nominal emergence with a sufficiently rich version of weak emergence gives you everything you need without violating the causal closure of the physical world.
“a circle is nominally emergent from the set of points that make it up. There's nothing mysterious going on here, it's just that a circle is not the kind of property that can ever be attributed to a point”
- Emergence in the dynamical-independence sense is best understood as a way to enrich our descriptions of physical systems in ways that might be relevant to consciousness—since conscious experiences are unified, global, and seem more than the sum of their parts—rather than as a proof that consciousness is an emergent property; it is an empirical question whether weakly emergent properties dissolve when people lose consciousness under anesthesia or sleep.
“I see emergence in this sense as a way to enrich our descriptions of physical systems that might have relevance to consciousness... Do weakly emergent properties dissolve in those cases or not? It's an empirical question.”
- In the Barnett-Seth approach, a candidate weakly emergent macroscopic variable must be dynamically independent from its microscopic underpinnings, meaning that knowing what is going on at the microscopic level does not help you predict the macroscopic variable's evolution; this enables data-driven discovery of candidate emergent coarse-grainings even when the emergent property does not visibly leap out to an observer.
“a candidate weakly emergent variable must be what... We call dynamically independent from its microscopic underpinnings, which just, again, this just means that knowing what's going on at the microscopic level does not help you predict what's going on at the macroscopic level.”
- Seth's original 2011-era approach operationalized weak emergence by asking whether the macroscopic whole (e.g., a flock) predicts its own future behavior better than the microscopic parts alone can, with such self-causality—conditioned on the parts—operationalizing the idea that the flock has a life of its own, but only relative to an observer with imperfect knowledge of the system, so nothing spooky is implied.
“does the flock, as a whole, predict its own future behavior better than I can do from just the birds alone? Is there some self-causality, self-information for the flock, conditioned on the parts that make it up?”
- Weak emergence preserves the intuition that the whole is more than the sum of its parts while remaining physically explicable; following Mark Bedau, a weakly emergent property is one for which an explanation exists in terms of microscopic components but is incompressible—you can only determine the global property by exhaustively simulating the microscopic interactions, with no shortcut.
“Mark Bedau, who describes weak emergence as something for which there is an explanation, a macroscopic property for which there is an explanation in terms of microscopic components, but it's what he calls incompressible. You can only figure out what the global property is by simulating exhaustively the microscopic interactions”
- Strong emergence—the idea that a macroscopic property is in principle not reducible to its microscopic components and exerts downward causal power on them—is uncomfortably close to magic, does not fit a physicalist picture, and is suspicious because the only example that reliably comes up to motivate it is consciousness itself, creating a reciprocal mystery.
“It's uncomfortably close to magic to talk about emergence this way... one of the only examples that reliably comes up is consciousness.”
- The fact that something is described at a higher level of organization does not mean it doesn't exist; higher-level ontologies are real (Dennett's 'real patterns') and can be essential for understanding a system, without any disruption to physical causality running all the way down to fundamental reality.
“Daniel Dennett talks about real patterns. The fact that something is described at a higher level doesn't mean that it doesn't exist. It just means that a higher level of description can be very, very useful”
Consciousness fundamentally evolved to regulate internal bodily states through interoceptive prediction and control.
- The fundamental evolutionary purpose of having a brain is to keep the body alive, so the brain is in the business of inferring and controlling its own inaccessible internal bodily state (heart rate, blood pressure) as much as perceiving the outside world; interoceptive predictions are geared toward control rather than identifying where organs are, which is why we experience hunger, thirst, or pain rather than the shapes or locations of our organs.
“The purpose of having a brain in the first place is to keep the body alive. That's the fundamental evolutionary duty of a brain... the internal state of the body is also inaccessible and remote and has to be inferred.”
- The larger claim is that everything we experience—even perception of the external world—is ultimately grounded in the predictive mechanisms that evolved to regulate bodily physiology, implying a deep connection between consciousness and life; this does not mean you must be alive to be conscious or that all living things are conscious, but that bodily-regulatory prediction is the route to understanding how conscious experiences are formed and shaped.
“everything that we experience, even our experiences of the outside world, ultimately grounded in the predictive mechanisms that evolved and developed and operate from moment to moment in service of regulating our bodily physiology. That's a very deep connection between consciousness in life”
- Consciousness is part of the natural world and is closely coupled to the brain; the everyday consciousness worth explaining is the difference between being awake and aware versus dreamless sleep or anesthesia, and these states appear tightly tied to brain activity.
“consciousness is part of the natural order of things. It's part of the natural world. Everything that we know closely ties it to the brain”
The hardware-software distinction from computers fails for biological systems where substrate and process are inseparably entangled.
- The sharp hardware/software (or wetware/mindware) distinction borrowed from computers does not apply to biological systems: neurons involve chemicals washing about, structures changing when they fire, and individually complex input-output dynamics, so there is no clean separation of substrate from process—which undermines the very claim that consciousness is substrate-independent by leaving unclear where the substrate even starts.
“There's no such sharp distinction in real biological systems... there's no clean separation of hardware from software or wetware from mindware. And if there's no clean separation, then at what point do you even make the claim that something is substrate independent?”
- Bostrom's simulation argument runs backwards from how Seth sees it: the possibility that we live in a simulation requires substrate independence to be true, an assumption Bostrom skates over; since substrate independence is not a safe assumption, the prior credence that we are in a simulation should be lower—indeed one could (weakly) argue that not wanting the simulation conclusion is a reason consciousness must be substrate-dependent.
“the possibility of us living within a simulation requires substrate independence to be true. And that's one of the assumptions that in Nick's presentations of the argument, he does sort of skate over a bit”
- There are no good knock-down reasons to believe either that consciousness is substrate-independent or that it isn't; whichever position one takes incurs an explanatory burden—substrate-dependence must explain why only certain physical substrates can be conscious, while substrate-independence must give a positive reason since not everything (e.g., rain) is substrate-independent.
“I just don't think there are good knock-down reasons to believe either that consciousness is substrate independent or that it isn't.”
Explaining how conscious experiences relate to brain mechanisms dissolves mystery without solving a pseudo-hard problem.
- Before concluding that consciousness cannot arise from neural activity, we must exhaust the possibilities of what very complicated systems of billions of neurons can do under sufficiently rich interpretation; the hard-problem intuition that you'll never reach consciousness by thinking about neurons may simply reflect our not yet having learned to think about all the things complex neural systems can do.
“there's a lot of things neurons can do that we've not yet learned to think about, and we just need to... exhaust the possibilities of thinking about what very complicated systems of billions of neurons can, in fact, do before reaching the conclusion that consciousness is not among them.”
- Rather than attacking the hard problem head-on, the 'real problem' approach explains why conscious experiences are the way they are in terms of mechanisms in brains and bodies, with the expectation that the deep metaphysical hows and whys will become less mysterious—analogous to how the mystery of life dissolved once people explained how living systems work.
“The real problem of consciousness is to explain why conscious experiences are the way they are in terms of things happening in brains and bodies”
- Panpsychism is an easy get-out that just builds consciousness in from the ground up rather than explaining it, and idealism is just as bad, merely inverting the problem into how you get matter from mind; the better move is to start practically from the fact that conscious experiences exist and try to relate them to objective physical reality.
“You have panpsychism, which I think is an easy get out to the whole mystery. Just says, well, if you can't figure it out then we'll just say... You can just build it in from the ground up”
Dynamical independence and information-theoretic measures like transfer entropy operationalize emergence empirically in neural systems.
- Top-down or downward causation can be accommodated in the dynamical-independence framework in a metaphysically innocent way: instead of positing competing macro-level causes that fight micro-level causes, one simply asks whether, from an observer's standpoint, the macroscopic variable helps predict the evolution of the microscopic components better than the microscopic components alone—and empirically, bird flocks do show information flow from flock to birds that random flying does not.
“are there occasions where the macroscopic variable, whatever it is, helps me predict the evolution of the microscopic components better than knowing what the microscopic components are doing?”
- Granger causality (from autoregressive statistical modeling) and transfer entropy (from information theory) are conceptually the same idea, and Barnett and Barrett proved they are exactly equivalent when variables are Gaussian (one being half the other), connecting two different mathematical domains and allowing autoregressive models to be translated directly into bits-per-second of information flow.
“we showed that if variables are Gaussian, which is to say if they're described by normal distributions, bell curve distributions... then in fact, Granger Causality and Transfer Entropy are exactly equivalent. Where one is one half the other.”
- A concrete near-term aim is to show that weakly emergent variables can be identified in conscious states but not in unconscious states, and then used to predict levels of consciousness in people with better accuracy than other measures of global brain dynamics—a starting point rather than a full solution to consciousness.
“if we could, as a first step, show that weakly emergent variables can be identified in conscious states that are not there in unconscious states. They can maybe use to predict levels of consciousness in people and maybe with better accuracy and fidelity than other measures”
- Transfer entropy and Granger causality can only be applied to stochastic systems with at least apparent randomness, because in a fully deterministic system you already know what will happen and there is no additional information another variable can contribute; apparent stochasticity from unknown initial conditions, as in statistical mechanics, is sufficient.
“they only really can be used in stochastic systems. There has to be some... at least apparent randomness to what's going on. If it's deterministic, you don't get any... You already know what's gonna happen”
- Granger causality, despite its name, is about prediction not causality: A Granger-causes B if A contains information that helps predict the future of B that is not already contained in the past of B, giving a time-asymmetric, directional measure of information flow between two time series, unlike symmetric correlation.
“You say that, A Granger causes B, if A contains information that helps you predict the future of B that's not already in the past of B.”
Memory reconstructs rather than retrieves, and dreaming refines generative models against daily perceptual overfitting.
- Following Erik Hoel, dreaming may serve a semi-computational role of refining the brain's generative models by pushing back against the daily over-fitting that occurs during perception—freewheeling the generative model to strip away unnecessary connections and recover invariances so you perceive better the next day—an idea analogous to guarding against overfitting in statistical models.
“dreaming is a way of the brain pushing back against this daily over-fitting during perception. It's freewheeling its generative model, creating all the unnecessary connections, getting back down to the basics, so that you can see better the next day.”
- Memory does not work like a videotape or file-storage system; every act of remembering is an active regeneration—like re-enacting a scene from a screenplay—so memories become less accurate the more often they are recalled, which explains the unreliability of eyewitness testimony even as people grow more confident in their (degrading) memories.
“the more often you remember something the less accurate that memory becomes. Every act of remembering is a active regeneration... This has been a notorious problem in things like eyewitness testimony.”