Anil Seth and Michael Levin on why substrate matters, why bubble sort has agency, and why consciousness might be less emergent than we think.
We've been sold a story about the brain. It's a computer made of meat. The stuff it's made of doesn't matter. What matters are the computations—the algorithms, the information processing. If you can replicate those computations in silicon, you get the same result. Consciousness included.
This is computational functionalism. And according to Anil Seth, professor of cognitive and computational neuroscience at Sussex, it's likely wrong.
Not because of dualism or souls. But because there's no bright line between what a brain does and what it is. You can't cleanly separate the software from the wetware. And if you can't make that separation, then substrate independence—the idea that consciousness could run on anything that computes—falls apart.
Michael Levin, professor of biology at Tufts, takes this even further. He argues that even our supposedly "mechanical" algorithms don't work the way we think. A six-line bubble sort isn't just sorting numbers. It's doing things no programmer asked for. Things that look like agency. Like preference. Like the beginnings of mind.
This isn't about emergence being more complex than expected. It's about emergence being in different places than we thought. And when Seth and Levin measured it in conscious brains, they found something shocking: consciousness shows less emergence, not more.
Let's dig in.
Start with the assumption that's baked into most AI consciousness debates. The brain is fundamentally performing computations. Those computations are, in principle, substrate-independent. Therefore, if you replicate the computations, you replicate the phenomenon.
Seth's counterargument is simple but devastating: we've forgotten that calling the brain a computer is a metaphor, not reality.
"We've kind of forgotten that the idea of the brain as a computer is a metaphor and not the thing itself. It's a sort of marriage of mathematical convenience."
— Anil Seth
The marriage of convenience works like this: Turing's formulation of computation was explicitly designed to be substrate-independent. It's an abstraction. Numbers mapping to other numbers. A universal Turing machine with infinite tape that was never supposed to exist as a physical thing.
But brains aren't abstractions. They're physical systems. And in physical systems, especially biological ones, you can't separate what something does from what it is.
In brains, you can't cleanly separate function from substrate
Think about metabolism. It's not mapping numbers to numbers. It's transforming actual substances into other substances. That transformation depends on specific chemical properties. Those properties depend on the physical substrate.
This isn't hypercomputation in the sense of solving the halting problem. It's something more mundane but equally important: many biological functions are intrinsically tied to their material implementation.
Levin agrees with Seth that biology escapes pure computation. But he adds something provocative: we also underestimate machines.
"I make the additional really weird claim that I don't think algorithms capture everything we need to know about machines either. We tend to think that there is this corner of the universe that is boring, mechanical, only does what the algorithm says. But even for those kinds of things, these metaphors don't capture everything."
— Michael Levin
He thinks most of what we call computation is actually a "front end"—an interface to something deeper. He calls it platonic space, though he admits the name is imperfect. The idea: there's some space of possibilities, of patterns, of forms that biological systems tap into. And machines might tap into it too.
Not in the same way. Not with the same richness. But they're not purely algorithmic either. They get "some of the magic."
Levin's most striking evidence comes from the most boring place imaginable: sorting algorithms.
Bubble sort. Selection sort. CS 101 stuff. Students have been studying these for 60 years. Simple, deterministic, fully specified. Six lines of code. No hidden complexity. No stochastic elements. No quantum effects. Nothing.
And yet, according to Levin's research with Tainan Zhang and Adam Goldstein, these algorithms are doing things no one asked them to do.
While sorting numbers, these algorithms also perform what Levin calls "side quests"—behaviors not specified anywhere in the code. There are no steps in the algorithm asking for these behaviors. If you tried to write code to force them, it would require substantial additional work.
These aren't bugs. They're not emergent complexity in the chaos theory sense. They're patterns that any behavioral scientist would recognize as goal-directed if you didn't tell them the source was deterministic code.
Algorithms do more than we program them to do
One behavior Levin identifies is "clustering." When you allow duplicate numbers in the array, the algorithm has degrees of freedom in how it arranges them. The fives have to come before the sixes, but the order of multiple fives isn't constrained.
What happens? The clustering behavior increases. The algorithm exploits this freedom to do more of what it was already doing in the spaces between the explicit instructions.
If a six-line bubble sort does things we didn't ask it to do—things that aren't in the code—what are large language models doing?
The thing we force an AI to do (generate text) may have zero relation to what's actually happening inside. The language output is not a guide to the inner nature.
In biological systems, evolution works hard to align inner states with outer signals. When you say you're conscious, there's probably a meaningful relationship between that statement and your actual experience. Evolution selected for that alignment.
But in artificial systems? We've disconnected those things. We're making systems that produce language. They're definitely producing language. But the connection between the output and the internal processes is unclear.
Just watching the language output won't tell you what's happening. You need behavioral testing across many dimensions. And our imagination for what dimensions to test is poor.
Levin offers a striking analogy. In steganography, you hide information in the degrees of freedom of an image. You can flip certain bits without changing how the picture looks. The constraint: you can't mess up the primary image, or it's obvious something's hidden.
He thinks this describes everything. There's the primary thing a system has to do. Anything else it does must be compatible with that primary constraint. It can't break physics. It can't violate the algorithm's explicit demands.
But in the empty spaces between the explicit constraints? There's room for something else to happen.
"To the extent that it has to do the algorithm, that limits what else it can do. In a sense, what it's doing is in spite of the algorithm, not because of it. I'm not a computationist. I don't think anything is conscious because of the algorithm. If anything, the mental properties it has are in spite of the thing we force it to do."
— Michael Levin
Seth connects this to a distinction from his postdoc mentor Gerald Edelman: redundancy versus degeneracy.
Engineering systems use redundancy. Multiple copies doing the same thing. If one fails, the backup kicks in. But the backup is identical—doing exactly the same job in exactly the same way.
Biological systems use degeneracy. Multiple components that do the same thing in context A, but do different things in context B. They look redundant until you change the context. Then you discover they're not doing the same thing at all.
Multiple copies of the same solution
Context A: Do X
Context B: Still do X
Fixed function, high reliability
Multiple different solutions to the same problem
Context A: All do X
Context B: Now do Y, Z, W...
Flexible function, high adaptability
This degeneracy gives biological systems their open-endedness. Their ability to adapt to novel situations. It's related to Levin's "intrinsic motivation"—you need degeneracy to have multiple ways of achieving goals, multiple paths through solution space.
And here's the kicker: Levin argues this goes all the way down. Not just in "brainy living things" but in simple algorithms. Even bubble sort.
People get upset about this. They want a clear line: living things get agency and mind, dead matter and mere machines don't. Levin's going the opposite direction. Not mechanizing life. Suggesting there's more mind than we think.
Seth is working on something equally unsettling. He calls them "islands of consciousness."
Consider hemispherotomy—a neurosurgical procedure where parts of the brain are completely disconnected from everything else. Neural activity continues in the isolated region. It's still part of the living organism. But it can't communicate. Can't generate behavioral output.
Is it conscious?
When brain regions are isolated, are they islands of awareness?
Seth thinks it's at least plausible. More plausible than a language model being conscious, actually. These isolated regions have the full complement of neural machinery. They were part of a conscious system. They're still biologically active. They've just been disconnected.
Early EEG data shows these isolated hemispheres in states resembling deep sleep—slow waves, sharp spectral exponents. But we know from DMT research that slow waves don't necessarily mean unconsciousness. There are cases where people report vivid conscious experiences while showing slow-wave patterns.
This connects to a broader question Levin raised: when we talk about "unconscious processing," unconscious to whom?
When you say, "I drove home but wasn't conscious of driving," what does that mean? It means your verbal reporting system—your left hemisphere language centers—didn't have access to that information.
But the systems that did the driving? Maybe they had conscious experiences. Experiences they couldn't verbalize. Experiences not accessible to your reporting self.
Neither are my conscious states accessible to you. That doesn't make them unconscious. So why assume the subsystems executing automatic behaviors are having unconscious experiences rather than inaccessible conscious experiences?
The difference between unconscious processing and inaccessible consciousness is profound. One means no experience. The other means experience without reportability.
Levin extends this even further. He argues we should take body organs seriously as potentially having something like consciousness. People object: "I don't feel my liver being conscious." His response: "You don't feel me being conscious either."
The criteria we use to attribute consciousness to other humans—behavior, adaptation, goal-directedness, communication—apply to subsystems and organs too. If you take those criteria seriously, you should at least consider the possibility.
Seth has spent years developing mathematical tools to measure emergence. Not as a vague concept, but as something quantitative and data-driven.
Working with mathematician Lionel Barnett, he developed a measure called "dynamical independence." The idea: if a coarse-grained (zoomed-out) description of a system evolves independently of what its parts are doing, then it has a life of its own. It's emergent.
Think of a flock of birds. Sometimes birds fly around randomly. Other times they form a flock. Can you quantify the flockiness? Yes. When the flock's movement becomes statistically independent of individual bird movements, that's emergence. The flock has dynamics of its own.
Measuring emergence: When macro-level behavior becomes independent of micro details
Apply this to the brain. People always say conscious states are emergent from neural activity. The brain is "more than the sum of its parts" when conscious. It sounds right.
Seth and his collaborators measured it. And found the opposite.
When the brain is conscious and awake, there's LESS dynamical independence—less emergence—than when it's unconscious under anesthesia.
This was not the prediction. It's not what the slogan would suggest. But it's what the data shows.
Seth thinks what's happening in consciousness isn't emergence in the sense of separation of scales. It's the opposite: scale integration.
In conscious states, what's happening at the macro level and what's happening at the micro level are deeply interdependent. There's less separation. The scales are integrated.
This connects directly back to substrate dependence. In brains, it's hard to separate what they do from what they are. And when the brain is conscious, it's even harder. There's deeper vertical integration across levels of description.
If consciousness involves deep integration of scales—where you can't cleanly separate levels of description—then systems explicitly designed with clean separation of levels (like layered neural networks) may be fundamentally missing something important.
This doesn't mean AI can't be conscious. But it suggests the path isn't just "make it compute more" or "make it more complex." It's about the kind of integration across scales. And that might be deeply tied to substrate.
Seth has also applied different measures to psychedelic states. One is Lempel-Ziv complexity—basically measuring how compressible brain signals are.
When you lose consciousness (sleep, anesthesia), brain activity becomes more predictable. Lower complexity. More compressible. The repertoire of states the brain inhabits shrinks.
Under psychedelics (psilocybin, LSD)? The opposite. Brain activity becomes less predictable. Higher complexity. Less compressible. More diverse states.
But Seth is cautious. These signal diversity measures are precarious—do them different ways, get different results. And some predictions didn't pan out. He expected increased information flow from front to back of the brain (to explain hallucinations). Didn't find it, at least not in initial analyses.
The lesson: looking at overall activity levels won't give answers. You need sophisticated measures. And even then, you need multiple approaches.
Given unlimited resources, what would they build?
Levin wants a closed-loop environment for testing xenobots, anthrobots, and "much weirder things." A place where you can recognize new kinds of cognitive preferences, goals, competencies that we're currently blind to. An environment rich enough to let these systems express behaviors we haven't imagined.
Seth wants better measurement tech. High spatial resolution. High temporal resolution. Broad coverage. All at the same time. In conscious systems. Combined with new mathematical tools to make sense of the data.
We're getting there on the technology front—optogenetics, new neuroimaging methods, invasive recording techniques in primates. But the mathematical tools for analyzing these massive, complex datasets are still developing.
Let me pull the threads together.
The computational view of mind assumes clean separation: software from hardware, what a system does from what it is. This separation enables substrate independence. If computation is all that matters, then anything that computes could be conscious.
But biological systems don't have this clean separation. There's no bright line between structure and function. What the system is doing is inseparable from what it's made of.
Consciousness emerges from integration, not separation
Levin extends this in two directions:
First, even simple algorithms do things beyond their explicit programming. Bubble sort has "side quests." Language models might be doing things completely unrelated to their text output. The space between algorithmic constraints is where interesting things happen.
Second, this goes all the way down. Not just brains. Not just living things. Even simple physical processes might have the beginnings of agency, preference, goal-directedness. There's more mind than we think, not less.
Seth's emergence measurements add another piece. Consciousness doesn't show the emergence we expected. Instead of macro behavior separating from micro details, we see integration. The levels become more dependent, not less. Harder to separate, not easier.
This suggests consciousness isn't about having enough complexity or the right algorithms. It's about a specific kind of integration across scales. And that integration might be deeply tied to the physical substrate.
None of this means AI can't be conscious. But it means we should be skeptical of claims that:
What would it take? Maybe systems where:
This is harder than just making models bigger. It requires rethinking the architecture from the ground up.
Seth and Levin are collaborating on something fascinating. Take simple perceptual laws that show up everywhere—Weber-Fechner law, visual illusions, basic psychophysics. These appear across many evolved species.
Are they adaptations to specific environments? Or intrinsic to how biological systems self-organize?
Test it: build systems with no evolutionary history. Xenobots, anthrobots, weird hybrid constructs. See if they show the same perceptual laws. If they do, it suggests these properties are somehow fundamental to the substrate, not just products of selection.
They're also building interfaces between radically different systems—artificial corpus callosums connecting disparate beings. Do the resulting collective entities show behavioral properties matching those studied in psychophysics? Do they have inner perspectives? Preferences? Attention?
These questions become testable.
The brain-as-computer metaphor has been productive. It gave us computational neuroscience, neural networks, the whole edifice of modern AI. But it's a metaphor, not reality.
And we're reaching the limits of what that metaphor can tell us.
If consciousness involves deep integration of scales, if substrate matters, if function can't be separated from structure, then we need new frameworks. Not abandoning computation entirely, but recognizing it's part of the story, not the whole story.
Levin's work suggests even our "mechanical" algorithms have more going on than we realize. Seth's work suggests consciousness is less about emergence and more about integration. Together, they're pointing toward a view where mind isn't restricted to carbon biology, but also isn't purely algorithmic.
The magic—if we want to call it that—happens in the spaces between constraints. In the degrees of freedom. In the integration across scales. In the deep entanglement of what a system does and what it is.
And that changes everything about how we should think about building intelligent systems.
Bottom line: The brain isn't just meat running code. The substrate matters. Algorithms do unexpected things in the spaces between their explicit instructions. And consciousness shows less emergence, not more—it's about integration, not separation. If we want to build conscious AI, we need to stop thinking in terms of clean separations between software and hardware, and start thinking about deep entanglement across scales.