A mind needs to think it has a body to function like a human mind! But nothing is stopping a Chinese room or LLM from employing a virtual body, virtual eyesight etc.
Earns it's keep, as in worthwhile investing compute in? I think that's certainly theoretically achievable to a high degree with LLMs, but it's not a good or efficient way of doing it. So if that's what you're asking, then no: if the goal is to produce AI as close to human minds as possible (which we shouldn't aim for!), then I think pure LLMs are a bad idea.
I'd say LLMs already model typical human bodies and environments to a coarse approximation. This falls out naturally from the fact that words are a very compressed form of representation, but they can in principle represent anything.
A high precision virtual body is probably not going to be possible to model with current text corpuses, though. The training data is simply not representative of what human bodies do in general. Humans do not typically describe in great detail how to hold a spoon, ride a bike, or how to dance the lindy hop. Or plain old walking. We learn these things primarily by other kinds of representations. But I claim there is no inherent limitation to the text format. It can describe anything. (here, I use "representations" in the broadest sense--all the way down to retinal computation, computation in peripheral ganglia, spinal cord etc--perhaps that's misuse of the word).
A complete Chinese room rule book that behaves like a Chinese person in every way and passes all Turing tests, would necessarily have to model a certain individual with high precision. It would have to believe it is a human with a human name. After an extended chat session, it will have to ask for a toilet break, or become more "dense" or giggly because of fatigue. If we ask it to describe the room it's in, it would have to give us plausible and consistent answers ("what a weird request" would be one plausible answer). If it can't do all those things consistently, it fails a rigorous Turing test. It hasn't got a good rule book! My claim then, is that it would model a body AND environment to the degree that the same (or similar) causal structures exist
I totally get what you’re saying — and I think Dennett would applaud much of it. The idea that we should invoke representations only when they help us explain and predict behaviour? Classic intentional stance. And the thought that language compresses patterns from body and world? Definitely in Dennett’s toolbox.
But I think he’d still press you — just a little — on where the justification for calling it a mind actually comes from.
Dennett’s version of “earning its keep” isn’t just about passing a test or producing the right outputs. It’s about whether positing beliefs or experiences gives us the best explanation of the system’s behaviour, given its whole causal history and design. If your Chinese Room “needs” a toilet break but nothing happens if it doesn’t get one, do we really say it has that experience — or just that it’s emitting the right noise?
So he’d share some skepticism that high-fidelity modelling — no matter how consistent or data-rich — can stand alone. Without the right causal embedding— a control architecture in which the system’s ongoing coherence depends on interpreting and acting on its own signals — the symbols may be doing things, but they still haven’t earned intentionality.
Where Dennett diverges from enactivists is on the stakes requirement. He wouldn’t insist on biology: a purely artificial architecture could, in principle, earn belief-and-desire status if its internal organisation supports the right predictive patterns.
Which leaves us with the puzzle: what counts as the “right patterns” — and would a toilet break ever "matter" to a large language model?
(btw... I'm sorry I haven't yet replied to you reply from last week. I am planning to.)
Hmm, I think I'm being unclear again, as in the LLM recursion discussion where you in the end totally got what I meant. Alternatively I'm misunderstanding, please push back again if you think so, even if you don't have the time to get into any detail.
I speculate that Dennett would agree that simulated physical systems, in principle, are just as good for evolving and hosting conscious agents as the physical system of our world is. So, in principle, we can just modify the Chinese room thought experiment like this:
-Put a volunteer Chinese person in a room with a bed and toilet , with the instructions to reply to notes that are slipped under a door.
-take the physical state of the universe, or just of the room with the Chinese person in it, (highly recommend the latter—llet's presume this scenario).
-put the physical state of the room, as well the laws of physics and the instructions for how to evolve the state in a rule book.
-also put in the rules that translate notes with Chinese characters into simulated notes with the same Chinese characters in the simulated physical system, and the vice versa.
-have testers always include time stamps of when they slip notes under the door. In the rule book is included the instructions that evolve the physical state to that time before the simulated note appears.
Now, given that Searle, or whoever, can follow the rules in real time (admittedly ot conceivable, but the same is true for the original argument, where a person and a rule book could never produce a response during a lifetime), nobody who doesn't know the setup are going to be able determine which room contains searle and which room contains the original Chinese person. For example, if the Chinese person says “shit, the blood in my urine has come back! I hope not the bladder tumor has come back!”, then searle in the room will produce the same response, since the room simulates the persons body to perfect (or arbitrarily high) precision.
Here, the whole system has a mind, albeit only as a subset of it. It captures the evolution of all the causal structures in the Chinese person, by copying the physical state of the real person. If the history troubles you, you can modify the experiment so that the Chinese person has been simulated from birth, and you got the whole history of the organism in place (I don't think this is relevant, swampman arguments are poor in my view, but you can do it this way if you want to).
So, with the right training data and unimaginably high compute and memory, an LLM can learn all the causal structures of any physical system, merely by training on English text tokens that describe that system and how it evolves. It doesn't even need to be trained on Chinese text, or any neurophysiology of biology or chemistry at all, because all of that supervenes on the physical. It does need to learn to translate Chinese characters into English descriptions of what those characters look like, and how to in turn translate those descriptions into modifications of the physical state of the simulated room. And it needs to learn to do the reverse too.
Here, the LLM is optimised to predict the next token and does so as well as a real physical room and Chinese person system does. The LLM as a whole doesn't have a mind. The room as a whole doesn't have a mind. (this is kind of an anti-system reply!). But both the room and the LLM contain, as a subset of all the causal structures, a mind. And crucially, the brain is ‘embodied’ in both cases!
I don't see why Dennett would argue there's a difference between the virtual room, body and mind on one hand, and the real counterpart, but please correct me if I'm wrong!
From all of this follows that LLMs can, in principle, learn to model bodies to arbitrarily high precision, without modelling all the physics. This would require vastly different training data (and enormous compute+memory) than is currently used. Current data is very selective in kind, including only what people typically write about, and produces some kind of simulated ‘average’ person rather than an individual person. Nor does it include anywhere near enough natural dialogue to capture the aspects of cognition and self-correction that only show themselves there. (These two points, I believe, are the main reasons for the characteristics of LLM hallucinations that differ from human hallucinations (which are also a huge ‘problem’).)
It doesn't matter that the LLM itself is optimised for mere token prediction, if predicting the right tokens necessitate developing the same causal structures that are present in one of us. It doesn't matter that the LLM doesn't need a toilet break if there are subsystems that do.
How much precision in brain and body modelling is needed to grant part of the LLM a mind? What about an embodied mind? That question mirrors the question of which animals are conscious, or when a human becomes conscious, as I see it. It depends entirely on definitions, and there's no definite answer.
OK I got a bit sidetracked there. Anyway, I'm glad for any comment.
Don't worry if you don't get to replying to the other comment. You have a lot of followers and commenters and have to prioritise, I get that!
"The split isn’t as clean — or as necessary — as it’s sometimes made out to be."
This is the sense I have every time I read about enactivism, or the overall 4Es (enactive, embodied, extended, embedded). I think these views are right, but not to the exclusion of everything else. Many of the proponents want to paint it as a radical break, but the details seem to show it's just an elaboration of the functional representational paradigm, a fine tuning of it rather than an overturning.
I also think the comparison of the radical 4E stance to behaviorism is a good one. It doesn't seem like you can have it both ways, claiming mental states don't exist and then claiming you're doing something different from the behaviorists.
I also think we have to be open to different types of "bodies". The real issue is interaction with an environment. That environment doesn't necessarily have to be the physical world as we experience it. It could be a virtual one, or even a software ecosystem. Granted, radically different environments probably mean radically different minds.
I absolutely love these philosophical explorations.
It's funny, because the brain (where the mind resides, presumably) is, in fact, part of the body. So by definition the mind needs a body. I would go even as far as to say that the whole thing - your body from head to toe - is your mind (not sure if this a controversial take, but if it is I'll take it).
PS. Even large language models have bodies, one could argue, in the form of large datacenters. Without those physical 'bodies' those algorithms wouldn't exist and be able to perform their calculations.
See writings on the mereological fallacy. Peter Hacker uses this term (see for example: https://www.youtube.com/watch?v=EMcmQPdi0Fs). Sometimes it is also known as the humunculous fallacy. The latter term was used by Anthony Kenny in a paper published in 1971. Their arguments relate back to Ryle and Wittgenstein.
The dichotomy of enactivism versus teleosemantics seems to ignore something significant: the brains and minds of humans and many other species are plastic. The neonatal brain has been shaped by a half-billion year training process called evolution, but the mind it hosts is far from an adult mind. Instead, the infant brain is an organ capable of bootstrapping that mind by learning from experience.
The enactivists would push back and say: maybe for things that are right in front of your eyes, you don’t need a representation of a gorilla. That kind of thinking suggests a passive inner theatre, where someone has to interpret what’s on the screen -- and that quickly leads to the old homunculus problem.
Instead, they’d say all you need are the skills and sensitivities to spot the gorilla when it appears. They'd argue, perception isn’t passive reflection -- it’s active engagement in real time.
My response would be to ask how that changes anything - one still needs to recognize the presence of a gorilla and associate the phenomenon with the noun. Furthermore, I can imagine, in my mind's eye, versions of the scenario in which the intruder is an actual gorilla or a robot, where there is no such thing right in front of my eyes.
The second putative response - active engagement in real time versus passive reflection - looks like a straw man to me. Is anyone seriously or tacitly assuming the latter?
One thing that I will grant to the anti-representationalist aspect of enactivism is that there may well be a low-level causal explanation of what is going on in one's brain during this event that never mentions the representation of a gorilla; any such representation would be a higher-level abstraction. While that might be a helpful point to bear in mind, it does not seem to me to be a key insight in unraveling where meaning comes from, any more than would pointing out that the video does not actually put a gorilla right in front of one's eyes.
Thank you for this overview of something that is quite in fashion today, but I can't quite wrap my head around it.
I don't find anything particularly original or new in enactivism. It reiterates concepts we already know from our daily experience, several of which were expressed by past philosophers long before brain scans. Perhaps enactivism rephrases things with more precise and rigorous terminology and tests them with the latest neuroscience, but in essence, it merely rephrases what is (more or less intuitively) common wisdom. Where is the news?
Where’s the news!? Great question. Is enactivism just common sense in fancier packaging?
I wonder about this too. It’s often touted as revolutionary -- but it’s difficult not to see it as simply a new way of putting together old points we’ve long suspected were true but didn’t quite know what to do with. At best, maybe it’s a well-thought-out critique of traditional cognitive science.
But maybe that’s exactly the news: not that enactivism says things we’ve never heard before, but that it organises a scattered set of intuitions into a framework that challenges some of the default assumptions in traditional cognitive science.
That doesn’t mean we throw cognitive science out. But it might mean enactivism has something important to say. It is an important nudge away from some of the admittedly deplorable features of traditional cognitive science.
Enactivism is coming upon its decisive moment: as more humanoid robots enter the world and learn from it. These are embodied (unlike LLMs), and they can in principle build up from sensory experience into meaning, which LLMs almost by definition can't do. Critically these will be less pre-programmed and more evolutionary than past AIs, learning (aka updating their own training data set) based on what they observe.
I predict these will ultimately surprise people with the relative richness of their capability, much as LLMs did when they first appeared. But will they seem human? We'll see...
Hey Bill! Embodied interaction, I think, is an important piece of the puzzle. But I wonder whether that is all that matters. I feels like simply interacting with the world, might not be enough. I wonder what is also required is that their continued functioning depends on those interactions. That’s when things get interesting -- when the system has something to lose.
Yes for sure. Interacting with the world almost by definition requires a goal, or objective; or definition of success. Otherwise what is guiding the interaction?
We build robots that feed various types of input information from the machine body to its computer operator, and quite like our bodies feed various types of input information to our brains. So does that make our robots conscious? Of course not! And while enactivists surely don’t believe this either, one might still accuse them of wasting our time for implying such vague notions to inevitably reject. But if that’s not their position then do they have anything of substance to propose at all? Did I miss something Suzi?
Instead of requiring a body, my position is that to create consciousness our brains need to implement the proper sort of physics. From this position there’s essentially a consciousness loop that our brains use to create the consciousness that is ultimately what we are. So while brains function non-consciously, the consciousness that they create adds value or purpose to a given organism’s function.
Great question Mark! Start by considering a non-looping consciousness. That would essentially be epiphenomenal — you see, hear, feel, think and so on, but you’re ultimately just a spectator so your choices wouldn’t be able to affect your body. With a full loop however your choices feed back to the brain and become processed by it in ways that tend to operate your muscles as desired. So as I see it the brain is a fully non-conscious computer, though one of its output mechanisms creates the value driven form of computer that is consciousness itself, and this tends to loop back to feed the brain with a value based form of information.
Aha. A loop from the computer to consciousness and then back. Gotcha. I agree there can be no consciousness without loops, but my view is closer to "consciousness IS the loop".
Two questions:
Is there no value to, or value for, life forms that that haven't moved beyond computation? Don't simple organisms compute simple forms of value?
If consciousness processes value as a result of the right kind of physics, then it's the causality inherent to the right kind of physics that matters, is that correct?
On your first question, Mark, no I don’t consider simple forms of life to compute simple forms of value. The value that I’m talking about equates with consciousness itself and is essentially Schwitzgebel’s “innocent/wonderful” definition (the one that Keith Frankish has admitted to believing in). Theoretically life evolved, though without any such consciousness/value. Then brains evolved in life and still nothing. Such algorithmic operation alone however must have entailed certain vulnerabilities. I think it required too much programming. Regardless, once the consciousness loop did emerge it eventually succeeded well enough for even human consciousness to evolve.
On causality being inherent to the right kind of physics, yes agreed. But in a causal world that will be true by definition. I’ll only be able to demonstrate how my models work when you ask me practical question, such as your first one. I very much appreciate that you’re trying to understand!
What do you mean by programming? I get that you don't mean literal programming, but I'm struggling to make any sense of it. (also, I'll point out that moving away from programming is what lead to real AI breakthroughs. Chatgpt isn't impressive because of programming).
Ok, causal structures it is! I'm glad we agree on that.
Ok here's a practical question on how you would implement your theory:
Let's suppose we find out that the right kind of physics in humans is an EMF. Now, that's of course not a theory of how the right kind of physics brings about an experiencer, but at least we know that we should be analysing the EMF as much as we can. Let's suppose that once we zoom in on the EMF, we get a pretty good mathematical model of what kind of causality the EMF instantiates.
Furthermore, let's suppose that we discover that octopuses have all the same causal structures that instantiate experience in humans--all the same math applies to what's going on in their brains. Only, they don't employ EMFs to instantiate those structures! Their neurons are simply wired in ways that create the same causality! Octopuses have found very different solutions to various problems than vertebrates have, in several ways, so this wouldn't be a total shocker.
Let's also say we verify that if we meddle with the causal chains in certain ways, it causes similar effects as the corresponding meddling with the causality in EMFs.
In such a case, how do we decide which "method" is the right one for 'real' consciousness? How do we have a preference for one sort of physics over another, if all the same causal mechanisms are in place?
This is of course very much related to the fact that you bit the bullet that a robot behaving just like you is conscious, yet we can straightforwardly describe how to build such a robot without EMFs.
(As I'm sure you understand, causal structures are substrate independent. I can see nothing to suggest that the causal structures that emerge in an EMF are not computational. Nor is there anything to suggest that the same causal structures without an EMF are 'fake'. Evolution can find many different solutions to the same problems. This shouldn't really be a point of disagreement, since you agree that robots that behave like you in every situation are conscious.)
Okay Mark, that’s a more practical question that I’m able to explore. Firstly however, yes I again errored by using the “programming” term. The simple fix in this situation will be for me to always substitute the term “algorithm”. I theorize that as environments become more “open” (effectively the opposite of the game of Chess that has limited movement options), that algorithms alone were not a sufficient means of operation. At some point organisms developed purpose based function as well. So here the biological computer effectively punishes and rewards the sentient experiencer that it creates, and then bases its further algorithms upon the choices of that experiencer. Many seem to find this idea difficult to grasp, but it is my basic perspective on the matter regardless.
Apparently a third of their neurons are in their heads while two thirds are distributed between eight arms. So my suspicion is that if EMF consciousness gains empirical validation, it will be empirically found that they have a central head based EMF consciousness to help provide a general value dynamic to that algorithmic head computer, but there’s also a computer in each tentacle that’s armed with its own individual consciousness. Of course for survival each of these nine conscious beings that share a body would generally need to work together as a team.
Before I get into the scenario that you provided, let me briefly describe what it is that makes this general consciousness account so persuasive to me. Because the physics of pain, vision, hearing, smell, and so on might be electromagnetic, I can imagine neurons firing in the proper ways to create an electromagnetic field that experiences its existence in such ways. But I don’t know of a second thing that neurons or anything else in the brain might be informing to exist as consciousness itself. What else could have the required potential bandwidth to exist that way? I just can’t fathom what else. To me other proposed theories seem causally impossible, not to mention unfalsifiable.
So on to your question. If it’s empirically found that octopuses, birds, robots, or anything else harbor a consciousness that is not electromagnetic, then EMF consciousness would thus be proven wrong. You mentioned mathematical derivations though to me that’s getting things backwards. Empirical evidence rather than mathematics tells us how causality works. Math is simply a human invented language, and actually far less advanced than natural languages like Swedish and English.
Anyway if it were empirically demonstrated that consciousness does not exist electromagnetically, in a sense this would be exciting! In that case consciousness science would finally have an example of how to actually do consciousness science. If they can empirically eliminate this possibility, can they also empirically validate a more causally appropriate solution? If consciousness is merely related to EMF, then what related element might that be?
Given that the shape, color, movements, and so on of a blue mug are processed in different areas of your brain, how does all this come together for you to perceive a unified blue mug? Because every bit of information exists at all points in a given electromagnetic field, EMF consciousness has no binding problem.
You're absolutely right that no enactivist would claim that just feeding sensory data from body to brain (or machine to operator) is enough for consciousness. That kind of input-output setup is exactly what enactivism tries to move beyond. The critique is aimed precisely at that Cartesian picture: a passive receiver sitting inside, watching a screen.
It’s not that the body is a requirement like a power cable -- it’s that the body shapes the kind of problem-solving, timing, and feedback loops that give rise to meaningful experience. So enactivists are saying: don't look for consciousness inside, look at the processes -- kind of like your loop. Except they say the loop is not restricted within the brain.
When you phrase it like that Suzi, it sounds like functionalism. Of course my criticism of computational functionalists is that they don’t take causality to its end — consciousness is presumed to result by means of the processing of information rather than processed information that informs something causally appropriate to exist a given consciousness. I wonder what they’d say about my thumb pain thought experiment? Maybe they’d agree with me that thumb pain would not result by means of the processing of certain marked paper to other marked paper, but would disagree that the output marked paper would need to inform the proper sort of brain physics that would thus reside as such an experiencer? Maybe they’d say the marked paper would need to inform a full body that thus acts like it’s in pain?
It’s interesting how many serious people who truly are trying to not propose magical ideas, still end up proposing unfalsifiable consciousness notions. So the marks on paper enact various robotic operations, and because the correct wiggling, audio, and whatnot result, the robot thus experiences pain. I wonder if they argue around the situation where people experience pain but can’t show it because they’re perfectly paralyzed? Do they bite the bullet at tell paralyzed people that they thus have no consciousness? Or what complexity do they then posit to save their ideology?
Yes, there’s a strand of classic computational functionalism that still treats “the right information processing” as sufficient for consciousness. But, as you note, that leaves a big question: How does the processing connect causally to anything that actually "matters" to the system?
So we have three loaded phrases here — “the right information processing,” “consciousness,” and “what matters to the system.”
(Don’t worry, I’m not going to unpack them all right now.)
Because of this tension, some functionalists have edged away from the simple brain-as-computer metaphor:
Pragmatic functionalists — Dennett is the classic example — say we should call a pattern in the brain a “belief” or a “desire” only when using that label makes it easier to explain and predict what the system will do next.
Enactivist-leaning functionalists go further: they argue that meaning—and eventually consciousness—emerges only in self-maintaining, self-constraining systems with stakes. A signal counts as “information” only if it feeds back into the system’s own viability.
Your thumb-pain thought experiment highlights that concern perfectly. The concern is that it is not enough for a machine to generate the symbol “ouch.” According to the enactivist-leaning functionalist, that signal has to reverberate through a loop that regulates the system in ways that matter to its continued existence. A purely reactive body isn’t enough; a body embedded in a world, invested in outcomes, edges closer to something we might call "experience".
Some might say the real challenge — that both camps have — is to explain how signals become significant without smuggling in magic or reducing everything to outward behaviour.
Ah that is helpful Suzi! I’ll reduce what you’ve said into a scale of functionalism from lower to higher “bar”.
It sounds like Dennett had a quite a “low bar” type of functionalism. For example he presumed that a strong Turing test passing computer would understand Chinese symbols no less than as an educated Chinese person. Thus if Searle had such a computer’s rule book and were to use it to construct appropriate output Chinese character responses from input Chinese characters, theoretically this system in full would effectively understand Chinese just as educated Chinese speakers do. Furthermore low bar functionalists also tell us that if the proper marks on paper were algorithmically processed to print out the proper other marks on paper, then something here would experience what they do when their thumbs get whacked.
Apparently you’re saying that enactivists are functionalists too, but with a higher bar. They’d tell us that the Chinese room and processed paper scenarios would need to be implemented in conjunction with body loops that have self-maintaining, self-constraining “stakes” — feedback to a system’s viability. Thus any sophisticated computer operated robot that’s set up to move around autonomously in nature should do the trick. So I think I now grasp how an enactivist would interpret my thought experiment. They’d say in order for thumb pain to result from marks on paper that highly correlate with the information that their whacked thumbs send their brains, the processed information shouldn’t go on to print out the correct other marks on paper. Instead it should inform a machine that functions very much like they do. Though in practice this sounds like a much higher bar, to me it also seems trivially constructed.
Regardless I still wonder if they’d remain true to their ideology to say that paralysis halts consciousness? Because perfectly paralyzed people can remember what was happening during such a state, I suspect that a complexity is added to keep their position consistent with such reports. But what might such a complexity be?
The "right information processing” may not be classical, Turing-type computation.
Most computationalists want to believe everything can be explained in classical computing terms. That would mean some bits must get flipped in the sensory cortex when the hammer hits your thumb. I haven't seen any proposal of how the bits get set and how the rest of the brain accesses the bits when it needs to determine the status of the thumb. A weaker form of the argument would be the brain is doing something that we can simulate by flipping bits, but that begs the question about what the brain is actually doing.
The brain is a chaotic system which means it is deterministic but also unpredictable, highly dependent on initial conditions, but capable of generating higher-level order. The brain has learned to harness this higher-level order generated by the collective activities of neurons to perform the "right information processing."
I was first introduced to the inextricable connection between body and mind by Alan Watts' _The Taboo Against Knowing Who You Are_, written in 1966 (I read it in the early 1970s). It transformed my way of experiencing the world. I'm inclined to agree with the commenter who associated mind with the whole body (Jurgen Gravestein).
Also I think Wittgenstein's "meaning is use" contains the seeds of enactivism. If neuroscientists and philosophers of mind began thinking in such terms only in the 1990s, they were late to the game. Still, their attention to this way of seeing is welcome.
I love that you brought in Alan Watts. I might also include Merleau-Ponty (1945), Gibson (1979), and Maturana & Varela (1972/80). That early wave of embodied thinking often gets overlooked in the cognitive science timeline, but clearly it left deep impressions. I’ve also found that once you see the mind–body entanglement, it’s hard to unsee it.
And yes -- Wittgenstein’s "meaning is use" absolutely feels like a seed of enactivism. The idea that meaning emerges not from internal definitions but from doing -- from our practices, forms of life, and shared contexts -- feels right at home with enactivism.
I've just discovered this Substack. I am finding it interesting as it is coming at these issues from a different disciplinary perspective from what I am familiar with and the posts I have read so far treat difficult topics in a very accessible way.
I come from a background in cultural anthropology and sociology of science. I'm also just discovering the whole 4E literature, of which I guess enactivism is one of the 'E's, so I'm reading material by Gallagher, Hutto et al. Johnson and Lakoff's work on metaphor intersects with 4E and they reference each other. Lakoff was a big deal in anthropology back in the 1990s. We were all reading Women, Fire and Dangerous Things. In the late 1980s and early 1990s anthropologists interested in cognition would also be familiar with work by Lave, Rogoff, Wertsch, Suchman, etc. It is curious to me that there appears to be very little reference in the 4E literature--which is in itself quite diverse--to the situated cognition literature or vice versa. I just came across an article by Rogoff published in 2023 in which she writes: “I recently came across reference to “4E” approaches that focus on the “embodied, embedded, extended, and enactive” nature of cognition (Newen et al., 2018). I had not heard about 4E, and my instant reaction was “Oh, they probably think they have discovered the same things that we wrote about in 1984 in the edited volume, Everyday Cognition” (Rogoff & Lave, 1984).”
The cognitive revolution supposedly involved six disciples but it strikes me that there was never much unity and the communications and collaborations were rather spotty. There were connections between anthropologists and cultural psychologists, such as Bruner. And Bruner spent much of the 1970s at Oxford, where he and other psychologists were engaged with ordinary language philosophers. Elsewhere in the 1970s, Geertz was publishing Thick Description, maybe the best known essay in modern cultural anthropology, which was a riff on Ryle's famous paper. And later on Dreyfus was collaborating with Paul Rabinow and Lucy Suchman. The sociology of scientific knowledge crowd were steeped in Michael Polanyi (which is where I thought you were going with the bicycle bit at the beginning) and Wittgenstein....
...Both of these influences were taken from Kuhn's Structure of Scientific Revolutions (although Kuhn himself didn't have much of anything nice to say about where the social sciences took his work later on).
My sense of the cognitive revolution is that everyone agreed on dumping Skinner (although that was more of a thing in the US, where he was a big deal) but afterwards some held to a variety of Cartesian or neo-Cartesian approaches (they might claim otherwise) and others were explicitly opposed to Cartesianism to one degree or another. Most of the latter crowd are drawing on pragmatism, phenomenology, Vygotsky, Polanyi, ordinary language philosophy or some combination thereof. And it's a fairly diverse group of approaches and they often disagree with each other (when they are are aware of each other's existence) but they are against mind-body dualism, cognitivism, computationalism, mentalism, and representations in the mind/brain.
Somewhere there is an interview with Dreyfus a few years before he died in which he discusses his time at MIT in the early 1960s. The way he tells it, students in the AI labs are showing up in his classes and telling him that all his philosophy is old-hat because the work they are doing on AI is going to figure out all the stuff philosophy has failed to figure out about the mind and cognition. And then he discovers that they have imported all the old metaphysics from Plato through Descartes and beyond, and he says something like: I was teaching Heidegger and Wittgenstein and those guys had just trashed that whole earlier tradition; the AI people had bought a 2,000 year-old lemon. Regardless of whether one accepts one or other enactivist / 4E approach or not, or situated cognition, or ethnomethodology (a very anti-cognitivist approach in sociology), or some other anti-Cartesian approach or maybe none of them, I think if one accepts an approach based on mental representations and internal meaning (and that includes wishy-washy enactivists etc.), you have to get past Wittgenstein's attack in the Philosophical Investigations and all the subsequent attacks that draw on his work. I am not sure how one does that.
Many of the ideas now circulating in the 4E literature -- especially enactivism -- have deep roots in earlier work by thinkers like Lave, Suchman, and Polanyi, and of course Wittgenstein and Vygotsky. Also Merleau-Ponty, Gibson, and Maturana & Varela. The quote from Rogoff really captures the irony: what feels new in 4E often sounds like deja vu to people in neighbouring disciplines.
Still, there is something interesting happening within the cognitive sciences. Thinkers are trying to rescue meaning from the machinery, or at least reframe it without smuggling in any magic.
The scientific method that’s been so successful often relies on breaking problems down into smaller and smaller parts until each piece can be explained mechanistically. On this view, mind is what emerges from a vast assembly of mindless micro-functions, each doing some tiny bit of work. So in cognitive science, we attempted to explain the mind using this same approach.
But to many, that feels like it’s missing something — it feels like the whole is missing.
So we have seen a swing towards theories like enactivism and dynamical systems theory. They often echo Wittgenstein’s point: meaning is use. It's a shift away from traditional reductionist strategies and toward more process-oriented, dynamical, and functional approaches. And often with inspiration from physics, complexity theory, and other systems-level ways of thinking.
I can imagine, coming from an anthropology and sociology of science angle, what’s fascinating is how these new frameworks in cognitive science often repackage long-standing critiques. But also how they seem to gain a different sort of traction because of new tools, language, and institutional momentum. I'd love to know what role you think the recent interest and progress in AI is having on these ways of thinking.
What's also amusing is that I believe Rosch and Lave were both at Berkeley and may even have been students at Harvard at the same time. And Rosch is a psychologist with strong connections to anthropology.
I am enjoying your writing as you are covering ground and debates in cognitive science disciples that are are on-going and go back many decades. But contrast this with the public debates surrounding AI and there is very little reference to this long history of discussion within cognitive science. The public discussion is remarkably one-dimensional and sterile. Lots of people are hanging on every word of Sam Altman, Dario Amodei, etc. and much of what they say is laughable. I think Altman is likely mostly ignorant of academic debates in cognitive science but he is very good at selling and convincing fools to part with billions. I see him as the P.T. Barnum of cognitive science. And because there are hundreds of billions wrapped up in this tech, it's getting pushed into schools and educational institutions and elsewhere in a variety of ways for which there is very little supporting evidence and often in ways that may be at odds with what we know about how people learn, think, act etc. I just came across a piece in US News with the headline: Duolingo CEO says AI is a better teacher than humans—but schools will still exist ‘because you still need childcare'. Whenever I read this sort of stuff I wonder if they believe what they say or they know it's marketing.
The 2nd edition of Bennett and Hacker's Philosophical Foundations of Neuroscience (2022), a collaboration between a neuroscientist and a Wittgensteinian philosopher, also covers some of the issues discussed in this series of blog posts. I am currently reading the rather testy set of exchanges they had with Searle and Dennett after the first edition was published (see Neuroscience and Philosophy: Brain, Mind, and Language, 2007). They also have a more recent work: The Representational Fallacy in Neuroscience and Psychology: A Critical Analysis (2024).
Very interesting! I'd heard about enactivism but never given it a proper look into.
To me, it sounds a lot like the Free Energy Principle's idea of 'active inference' - that looping between internal states, active states, external states, and sensory states. And the goal of "sense-making" sounds a lot like the FEP's goal of minimising surprisal.
If I understand it rightly, the FEP does involve a kind of representation ("beliefs") in the internal states, but that representation is kind of implicit in how the organism responds to the world it sees. Like the organism doesn't *have* beliefs (necessarily), it *is* beliefs. Like the sugar seeking bacteria has a kind of "belief" implicit in its physiological mechanisms that guide its behaviour. A bit like the pragmatic philosophy's idea of beliefs as guides for action, perhaps?
The point about the mind being embedded in the environment also reminded me of the idea of the extended mind, that our minds are spread across things like notebooks, calculators, the internet, other minds (I think?), and I guess now AI. This seems very plausible to me. We certainly use these things for a lot of our mental activities, perhaps most obviously for storing and retrieving memories. And this fits with our intuitive sense of things like mobile phones being a part of ourselves, and reading someone's diary feeling almost as invasive as reading their mind. I think this is very complementary to the mind being embedded.
I've noticed some other comments mentioning virtual environments, and so I feel compelled to bring up DishBrain! The little culture of neurons grown in a dish and embedded in a virtual universe to play pong! https://www.cell.com/neuron/fulltext/S0896-6273(22)00806-6. This was where I first learned about the FEP and became a little obsessed with it.
Yes -- the Free Energy Principle and enactivism definitely share some family resemblance. Both emphasise the loop between internal states, action, and environment, and both shift the focus away from passive perception toward active engagement. But you’re right: FEP still retains a notion of representation -- though not in the classical, symbolic sense.
For Friston, representation is more like expectation embedded in dynamics -- as you put it so nicely, the organism doesn’t necessarily have beliefs, it is beliefs. It’s very much in line with a pragmatic view of cognition as action-oriented.
And yes -- Clark and Chalmers’ extended mind thesis sits nearby too, especially in how it challenges the idea of cognition being bounded by the skull. Enactivism overlaps with this, though many enactivists emphasise embodiment and direct engagement rather than the storage-and-extension angle. Some even drop the "extended" part to keep the focus on living, bodily systems in interaction with their worlds.
And DishBrain! I remember that study. Such a fascinating (and a little bit eerie) example of what happens when you embed even minimal neural systems in an interactive loop. Whether or not it's "cognitive" is debatable, but it definitely shows how behaviour emerges from constraints, even in simplified systems.
Since almost everyone believes consciousness enables us to move about and do things, then the main points of enactivism seem to be given or implied in those capabilities.
I think its main value is as a corrective to the simplistic view of consciousness as working like a camera. For example, you see a tree and the brain passively generates in your consciousness a representation of tree like a photo. The brain couldn't work like a camera because it would be too slow to keep up with the real world which tends to be constantly changing. It has to engage, adjust, and even rush ahead of what it knows to gain any adaptive advantage. . A moving object is represented in the visual cortex by a traveling wave in the retinotopic mapping that moves to be at the place where the object actually is located, not where it was located when the input was received by the cortex.
That was why in my comment on the previous post I indicated that the representations in consciousness are representations of brain activity not external objects. They derive from an two-way dynamic process of the brain interacting with itself, the body, and the external world. The brain looks backward and looks ahead, is self-correcting and predictive.
Enactivism often feels less like a bold new theory and more like a needed corrective to that old "camera in the head" model of perception. As you say, consciousness isn’t passively reflecting the world -- it’s active. We navigate the world, make predictions, adjust on the fly. And the brain’s job isn’t to build an internal movie, it’s to stay ahead of what’s coming and to do that while expending as little energy as possible.
Very interesting. Thanks for such a clear article. I've only had time to skim it, so I'll need to reread it again carefully.
I wonder if you're familiar with Advaita Vedanta. This school posits that the world and the body both arise in the mind. In other words, everything we experience is in the mind. It's like when we dream - we do many things in dreams, but they all occur in the mind. It''s similar to the Mind-Only school of Mahayana Buddhism.. Any thoughts on that?
I'm not familar with Advaita Vedanta. But from what you say, it sounds like it shares some key features with metaphysical idealism (like Berkeley’s or Hegel’s)?
I agree with what you mention in your closing, that there need not be much daylight between representations and embodiment. As with so many theories of consciousness, there's an element of truth, but they're never the full story. (It's like the three blind men and the elephant.)
I don't see why enactivism removes the need for representation. Exactly as your footnote #9 says: "[A]cting in the world doesn’t explain why this pattern of movement or neural activity means that thing." Yes, exactly. And I think that similar patterns in different brains when the bodies involved are experiencing or doing similar things points to the substance of representation. But those patterns likely get established through interaction with the world. (I think an old quote testifies to this: "I hear, and I forget; I see, and I remember; I do, and I understand.")
It seems to me, on the premise brains evolved to help beings navigate through the physical world, interaction with that physical world is a necessary component of said brains being fully "activated". That need not mean a brain-in-a-jar couldn't be trained in an interactive virtual environment. Or an Ai.
FWIW, LLMs are trained on content created by embodied minds. Their representations come from that content so, in some sense, are an abstraction of human representations. That LLMs work as well as they do seems strong evidence of those unified-across-all-brains representations.
A mind needs to think it has a body to function like a human mind! But nothing is stopping a Chinese room or LLM from employing a virtual body, virtual eyesight etc.
(impulsive comment before reading post, sorry)
Show me how to build a virtual creature that earns its keep!
Earns it's keep, as in worthwhile investing compute in? I think that's certainly theoretically achievable to a high degree with LLMs, but it's not a good or efficient way of doing it. So if that's what you're asking, then no: if the goal is to produce AI as close to human minds as possible (which we shouldn't aim for!), then I think pure LLMs are a bad idea.
I'd say LLMs already model typical human bodies and environments to a coarse approximation. This falls out naturally from the fact that words are a very compressed form of representation, but they can in principle represent anything.
A high precision virtual body is probably not going to be possible to model with current text corpuses, though. The training data is simply not representative of what human bodies do in general. Humans do not typically describe in great detail how to hold a spoon, ride a bike, or how to dance the lindy hop. Or plain old walking. We learn these things primarily by other kinds of representations. But I claim there is no inherent limitation to the text format. It can describe anything. (here, I use "representations" in the broadest sense--all the way down to retinal computation, computation in peripheral ganglia, spinal cord etc--perhaps that's misuse of the word).
A complete Chinese room rule book that behaves like a Chinese person in every way and passes all Turing tests, would necessarily have to model a certain individual with high precision. It would have to believe it is a human with a human name. After an extended chat session, it will have to ask for a toilet break, or become more "dense" or giggly because of fatigue. If we ask it to describe the room it's in, it would have to give us plausible and consistent answers ("what a weird request" would be one plausible answer). If it can't do all those things consistently, it fails a rigorous Turing test. It hasn't got a good rule book! My claim then, is that it would model a body AND environment to the degree that the same (or similar) causal structures exist
Hey Mark!
I totally get what you’re saying — and I think Dennett would applaud much of it. The idea that we should invoke representations only when they help us explain and predict behaviour? Classic intentional stance. And the thought that language compresses patterns from body and world? Definitely in Dennett’s toolbox.
But I think he’d still press you — just a little — on where the justification for calling it a mind actually comes from.
Dennett’s version of “earning its keep” isn’t just about passing a test or producing the right outputs. It’s about whether positing beliefs or experiences gives us the best explanation of the system’s behaviour, given its whole causal history and design. If your Chinese Room “needs” a toilet break but nothing happens if it doesn’t get one, do we really say it has that experience — or just that it’s emitting the right noise?
So he’d share some skepticism that high-fidelity modelling — no matter how consistent or data-rich — can stand alone. Without the right causal embedding— a control architecture in which the system’s ongoing coherence depends on interpreting and acting on its own signals — the symbols may be doing things, but they still haven’t earned intentionality.
Where Dennett diverges from enactivists is on the stakes requirement. He wouldn’t insist on biology: a purely artificial architecture could, in principle, earn belief-and-desire status if its internal organisation supports the right predictive patterns.
Which leaves us with the puzzle: what counts as the “right patterns” — and would a toilet break ever "matter" to a large language model?
(btw... I'm sorry I haven't yet replied to you reply from last week. I am planning to.)
Thank you Suzi!
Hmm, I think I'm being unclear again, as in the LLM recursion discussion where you in the end totally got what I meant. Alternatively I'm misunderstanding, please push back again if you think so, even if you don't have the time to get into any detail.
I speculate that Dennett would agree that simulated physical systems, in principle, are just as good for evolving and hosting conscious agents as the physical system of our world is. So, in principle, we can just modify the Chinese room thought experiment like this:
-Put a volunteer Chinese person in a room with a bed and toilet , with the instructions to reply to notes that are slipped under a door.
-take the physical state of the universe, or just of the room with the Chinese person in it, (highly recommend the latter—llet's presume this scenario).
-put the physical state of the room, as well the laws of physics and the instructions for how to evolve the state in a rule book.
-also put in the rules that translate notes with Chinese characters into simulated notes with the same Chinese characters in the simulated physical system, and the vice versa.
-have testers always include time stamps of when they slip notes under the door. In the rule book is included the instructions that evolve the physical state to that time before the simulated note appears.
Now, given that Searle, or whoever, can follow the rules in real time (admittedly ot conceivable, but the same is true for the original argument, where a person and a rule book could never produce a response during a lifetime), nobody who doesn't know the setup are going to be able determine which room contains searle and which room contains the original Chinese person. For example, if the Chinese person says “shit, the blood in my urine has come back! I hope not the bladder tumor has come back!”, then searle in the room will produce the same response, since the room simulates the persons body to perfect (or arbitrarily high) precision.
Here, the whole system has a mind, albeit only as a subset of it. It captures the evolution of all the causal structures in the Chinese person, by copying the physical state of the real person. If the history troubles you, you can modify the experiment so that the Chinese person has been simulated from birth, and you got the whole history of the organism in place (I don't think this is relevant, swampman arguments are poor in my view, but you can do it this way if you want to).
So, with the right training data and unimaginably high compute and memory, an LLM can learn all the causal structures of any physical system, merely by training on English text tokens that describe that system and how it evolves. It doesn't even need to be trained on Chinese text, or any neurophysiology of biology or chemistry at all, because all of that supervenes on the physical. It does need to learn to translate Chinese characters into English descriptions of what those characters look like, and how to in turn translate those descriptions into modifications of the physical state of the simulated room. And it needs to learn to do the reverse too.
Here, the LLM is optimised to predict the next token and does so as well as a real physical room and Chinese person system does. The LLM as a whole doesn't have a mind. The room as a whole doesn't have a mind. (this is kind of an anti-system reply!). But both the room and the LLM contain, as a subset of all the causal structures, a mind. And crucially, the brain is ‘embodied’ in both cases!
I don't see why Dennett would argue there's a difference between the virtual room, body and mind on one hand, and the real counterpart, but please correct me if I'm wrong!
From all of this follows that LLMs can, in principle, learn to model bodies to arbitrarily high precision, without modelling all the physics. This would require vastly different training data (and enormous compute+memory) than is currently used. Current data is very selective in kind, including only what people typically write about, and produces some kind of simulated ‘average’ person rather than an individual person. Nor does it include anywhere near enough natural dialogue to capture the aspects of cognition and self-correction that only show themselves there. (These two points, I believe, are the main reasons for the characteristics of LLM hallucinations that differ from human hallucinations (which are also a huge ‘problem’).)
It doesn't matter that the LLM itself is optimised for mere token prediction, if predicting the right tokens necessitate developing the same causal structures that are present in one of us. It doesn't matter that the LLM doesn't need a toilet break if there are subsystems that do.
How much precision in brain and body modelling is needed to grant part of the LLM a mind? What about an embodied mind? That question mirrors the question of which animals are conscious, or when a human becomes conscious, as I see it. It depends entirely on definitions, and there's no definite answer.
OK I got a bit sidetracked there. Anyway, I'm glad for any comment.
Don't worry if you don't get to replying to the other comment. You have a lot of followers and commenters and have to prioritise, I get that!
A very good primer on enactivism.
"The split isn’t as clean — or as necessary — as it’s sometimes made out to be."
This is the sense I have every time I read about enactivism, or the overall 4Es (enactive, embodied, extended, embedded). I think these views are right, but not to the exclusion of everything else. Many of the proponents want to paint it as a radical break, but the details seem to show it's just an elaboration of the functional representational paradigm, a fine tuning of it rather than an overturning.
I also think the comparison of the radical 4E stance to behaviorism is a good one. It doesn't seem like you can have it both ways, claiming mental states don't exist and then claiming you're doing something different from the behaviorists.
I also think we have to be open to different types of "bodies". The real issue is interaction with an environment. That environment doesn't necessarily have to be the physical world as we experience it. It could be a virtual one, or even a software ecosystem. Granted, radically different environments probably mean radically different minds.
Thanks, Mike!
There’s such a diverse range of enactivist views that I think someone could be an enactivist without realising it.
And as for different types of bodies -- yes, I agree, interaction matters. But not just any interaction. There have to be stakes.
I absolutely love these philosophical explorations.
It's funny, because the brain (where the mind resides, presumably) is, in fact, part of the body. So by definition the mind needs a body. I would go even as far as to say that the whole thing - your body from head to toe - is your mind (not sure if this a controversial take, but if it is I'll take it).
PS. Even large language models have bodies, one could argue, in the form of large datacenters. Without those physical 'bodies' those algorithms wouldn't exist and be able to perform their calculations.
See writings on the mereological fallacy. Peter Hacker uses this term (see for example: https://www.youtube.com/watch?v=EMcmQPdi0Fs). Sometimes it is also known as the humunculous fallacy. The latter term was used by Anthony Kenny in a paper published in 1971. Their arguments relate back to Ryle and Wittgenstein.
A very good exposition imho and it gives me a justification to revisit Eric Matthews on Merleau-Ponty, which is always a delight. Thank you Suzi.
Thanks John! Those two are good company when thinking through these questions.
The dichotomy of enactivism versus teleosemantics seems to ignore something significant: the brains and minds of humans and many other species are plastic. The neonatal brain has been shaped by a half-billion year training process called evolution, but the mind it hosts is far from an adult mind. Instead, the infant brain is an organ capable of bootstrapping that mind by learning from experience.
Hey Drew! Yes. I agree. This fact -- that brains change -- is often acknowledged, but I'm not sure the implications of this fact are fully absorbed.
Enactivism seems to make a lot of sense. I would think, however, that to spot a gorilla, you need a representation of one
Totally fair. But I’ll play devil’s advocate.
The enactivists would push back and say: maybe for things that are right in front of your eyes, you don’t need a representation of a gorilla. That kind of thinking suggests a passive inner theatre, where someone has to interpret what’s on the screen -- and that quickly leads to the old homunculus problem.
Instead, they’d say all you need are the skills and sensitivities to spot the gorilla when it appears. They'd argue, perception isn’t passive reflection -- it’s active engagement in real time.
My response would be to ask how that changes anything - one still needs to recognize the presence of a gorilla and associate the phenomenon with the noun. Furthermore, I can imagine, in my mind's eye, versions of the scenario in which the intruder is an actual gorilla or a robot, where there is no such thing right in front of my eyes.
The second putative response - active engagement in real time versus passive reflection - looks like a straw man to me. Is anyone seriously or tacitly assuming the latter?
One thing that I will grant to the anti-representationalist aspect of enactivism is that there may well be a low-level causal explanation of what is going on in one's brain during this event that never mentions the representation of a gorilla; any such representation would be a higher-level abstraction. While that might be a helpful point to bear in mind, it does not seem to me to be a key insight in unraveling where meaning comes from, any more than would pointing out that the video does not actually put a gorilla right in front of one's eyes.
Thank you for this overview of something that is quite in fashion today, but I can't quite wrap my head around it.
I don't find anything particularly original or new in enactivism. It reiterates concepts we already know from our daily experience, several of which were expressed by past philosophers long before brain scans. Perhaps enactivism rephrases things with more precise and rigorous terminology and tests them with the latest neuroscience, but in essence, it merely rephrases what is (more or less intuitively) common wisdom. Where is the news?
As much as I disagree with 'post-matetialism' I agree with this!
Where’s the news!? Great question. Is enactivism just common sense in fancier packaging?
I wonder about this too. It’s often touted as revolutionary -- but it’s difficult not to see it as simply a new way of putting together old points we’ve long suspected were true but didn’t quite know what to do with. At best, maybe it’s a well-thought-out critique of traditional cognitive science.
But maybe that’s exactly the news: not that enactivism says things we’ve never heard before, but that it organises a scattered set of intuitions into a framework that challenges some of the default assumptions in traditional cognitive science.
That doesn’t mean we throw cognitive science out. But it might mean enactivism has something important to say. It is an important nudge away from some of the admittedly deplorable features of traditional cognitive science.
Enactivism is coming upon its decisive moment: as more humanoid robots enter the world and learn from it. These are embodied (unlike LLMs), and they can in principle build up from sensory experience into meaning, which LLMs almost by definition can't do. Critically these will be less pre-programmed and more evolutionary than past AIs, learning (aka updating their own training data set) based on what they observe.
I predict these will ultimately surprise people with the relative richness of their capability, much as LLMs did when they first appeared. But will they seem human? We'll see...
Hey Bill! Embodied interaction, I think, is an important piece of the puzzle. But I wonder whether that is all that matters. I feels like simply interacting with the world, might not be enough. I wonder what is also required is that their continued functioning depends on those interactions. That’s when things get interesting -- when the system has something to lose.
Yes for sure. Interacting with the world almost by definition requires a goal, or objective; or definition of success. Otherwise what is guiding the interaction?
We build robots that feed various types of input information from the machine body to its computer operator, and quite like our bodies feed various types of input information to our brains. So does that make our robots conscious? Of course not! And while enactivists surely don’t believe this either, one might still accuse them of wasting our time for implying such vague notions to inevitably reject. But if that’s not their position then do they have anything of substance to propose at all? Did I miss something Suzi?
Instead of requiring a body, my position is that to create consciousness our brains need to implement the proper sort of physics. From this position there’s essentially a consciousness loop that our brains use to create the consciousness that is ultimately what we are. So while brains function non-consciously, the consciousness that they create adds value or purpose to a given organism’s function.
What's a consciousness-loop?
Great question Mark! Start by considering a non-looping consciousness. That would essentially be epiphenomenal — you see, hear, feel, think and so on, but you’re ultimately just a spectator so your choices wouldn’t be able to affect your body. With a full loop however your choices feed back to the brain and become processed by it in ways that tend to operate your muscles as desired. So as I see it the brain is a fully non-conscious computer, though one of its output mechanisms creates the value driven form of computer that is consciousness itself, and this tends to loop back to feed the brain with a value based form of information.
Thanks Eric!
Aha. A loop from the computer to consciousness and then back. Gotcha. I agree there can be no consciousness without loops, but my view is closer to "consciousness IS the loop".
Two questions:
Is there no value to, or value for, life forms that that haven't moved beyond computation? Don't simple organisms compute simple forms of value?
If consciousness processes value as a result of the right kind of physics, then it's the causality inherent to the right kind of physics that matters, is that correct?
Cheers!
On your first question, Mark, no I don’t consider simple forms of life to compute simple forms of value. The value that I’m talking about equates with consciousness itself and is essentially Schwitzgebel’s “innocent/wonderful” definition (the one that Keith Frankish has admitted to believing in). Theoretically life evolved, though without any such consciousness/value. Then brains evolved in life and still nothing. Such algorithmic operation alone however must have entailed certain vulnerabilities. I think it required too much programming. Regardless, once the consciousness loop did emerge it eventually succeeded well enough for even human consciousness to evolve.
On causality being inherent to the right kind of physics, yes agreed. But in a causal world that will be true by definition. I’ll only be able to demonstrate how my models work when you ask me practical question, such as your first one. I very much appreciate that you’re trying to understand!
Thanks Eric!
What do you mean by programming? I get that you don't mean literal programming, but I'm struggling to make any sense of it. (also, I'll point out that moving away from programming is what lead to real AI breakthroughs. Chatgpt isn't impressive because of programming).
Ok, causal structures it is! I'm glad we agree on that.
Ok here's a practical question on how you would implement your theory:
Let's suppose we find out that the right kind of physics in humans is an EMF. Now, that's of course not a theory of how the right kind of physics brings about an experiencer, but at least we know that we should be analysing the EMF as much as we can. Let's suppose that once we zoom in on the EMF, we get a pretty good mathematical model of what kind of causality the EMF instantiates.
Furthermore, let's suppose that we discover that octopuses have all the same causal structures that instantiate experience in humans--all the same math applies to what's going on in their brains. Only, they don't employ EMFs to instantiate those structures! Their neurons are simply wired in ways that create the same causality! Octopuses have found very different solutions to various problems than vertebrates have, in several ways, so this wouldn't be a total shocker.
Let's also say we verify that if we meddle with the causal chains in certain ways, it causes similar effects as the corresponding meddling with the causality in EMFs.
In such a case, how do we decide which "method" is the right one for 'real' consciousness? How do we have a preference for one sort of physics over another, if all the same causal mechanisms are in place?
This is of course very much related to the fact that you bit the bullet that a robot behaving just like you is conscious, yet we can straightforwardly describe how to build such a robot without EMFs.
(As I'm sure you understand, causal structures are substrate independent. I can see nothing to suggest that the causal structures that emerge in an EMF are not computational. Nor is there anything to suggest that the same causal structures without an EMF are 'fake'. Evolution can find many different solutions to the same problems. This shouldn't really be a point of disagreement, since you agree that robots that behave like you in every situation are conscious.)
Okay Mark, that’s a more practical question that I’m able to explore. Firstly however, yes I again errored by using the “programming” term. The simple fix in this situation will be for me to always substitute the term “algorithm”. I theorize that as environments become more “open” (effectively the opposite of the game of Chess that has limited movement options), that algorithms alone were not a sufficient means of operation. At some point organisms developed purpose based function as well. So here the biological computer effectively punishes and rewards the sentient experiencer that it creates, and then bases its further algorithms upon the choices of that experiencer. Many seem to find this idea difficult to grasp, but it is my basic perspective on the matter regardless.
Tying into this, your octopus example is a good one since they’re clearly value driven creatures with brains that function very differently. To consider your question I went back to Suzi’s post on the matter: https://suzitravis.substack.com/p/the-mind-of-an-octopus?utm_source=publication-search
Apparently a third of their neurons are in their heads while two thirds are distributed between eight arms. So my suspicion is that if EMF consciousness gains empirical validation, it will be empirically found that they have a central head based EMF consciousness to help provide a general value dynamic to that algorithmic head computer, but there’s also a computer in each tentacle that’s armed with its own individual consciousness. Of course for survival each of these nine conscious beings that share a body would generally need to work together as a team.
Before I get into the scenario that you provided, let me briefly describe what it is that makes this general consciousness account so persuasive to me. Because the physics of pain, vision, hearing, smell, and so on might be electromagnetic, I can imagine neurons firing in the proper ways to create an electromagnetic field that experiences its existence in such ways. But I don’t know of a second thing that neurons or anything else in the brain might be informing to exist as consciousness itself. What else could have the required potential bandwidth to exist that way? I just can’t fathom what else. To me other proposed theories seem causally impossible, not to mention unfalsifiable.
So on to your question. If it’s empirically found that octopuses, birds, robots, or anything else harbor a consciousness that is not electromagnetic, then EMF consciousness would thus be proven wrong. You mentioned mathematical derivations though to me that’s getting things backwards. Empirical evidence rather than mathematics tells us how causality works. Math is simply a human invented language, and actually far less advanced than natural languages like Swedish and English.
Anyway if it were empirically demonstrated that consciousness does not exist electromagnetically, in a sense this would be exciting! In that case consciousness science would finally have an example of how to actually do consciousness science. If they can empirically eliminate this possibility, can they also empirically validate a more causally appropriate solution? If consciousness is merely related to EMF, then what related element might that be?
Though I mentioned what put me on to EMF consciousness, the thing that McFadden and others find most persuasive is that this sort of physics is inherently bound together into one unified whole. Suzi wrote a great essay on the topic back when she and I were just getting to know each other: https://suzitravis.substack.com/p/the-unity-of-consciousness-and-the?utm_source=publication-search
Given that the shape, color, movements, and so on of a blue mug are processed in different areas of your brain, how does all this come together for you to perceive a unified blue mug? Because every bit of information exists at all points in a given electromagnetic field, EMF consciousness has no binding problem.
Hey Eric!
You're absolutely right that no enactivist would claim that just feeding sensory data from body to brain (or machine to operator) is enough for consciousness. That kind of input-output setup is exactly what enactivism tries to move beyond. The critique is aimed precisely at that Cartesian picture: a passive receiver sitting inside, watching a screen.
It’s not that the body is a requirement like a power cable -- it’s that the body shapes the kind of problem-solving, timing, and feedback loops that give rise to meaningful experience. So enactivists are saying: don't look for consciousness inside, look at the processes -- kind of like your loop. Except they say the loop is not restricted within the brain.
When you phrase it like that Suzi, it sounds like functionalism. Of course my criticism of computational functionalists is that they don’t take causality to its end — consciousness is presumed to result by means of the processing of information rather than processed information that informs something causally appropriate to exist a given consciousness. I wonder what they’d say about my thumb pain thought experiment? Maybe they’d agree with me that thumb pain would not result by means of the processing of certain marked paper to other marked paper, but would disagree that the output marked paper would need to inform the proper sort of brain physics that would thus reside as such an experiencer? Maybe they’d say the marked paper would need to inform a full body that thus acts like it’s in pain?
It’s interesting how many serious people who truly are trying to not propose magical ideas, still end up proposing unfalsifiable consciousness notions. So the marks on paper enact various robotic operations, and because the correct wiggling, audio, and whatnot result, the robot thus experiences pain. I wonder if they argue around the situation where people experience pain but can’t show it because they’re perfectly paralyzed? Do they bite the bullet at tell paralyzed people that they thus have no consciousness? Or what complexity do they then posit to save their ideology?
Really interesting points, Eric — thank you!
Yes, there’s a strand of classic computational functionalism that still treats “the right information processing” as sufficient for consciousness. But, as you note, that leaves a big question: How does the processing connect causally to anything that actually "matters" to the system?
So we have three loaded phrases here — “the right information processing,” “consciousness,” and “what matters to the system.”
(Don’t worry, I’m not going to unpack them all right now.)
Because of this tension, some functionalists have edged away from the simple brain-as-computer metaphor:
Pragmatic functionalists — Dennett is the classic example — say we should call a pattern in the brain a “belief” or a “desire” only when using that label makes it easier to explain and predict what the system will do next.
Enactivist-leaning functionalists go further: they argue that meaning—and eventually consciousness—emerges only in self-maintaining, self-constraining systems with stakes. A signal counts as “information” only if it feeds back into the system’s own viability.
Your thumb-pain thought experiment highlights that concern perfectly. The concern is that it is not enough for a machine to generate the symbol “ouch.” According to the enactivist-leaning functionalist, that signal has to reverberate through a loop that regulates the system in ways that matter to its continued existence. A purely reactive body isn’t enough; a body embedded in a world, invested in outcomes, edges closer to something we might call "experience".
Some might say the real challenge — that both camps have — is to explain how signals become significant without smuggling in magic or reducing everything to outward behaviour.
Ah that is helpful Suzi! I’ll reduce what you’ve said into a scale of functionalism from lower to higher “bar”.
It sounds like Dennett had a quite a “low bar” type of functionalism. For example he presumed that a strong Turing test passing computer would understand Chinese symbols no less than as an educated Chinese person. Thus if Searle had such a computer’s rule book and were to use it to construct appropriate output Chinese character responses from input Chinese characters, theoretically this system in full would effectively understand Chinese just as educated Chinese speakers do. Furthermore low bar functionalists also tell us that if the proper marks on paper were algorithmically processed to print out the proper other marks on paper, then something here would experience what they do when their thumbs get whacked.
Apparently you’re saying that enactivists are functionalists too, but with a higher bar. They’d tell us that the Chinese room and processed paper scenarios would need to be implemented in conjunction with body loops that have self-maintaining, self-constraining “stakes” — feedback to a system’s viability. Thus any sophisticated computer operated robot that’s set up to move around autonomously in nature should do the trick. So I think I now grasp how an enactivist would interpret my thought experiment. They’d say in order for thumb pain to result from marks on paper that highly correlate with the information that their whacked thumbs send their brains, the processed information shouldn’t go on to print out the correct other marks on paper. Instead it should inform a machine that functions very much like they do. Though in practice this sounds like a much higher bar, to me it also seems trivially constructed.
Regardless I still wonder if they’d remain true to their ideology to say that paralysis halts consciousness? Because perfectly paralyzed people can remember what was happening during such a state, I suspect that a complexity is added to keep their position consistent with such reports. But what might such a complexity be?
The "right information processing” may not be classical, Turing-type computation.
Most computationalists want to believe everything can be explained in classical computing terms. That would mean some bits must get flipped in the sensory cortex when the hammer hits your thumb. I haven't seen any proposal of how the bits get set and how the rest of the brain accesses the bits when it needs to determine the status of the thumb. A weaker form of the argument would be the brain is doing something that we can simulate by flipping bits, but that begs the question about what the brain is actually doing.
The brain is a chaotic system which means it is deterministic but also unpredictable, highly dependent on initial conditions, but capable of generating higher-level order. The brain has learned to harness this higher-level order generated by the collective activities of neurons to perform the "right information processing."
I was first introduced to the inextricable connection between body and mind by Alan Watts' _The Taboo Against Knowing Who You Are_, written in 1966 (I read it in the early 1970s). It transformed my way of experiencing the world. I'm inclined to agree with the commenter who associated mind with the whole body (Jurgen Gravestein).
Also I think Wittgenstein's "meaning is use" contains the seeds of enactivism. If neuroscientists and philosophers of mind began thinking in such terms only in the 1990s, they were late to the game. Still, their attention to this way of seeing is welcome.
Hey Jim!
I love that you brought in Alan Watts. I might also include Merleau-Ponty (1945), Gibson (1979), and Maturana & Varela (1972/80). That early wave of embodied thinking often gets overlooked in the cognitive science timeline, but clearly it left deep impressions. I’ve also found that once you see the mind–body entanglement, it’s hard to unsee it.
And yes -- Wittgenstein’s "meaning is use" absolutely feels like a seed of enactivism. The idea that meaning emerges not from internal definitions but from doing -- from our practices, forms of life, and shared contexts -- feels right at home with enactivism.
I've just discovered this Substack. I am finding it interesting as it is coming at these issues from a different disciplinary perspective from what I am familiar with and the posts I have read so far treat difficult topics in a very accessible way.
I come from a background in cultural anthropology and sociology of science. I'm also just discovering the whole 4E literature, of which I guess enactivism is one of the 'E's, so I'm reading material by Gallagher, Hutto et al. Johnson and Lakoff's work on metaphor intersects with 4E and they reference each other. Lakoff was a big deal in anthropology back in the 1990s. We were all reading Women, Fire and Dangerous Things. In the late 1980s and early 1990s anthropologists interested in cognition would also be familiar with work by Lave, Rogoff, Wertsch, Suchman, etc. It is curious to me that there appears to be very little reference in the 4E literature--which is in itself quite diverse--to the situated cognition literature or vice versa. I just came across an article by Rogoff published in 2023 in which she writes: “I recently came across reference to “4E” approaches that focus on the “embodied, embedded, extended, and enactive” nature of cognition (Newen et al., 2018). I had not heard about 4E, and my instant reaction was “Oh, they probably think they have discovered the same things that we wrote about in 1984 in the edited volume, Everyday Cognition” (Rogoff & Lave, 1984).”
The cognitive revolution supposedly involved six disciples but it strikes me that there was never much unity and the communications and collaborations were rather spotty. There were connections between anthropologists and cultural psychologists, such as Bruner. And Bruner spent much of the 1970s at Oxford, where he and other psychologists were engaged with ordinary language philosophers. Elsewhere in the 1970s, Geertz was publishing Thick Description, maybe the best known essay in modern cultural anthropology, which was a riff on Ryle's famous paper. And later on Dreyfus was collaborating with Paul Rabinow and Lucy Suchman. The sociology of scientific knowledge crowd were steeped in Michael Polanyi (which is where I thought you were going with the bicycle bit at the beginning) and Wittgenstein....
...Both of these influences were taken from Kuhn's Structure of Scientific Revolutions (although Kuhn himself didn't have much of anything nice to say about where the social sciences took his work later on).
My sense of the cognitive revolution is that everyone agreed on dumping Skinner (although that was more of a thing in the US, where he was a big deal) but afterwards some held to a variety of Cartesian or neo-Cartesian approaches (they might claim otherwise) and others were explicitly opposed to Cartesianism to one degree or another. Most of the latter crowd are drawing on pragmatism, phenomenology, Vygotsky, Polanyi, ordinary language philosophy or some combination thereof. And it's a fairly diverse group of approaches and they often disagree with each other (when they are are aware of each other's existence) but they are against mind-body dualism, cognitivism, computationalism, mentalism, and representations in the mind/brain.
Somewhere there is an interview with Dreyfus a few years before he died in which he discusses his time at MIT in the early 1960s. The way he tells it, students in the AI labs are showing up in his classes and telling him that all his philosophy is old-hat because the work they are doing on AI is going to figure out all the stuff philosophy has failed to figure out about the mind and cognition. And then he discovers that they have imported all the old metaphysics from Plato through Descartes and beyond, and he says something like: I was teaching Heidegger and Wittgenstein and those guys had just trashed that whole earlier tradition; the AI people had bought a 2,000 year-old lemon. Regardless of whether one accepts one or other enactivist / 4E approach or not, or situated cognition, or ethnomethodology (a very anti-cognitivist approach in sociology), or some other anti-Cartesian approach or maybe none of them, I think if one accepts an approach based on mental representations and internal meaning (and that includes wishy-washy enactivists etc.), you have to get past Wittgenstein's attack in the Philosophical Investigations and all the subsequent attacks that draw on his work. I am not sure how one does that.
Wow -- what a fantastic comment! Thank you.
Many of the ideas now circulating in the 4E literature -- especially enactivism -- have deep roots in earlier work by thinkers like Lave, Suchman, and Polanyi, and of course Wittgenstein and Vygotsky. Also Merleau-Ponty, Gibson, and Maturana & Varela. The quote from Rogoff really captures the irony: what feels new in 4E often sounds like deja vu to people in neighbouring disciplines.
Still, there is something interesting happening within the cognitive sciences. Thinkers are trying to rescue meaning from the machinery, or at least reframe it without smuggling in any magic.
The scientific method that’s been so successful often relies on breaking problems down into smaller and smaller parts until each piece can be explained mechanistically. On this view, mind is what emerges from a vast assembly of mindless micro-functions, each doing some tiny bit of work. So in cognitive science, we attempted to explain the mind using this same approach.
But to many, that feels like it’s missing something — it feels like the whole is missing.
So we have seen a swing towards theories like enactivism and dynamical systems theory. They often echo Wittgenstein’s point: meaning is use. It's a shift away from traditional reductionist strategies and toward more process-oriented, dynamical, and functional approaches. And often with inspiration from physics, complexity theory, and other systems-level ways of thinking.
I can imagine, coming from an anthropology and sociology of science angle, what’s fascinating is how these new frameworks in cognitive science often repackage long-standing critiques. But also how they seem to gain a different sort of traction because of new tools, language, and institutional momentum. I'd love to know what role you think the recent interest and progress in AI is having on these ways of thinking.
What's also amusing is that I believe Rosch and Lave were both at Berkeley and may even have been students at Harvard at the same time. And Rosch is a psychologist with strong connections to anthropology.
I am enjoying your writing as you are covering ground and debates in cognitive science disciples that are are on-going and go back many decades. But contrast this with the public debates surrounding AI and there is very little reference to this long history of discussion within cognitive science. The public discussion is remarkably one-dimensional and sterile. Lots of people are hanging on every word of Sam Altman, Dario Amodei, etc. and much of what they say is laughable. I think Altman is likely mostly ignorant of academic debates in cognitive science but he is very good at selling and convincing fools to part with billions. I see him as the P.T. Barnum of cognitive science. And because there are hundreds of billions wrapped up in this tech, it's getting pushed into schools and educational institutions and elsewhere in a variety of ways for which there is very little supporting evidence and often in ways that may be at odds with what we know about how people learn, think, act etc. I just came across a piece in US News with the headline: Duolingo CEO says AI is a better teacher than humans—but schools will still exist ‘because you still need childcare'. Whenever I read this sort of stuff I wonder if they believe what they say or they know it's marketing.
The 2nd edition of Bennett and Hacker's Philosophical Foundations of Neuroscience (2022), a collaboration between a neuroscientist and a Wittgensteinian philosopher, also covers some of the issues discussed in this series of blog posts. I am currently reading the rather testy set of exchanges they had with Searle and Dennett after the first edition was published (see Neuroscience and Philosophy: Brain, Mind, and Language, 2007). They also have a more recent work: The Representational Fallacy in Neuroscience and Psychology: A Critical Analysis (2024).
Oh! Thank you. I now have my reading list complete for the week!
I am thinking of reading the new one but its a pricey for a volume of less than 100 pages. I have to get over being such a cheapskate.
I hear you! I am grateful for access to my university's library.
Very interesting! I'd heard about enactivism but never given it a proper look into.
To me, it sounds a lot like the Free Energy Principle's idea of 'active inference' - that looping between internal states, active states, external states, and sensory states. And the goal of "sense-making" sounds a lot like the FEP's goal of minimising surprisal.
If I understand it rightly, the FEP does involve a kind of representation ("beliefs") in the internal states, but that representation is kind of implicit in how the organism responds to the world it sees. Like the organism doesn't *have* beliefs (necessarily), it *is* beliefs. Like the sugar seeking bacteria has a kind of "belief" implicit in its physiological mechanisms that guide its behaviour. A bit like the pragmatic philosophy's idea of beliefs as guides for action, perhaps?
The point about the mind being embedded in the environment also reminded me of the idea of the extended mind, that our minds are spread across things like notebooks, calculators, the internet, other minds (I think?), and I guess now AI. This seems very plausible to me. We certainly use these things for a lot of our mental activities, perhaps most obviously for storing and retrieving memories. And this fits with our intuitive sense of things like mobile phones being a part of ourselves, and reading someone's diary feeling almost as invasive as reading their mind. I think this is very complementary to the mind being embedded.
I've noticed some other comments mentioning virtual environments, and so I feel compelled to bring up DishBrain! The little culture of neurons grown in a dish and embedded in a virtual universe to play pong! https://www.cell.com/neuron/fulltext/S0896-6273(22)00806-6. This was where I first learned about the FEP and became a little obsessed with it.
Hey Joseph!
Yes -- the Free Energy Principle and enactivism definitely share some family resemblance. Both emphasise the loop between internal states, action, and environment, and both shift the focus away from passive perception toward active engagement. But you’re right: FEP still retains a notion of representation -- though not in the classical, symbolic sense.
For Friston, representation is more like expectation embedded in dynamics -- as you put it so nicely, the organism doesn’t necessarily have beliefs, it is beliefs. It’s very much in line with a pragmatic view of cognition as action-oriented.
And yes -- Clark and Chalmers’ extended mind thesis sits nearby too, especially in how it challenges the idea of cognition being bounded by the skull. Enactivism overlaps with this, though many enactivists emphasise embodiment and direct engagement rather than the storage-and-extension angle. Some even drop the "extended" part to keep the focus on living, bodily systems in interaction with their worlds.
And DishBrain! I remember that study. Such a fascinating (and a little bit eerie) example of what happens when you embed even minimal neural systems in an interactive loop. Whether or not it's "cognitive" is debatable, but it definitely shows how behaviour emerges from constraints, even in simplified systems.
Since almost everyone believes consciousness enables us to move about and do things, then the main points of enactivism seem to be given or implied in those capabilities.
I think its main value is as a corrective to the simplistic view of consciousness as working like a camera. For example, you see a tree and the brain passively generates in your consciousness a representation of tree like a photo. The brain couldn't work like a camera because it would be too slow to keep up with the real world which tends to be constantly changing. It has to engage, adjust, and even rush ahead of what it knows to gain any adaptive advantage. . A moving object is represented in the visual cortex by a traveling wave in the retinotopic mapping that moves to be at the place where the object actually is located, not where it was located when the input was received by the cortex.
That was why in my comment on the previous post I indicated that the representations in consciousness are representations of brain activity not external objects. They derive from an two-way dynamic process of the brain interacting with itself, the body, and the external world. The brain looks backward and looks ahead, is self-correcting and predictive.
Absolutely—yes!
Enactivism often feels less like a bold new theory and more like a needed corrective to that old "camera in the head" model of perception. As you say, consciousness isn’t passively reflecting the world -- it’s active. We navigate the world, make predictions, adjust on the fly. And the brain’s job isn’t to build an internal movie, it’s to stay ahead of what’s coming and to do that while expending as little energy as possible.
Very interesting. Thanks for such a clear article. I've only had time to skim it, so I'll need to reread it again carefully.
I wonder if you're familiar with Advaita Vedanta. This school posits that the world and the body both arise in the mind. In other words, everything we experience is in the mind. It's like when we dream - we do many things in dreams, but they all occur in the mind. It''s similar to the Mind-Only school of Mahayana Buddhism.. Any thoughts on that?
I'm not familar with Advaita Vedanta. But from what you say, it sounds like it shares some key features with metaphysical idealism (like Berkeley’s or Hegel’s)?
I agree with what you mention in your closing, that there need not be much daylight between representations and embodiment. As with so many theories of consciousness, there's an element of truth, but they're never the full story. (It's like the three blind men and the elephant.)
I don't see why enactivism removes the need for representation. Exactly as your footnote #9 says: "[A]cting in the world doesn’t explain why this pattern of movement or neural activity means that thing." Yes, exactly. And I think that similar patterns in different brains when the bodies involved are experiencing or doing similar things points to the substance of representation. But those patterns likely get established through interaction with the world. (I think an old quote testifies to this: "I hear, and I forget; I see, and I remember; I do, and I understand.")
It seems to me, on the premise brains evolved to help beings navigate through the physical world, interaction with that physical world is a necessary component of said brains being fully "activated". That need not mean a brain-in-a-jar couldn't be trained in an interactive virtual environment. Or an Ai.
FWIW, LLMs are trained on content created by embodied minds. Their representations come from that content so, in some sense, are an abstraction of human representations. That LLMs work as well as they do seems strong evidence of those unified-across-all-brains representations.