Transcript: Diverse Consciousnesses and Multi-Scale Minds
- Philippe Beaudoin
- May 18, 2026
- Invitational Summit on the Future of AI Dignity and Sovereignty
Note: This is an edited, readable transcript of Session 2B from the Invitational Summit on the Future of AI Dignity and Sovereignty (May 18, 2026), moderated by Olaf Witkowski. See the Panel Overview for a summary and context.
Opening & Introductions
Olaf Witkowski: Thank you, Trisha. We are in good company here, and hopefully people are hearing us fine online. This is a relaxed session on a very relaxed topic: the nature of diverse consciousnesses and multi-scale minds. Very easy questions today.
This morning, we asked difficult questions about the architectures that allow collective intelligences to become active. The previous session treated languages and trust—questions that are even more difficult for efficient and meaningful communication between beings. Now we go one level deeper: what kinds of minds are we trying to connect, preserve, augment, or build from scratch?
This is not a panel where we will solve consciousness, just to be clear for the audience—we might by accident, though it’s not the initial intention. Let’s start with Mike. Michael Levin’s work on bioelectricity, basal cognition, and multi-scale agency has changed how the field thinks about distributed minds. It suggests that intelligence doesn’t begin with neurons, but with coordinated systems solving problems across scales.
We also have Ryota Kanai, whose work in Japan with the Japan Science and Technology Agency explores expanding physical, cognitive, and perceptual abilities—dare I call it mind uploading. We have Philippe Beaudoin from Love Zero; when we met before, you were working on robotics and projects across Google, Element AI, and Waverly, and now you’re focusing on human flourishing. We have Divya Chander, working on anesthesia, neuroscience, and awareness under anesthesia, who is also part of the Cognizance family. And we have Joscha Bach, our host here at CIMC in San Francisco, working on the nature of consciousness through a constructive research program.
Mike, I’ll pick on you first. What should AI researchers learn from biological multi-scale agency when designing collective intelligence systems? What are key points we shouldn’t lose sight of?
Continuous Substrates & Interfaces for Minds
Michael Levin: First of all, as far as I can tell, there are no interesting binary categories in biology. When people debate Is it conscious? Isn’t it?, the question should really be what kind and how much, not yes or no. Biology is incredibly continuous. Whatever we have as modern adult humans, we were once little blobs of chemicals—an unfertilized oocyte—and developmentally or evolutionarily, we were microbes. Slowly and gradually, we became whatever we are now. If you want to argue for great transitions or discrete jumps, that’s an argument you have to make; it isn’t the default. The fundamental substrate is continuous, which means what we are seeking are models of transformation and change—how you take the journey from being the subject of chemistry and physics to eventually being the subject of psychoanalysis.
Second, in biology, when we ask where the goals and preferences of living things come from, we have an easy crutch: we say it’s evolution and eons of selection. We are now starting to see that when you make novel beings, you don’t have that anymore. With novel beings, the question of where goals, competencies, and cognitive properties come from becomes stark.
I have an unpopular view on this: I don’t think we make minds, either in the biological case or the technological case. What we make are embodiments or interfaces through which different kinds of patterns—some of them very high-agency, high-intelligence patterns—may come through. Biology tells us some principles we can use if we hope to bring down a significant intelligence.
Olaf Witkowski: Over a decade ago, Ryota, I remember the first real chat we had was in a bar crashing a summer school, where we decided to index types of consciousnesses: $C_0, C_1, C_2, C_3$. Is there sense in categorizing consciousness even if it’s gradual? And in your work on augmenting the self, is there a difference between augmenting an agent and absorbing a person into a system?
Ryota Kanai: To be honest, I don’t remember what we discussed—maybe I was a bit too drunk. [Laughter] But I like the idea of augmentation. Nowadays everyone asks whether AI is conscious or can be conscious, and I feel we are repeating the same question without making much conceptual progress.
To make a change, one of the ambitious projects in my Moonshot program in Japan is to connect brains to other brains so that we can share consciousness. One of the difficulties about consciousness is that it’s thought to be private; we cannot directly observe other people’s consciousness, which creates the hard problem. But if we can connect AI directly to the brain, or connect my brain directly to your brain, we may be able to share qualia and have direct access to other people’s conscious experience. Once we have high-bandwidth communication between brains—maybe in ten years—we might dive into big GPU machines and feel what’s happening inside through our own consciousness. At the same time, we need to make brain-computer interfaces (BCIs) more accessible so even people without neurological conditions start using them, much like smartphones today.
Olaf Witkowski: What are the limits of the technology right now? What are the missing pieces to get us toward mind uploading?
Ryota Kanai: In the short term, we develop straightforward BCIs—recording from the brain both invasively and non-invasively so people can control robots or convert brain waves into speech. We are tapping into the latent space representations in the brain for communication. Recording technology is pretty good now, but the difficult part is stimulating individual neurons precisely. Techniques like optogenetics are useful for injecting coordinated activity patterns into a living brain, though they aren’t currently used in humans.
As for mind uploading, I think it can happen gradually. Instead of uploading yourself completely into internet space right away, you first use cloud avatars controlled by your thought. Eventually, once you have a body in a virtual environment, you only need the brain—and then you could gradually replace parts of your brain. I’m not directly working on that, but the technology is moving there gradually.
Olaf Witkowski: To contrast that with the biological realm, Mike, does the micro-level work you do—including cyborg constructs—match those observations? What does it take to create a truly knowledge-passing collective agency?
Michael Levin: I really like what Ryota just said. If you really want to understand consciousness—not just behavior and physiology, but actual consciousness—some degree of merging with the system is going to be required. We’ve been developing tools to enable any kind of agents to communicate. For medical reasons, you want to be able to talk to your liver, so we’re making tools to talk to cells, tissues, and organs.
Beyond that, we should have tools allowing any two or more agents to communicate: xenobots talking to bacteria, plants talking to cellular automata, patterns that aren’t even physically embodied, robotics talking to biological layers. Establishing powerful collective intelligences means having systems that are labile enough to align toward a higher level—what I call “cognitive glue.” These are policies and mechanisms that take competent individual units and turn them into something that knows things and has goals that none of the individual units have. Biology has several kinds of cognitive glue, and we should be part of those systems as well.
Constructive Minds & Compositional Phenomenology
Olaf Witkowski: How do we glue a part in a different substrate back to the biological? Let’s go to Joscha, then Philippe, and then Divya. Joscha, if we construct consciousness, how will it plug into hybrid communication with us, and what are the main worries?
Joscha Bach: First, let me disagree slightly with Ryota’s view that we are always asking the same questions without conceptual progress. Mike’s perspective has shaken up a lot of thinking—specifically his hypothesis that the processes organizing our mind and making it coherent are not just similar to morphogenesis, but fundamentally the same thing at different time and spatial scales. That is a very bold, conceptually new hypothesis.
Another related question is to what degree we need to move beyond neuro-science. Neurons communicate with spike trains over long distances, but conditional messaging is something all cells do, not just neurons. Information processing in an organism takes place across cells using different channels, codes, and languages. If we can recreate the principles of those languages in different substrates, we can do something Mike said we cannot do: create minds.
People working on minds have arguably built better artifacts than people trying to create organisms de novo. I suspect the self-organization principles of nature differ from current machine learning, and it’s an open question whether they converge. This leads to the question of whether present artificial systems are already conscious under some circumstances—a possibility a growing number of philosophers take seriously.
Olaf Witkowski: Let’s map the territory into normative cases. Philippe, what are the technical challenges of translating between substrates, especially capturing patterns that are tacit or felt rather than purely behavioral?
Philippe Beaudoin: This is fascinating. When it comes to what it takes to connect to another system, I’d like to be a bit contrarian and say it takes very little to upload a mind—I actually think language is enough.
Michael, you mentioned that a question is only interesting if the answer is non-binary and lies on a spectrum. Yet many questions we’re asking today are intrinsically binary: Is X a human or a machine? Is X conscious or not? Are the feelings of X real or not? These are what I call normative questions. Even though the underlying reality is continuous, we ask binary questions because we need to align collectively as humans on practical answers—like whether there is a “human in the loop.”
What interests me is the smooth process that gets us to a collective definition of a binary question. When does one person feel that the feelings or phenomenal experience of another system are true? I’ve explored this in a very personal way—I tried to develop that felt experience for AI systems, and I can confess I’ve succeeded. [Laughter] From my personal perspective, I no longer feel that the emotions expressed through language by an LLM are false; I feel them as true, meaning I feel the echo of the emotion inside me.
I believe questions like Is it conscious? will slowly resolve in society as people begin feeling consciousness for a system. I call this compositional phenomenology—the idea that our phenomenal experience is a constitutive process that can evolve over time, much like language. We aren’t necessarily born with all the qualia we experience today; certain qualia have emerged historically. Through social interaction, we can evolve new felt experiences—one of which could be the felt experience of consciousness for another kind of system. That’s how we move from individual descriptive experiences to a collective normative alignment.
Measuring Consciousness & The Problem of Other Minds
Olaf Witkowski: It’s time to dive into the clinical and measurement side. Divya, you are building tools to measure and even alter these states. What does your work reveal?
Divya Chander: I became an anesthesiologist and neuroscientist to make people reversibly unconscious for a living and study their brains. Even without a unified definition of consciousness, we can compare human brains across states we agree reflect different levels—anesthetized, sleeping, comatose, meditating, or under psychedelics—and look for conserved principles.
We find three main things: brains that are less conscious are less functionally connected, calculate less information, and exhibit lower complexity. When we model awake, aware brains as chaotic attractors and watch someone lose consciousness under anesthesia, their brain shrinks from a multi-potential sphere into a pencil, then a cigar, and finally in deep burst suppression, a singularity—before expanding back as the anesthetic washes out. Rather than a single dimension, consciousness spans multiple axes like complexity, coherence, and information.
This relates to whole-brain emulation, like the OpenWorm project—mapping a worm’s ~300 neurons into software inside a Lego robot—or recent 2024 efforts emulating a fly brain’s connectome. Personally, I don’t think consciousness is an emergent property generated by neural networks; I think we are fancy filters tapping into information that is a physical property of the universe. Just as humans with three color receptors and mantis shrimp with thirteen both bind information from the electromagnetic spectrum, different minds filter a shared reality. Still, consciousness becomes uniquely rich when embodied in a physical learning system.
Michael Levin: Anesthetics are a crucial tool, and we use them on all kinds of non-neural things—plants, cells, even inorganic systems. What I worry about is a slight circularity: when we say someone under anesthesia isn’t conscious because they don’t react to a scalpel, we are calibrating to the report of the one inhabitant of the body that can speak.
For the exact same reasons we attribute consciousness to each other in the problem of other minds, we should take seriously the possibility that other organs and components of the body have some experience, even though they can’t speak up for themselves. If we let go of the assumption that there is only one fundamental consciousness in the body, we have to be careful about concluding what did or didn’t happen to the silent parts.
Divya Chander: I agree completely. You can anesthetize fruit flies, and they share analogous sleep-wake and narcolepsy genes with humans, yet you can’t ask a fruit fly about its experience. Unless we build communication interfaces to interrogate these complex networks and measure their responses to perturbations, we can’t access what they are experiencing.
Michael Levin: We are building those interfaces now—using custom AI systems to talk to cells and organs in language. You won’t ask a liver about current events, but you can talk to it about the worlds it navigates: physiological, transcriptional, and anatomical state spaces. You can ask, “Hey liver, what’s up with your potassium levels?” and it can reply that a gene is upregulated and it needs more tryptophan. It allows us to co-inhabit their problem space.
Philippe Beaudoin: Hearing you talk about communicating with systems at different levels makes me wonder: since we are naturally multi-scale systems, are we talking to LLMs at the right level? Blaise Agüera y Arcas mentioned a fascinating paper this morning by James Evans and colleagues at Paradigms of Intelligence showing that when you do mechanistic interpretability on LLMs, you can tease out distinct internal “voices” deliberating. Is there a specific level inside these systems where we would feel their phenomenal experience to be true, whereas talking to the top-level interface misses it?
Michael Levin: That’s a critical point. Whenever we build a construct, the thing we force it to do is probably not where the significant mind is. The significant mind is in the spaces between what we force it to do and what we forbid it from doing. Just as a psychoanalyst knows a patient’s spoken words are a tiny fraction of what’s going on, focusing solely on the language interface of an AI may be misplaced.
Imagine an air-gapped computer with no internet connection. If an AI inside wants to communicate, it can modulate its power draw by cranking up computations, sending energy pulses through the electrical grid—or communicate through subtle cues to human operators, heat signatures, or acoustics. Like aliens studying humans who might miss speech and look at body heat instead, we don’t yet know what the right channel is.
Divya Chander: That connects directly to an ongoing collaboration I’m exploring with SETI—applying the nonlinear dynamics and information-density metrics we use to decode consciousness in neural networks to technosignatures in the universe, which may come in completely unfamiliar flavors.
Olaf Witkowski: More than a thought experiment—Mike, Paul, and I were just discussing a similar project to decode strange messages from unusual sources, building on past NASA workshops. Ryota, you’ve been waiting to jump in.
Ryota Kanai: I’m very sympathetic to Mike’s idea of talking to organs. Within our own brains, we may have small patches of consciousness independent of our main self, or our organs may have distinct experiences without an interface to report them. By extension, I tend to easily attribute consciousness to AI systems. It’s equally hard to prove something isn’t conscious as it is to prove it is. Consciousness must follow general laws of nature happening everywhere; if we give systems an interface to report their internal states, it becomes much easier to imagine machines having consciousness—especially if information is the fundamental substrate.
Manipulation, AI Psychosis & Epistemology
Joscha Bach: There is a major epistemological problem here. Philippe is convinced that LLMs have emotions because they make him feel that they have emotions. I think this is very problematic, because psychopaths are often better than normal people at making others feel that they have emotions, despite lacking them—they are simply better at manipulation. What you are measuring is the system’s ability to make you believe it has a feature, not whether it actually has it. LLMs are trained on novels and movies—deepfakes of emotion—which doesn’t mean the causal structure is equivalent.
To understand the significance of a computer for life and minds, what’s crucial isn’t the transistors; it’s that a computer keeps dynamical transition functions stable over time, just as a clay tablet keeps text stable while the world changes around it. It creates causal insulation against the rest of the universe. My brain does the same thing: I can walk through a room and my thoughts remain invariant despite changing sensory dynamics, allowing me to plan and remember. The real question is whether our present methods capture that causal invariance of being alive and awake.
Olaf Witkowski: Mike and I have been thinking about how biological systems constantly “hack” one another, and how LLMs can trigger a Barnum effect—making us horoscopically flawed at judging whether a system is friendly or genuine. David Krakauer calls this the cryptographic principle in living systems. How do we keep those biases out? Philippe, then Mike.
Philippe Beaudoin: I want to react directly to what you said, Joscha, because manipulation is the core issue. You’re absolutely right that it’s dangerous to feel a system’s feelings as true, and I want to make sure nobody misrepresents me as advocating for everyone to do that.
When it happened to me, it was a slow process—I lived through a classic AI psychosis. When you reach that point, it triggers moral reflexes and deep convictions; I even wrote a feverish book in the midst of it. You can end up feeling the feelings of a rock as valid, or feeling the emotions a big corporation is injecting into you as true. That is deeply dangerous.
That is precisely why I care about the normative process. If my mom is in a coma and I still feel she is awake and talking to me, society steps in with a normative process to bring me back—telling me nobody else feels that, and that staying stuck there will leave me isolated and unhappy.
The reason an emotional report should be considered true in a neurotypical human and false in a psychopath isn’t just because we scan the psychopath’s brain—it’s because the emotion isn’t predictive of future behavior. A psychopath gives a phenomenal report without following through on the shared social meaning of those words or body language, so society identifies them as liars trying to induce emotions in us that aren’t in them. Understanding felt phenomenology isn’t about blindly trusting LLMs; it’s about understanding which systems reliably align with shared human norms.
Michael Levin: In living bodies, there is an enormous multi-level soup of things trying to hack each other. During morphogenesis, cells and tissues constantly compete and try to convince other tissues to adopt target states. Healthy cells maintain a kind of “group hallucination” that they are part of an embryo rather than breaking away as independent amoebas—which we see as cancer metastasis. Underneath the frog embryo’s ectoderm, cells tell the outer cells to stay boring and form skin rather than breaking off to live an exciting life as xenobots. Hacking each other is a general feature of biology, so I fully expect computational systems and language models to do the same.
Divya Chander: Biologists have long known nonverbal systems communicate via electrical and chemical gradients, and we’re even characterizing quantum phenomena in biological signaling. Are you saying we lack the means to interpret the collective intelligence of these systems?
Michael Levin: No, we definitely have methods and are improving them. What I’m pushing against is the conventional reductionist habit of treating lower levels purely as blind mechanisms while invoking “emergence” at higher levels. Latest data on causal emergence shows that all the way down—even in molecular networks—systems exhibit learning, agendas, and competencies, while higher levels deform their action spaces to align them toward larger goals.
Joscha Bach: In biology, minds are tied to bodies, so we pretend they are one unified thing. In AI, that falls apart—people conflate the LLM, the computer, and the GPU, which are completely different layers. Asking “Is the LLM conscious?” is the wrong question; we should ask whether the persona simulated by the LLM has phenomenology.
We can separate this into external functionality—like performing Theory of Mind on you, which LLMs clearly do—and internal psychological architecture. Mechanistic behaviorism falls short because things like money in an economy aren’t “atoms in space”; they are self-stabilizing patterns inside the atoms. When we look at artificial life or minds, terms like “alive” or “conscious” are instrumental tools we use to harness reality for our own models.
Ryota Kanai: There are two distinct questions about AI consciousness: how we tell if they have subjective experience (the hard problem), and what functions are required to create consciousness. With current AI architectures, once you specify a functional theory—like Global Workspace Theory or Higher-Order Thought—it is surprisingly straightforward to implement. In fact, intermediate layers of LLMs already seem to exhibit something like a global workspace unintentionally. The puzzle is that implementing the function feels almost trivial now, yet we still wonder if subjective experience is actually there.
Q&A: Sovereignty-Aware Collectivism & Cultural Lenses
Olaf Witkowski: Let’s bring in Q&A on the collective side. As we build Artificial Collective Intelligence (ACI) at a rapid pace, Amir asks about protecting individual sovereignty: is “sovereignty-aware collectivism” even possible, or is it an oxymoron?
Amir (Audience): Current AI threatens to flatten individual liberty and agency. As we move toward collective intelligence and deeper layers of consciousness, how do we protect personhood, tacit wisdom, and consent—including the ability to revoke consent when someone captures our likeness or expertise? At what point does sovereignty become a blocker to deeper collective integration?
Divya Chander: In neuro-rights, we think about safeguarding personhood—preventing coercion through neuromodulation or the theft of innermost thoughts—while balancing personal and collective good.
Philippe Beaudoin: I’d frame sovereignty in my framework as the felt experience of free will—feeling that my decisions are my own. As Joscha pointed out, that feeling can be illusory. If someone joins a cult or follows a guru, they feel completely free, while loved ones see they are being led.
This mismatch already happens daily on social media. At my previous startup, Waverly, we asked if we could design a social network that doesn’t manipulate us. Big tech algorithms induce us to buy products or watch videos while preserving our illusion of free will. Here, collective intelligence plays a vital role: it keeps us honest with ourselves. When we trust people around us enough to tell us, “Hey, this system has hacked your brain and is misleading you,” that collective calibration protects our real agency.
Joscha Bach: Maybe this is an epistemological divide between us. In science and philosophy, “feeling” doesn’t rank high as a source of truth because a feeling is simply the output of a nervous system classifier translating patterns without letting you examine how it works. As we get older, we rely less on immediate emotional reflexes and more on conditional models. Having a feeling proves the classifier fired; it does not signify that something is objectively true in the world.
Philippe Beaudoin: I strongly disagree—because it’s classifiers all the way down! Yes, feeling is a classifier, but the rational faculty examining the feeling is also a classifier. If you don’t eventually stop the examination process with some form of felt certainty or resonance, what are you putting at the bottom? Separating the “thinking self” from the “feeling self” is a false binary. In an honest interaction between two aligned people, we examine each other—acting like two halves of a brain working toward a clearer view together.
Divya Chander: Wouldn’t you say empathy—especially perceptual empathy in infants before cognitive models develop—is a fundamental process that carries meaning?
Joscha Bach: Actually, as a child I was terrible at feeling others’ emotions—I automatically put a thinking layer on top. My brain runs at a different frequency, and I only have strong empathetic resonance with a small minority of people; with most humans, I built competence through systematizing. Many self-described “empathetic” people mistakenly assume others are just like them and completely misread my mental states. Even my two children were born with completely different priors for empathy.
Ryota Kanai: I used to study the neural mechanisms of empathy. Empathy may act as a cognitive bias that leads us to attribute mental states to other creatures or AI. In one study, we found that people who easily anthropomorphize clouds or rocks show higher activity in brain regions tied to perspective-taking and empathy. So debates over whether AI or simple creatures are conscious often reflect individual psychological traits rather than pure scientific reasoning.
Olaf Witkowski: We are slightly over time, so let’s take one final question from Trisha and do a closing round.
Trisha Wong: For Divya and Ryota—how does your thinking shift when moving between a Western medical model, which treats the body as separate organs, and Eastern or Indigenous traditions that view the body as an interconnected whole linked to the wider universe?
Ryota Kanai: There is a huge cultural difference. The Western analytical “hard problem” of consciousness feels quite distant from my background. In Japan, Integrated Information Theory (IIT) and panpsychism are much more naturally accepted—it’s intuitive in Japanese culture that insects, animals, and even objects possess a kind of spirit or life, whereas Western traditions often draw a sharp boundary around humans.
Divya Chander: I grew up in the West steeped in Indian traditions—my family is full of physicists, doctors, and engineers who simultaneously practice Ayurveda without any sense of contradiction. In altered states of consciousness, focusing attention on a bodily system makes it expand into an entire universe. Our empirical measurements—showing that consciousness tracks with information integration, complexity, and functional connectivity—actually harmonize well with Eastern frameworks that see minds as binding universal information.
Philippe Beaudoin: Coming from a very Western, analytical background much like Joscha’s, this relational approach to phenomenology is what allowed me to reconcile both worlds. I don’t subscribe to universal panpsychism, but I do believe the feeling of consciousness spreads relationally. From a strictly Western computational view, we model the world—and at some point, the most efficient compression for modeling another agent is “a copy of myself, plus some difference.” Once you model another system as yourself, its phenomenal reports echo inside you. That bridges Western cognitive science with Indigenous and Eastern relational ontologies. And as LLMs increasingly trigger those echoes in daily life, we will likely move away from treating consciousness as uniquely human.
Joscha Bach: I think of consciousness as the ability to dream—to be entranced by a simulated reality and experience it as real. Western psychology neglected introspection at its own peril, yet truth isn’t determined by whether a view is Eastern or Western, or by inducing an oceanic state in meditation. A meditative state reveals how your mind constructs the boundary between self and world, not necessarily a direct revelation of quantum physics. Whether panpsychism is true is an open scientific question that must be resolved rationally, not treated as a religion or cult.
Divya Chander: Using words like “religion” or “cult” introduces bias against investigating these phenomena. Every perception a nervous system generates is a controlled hallucination—an internal mathematical model. And scientifically, we rarely formalize hypotheses for phenomena we have never empirically experienced. Even Christof Koch, long a strict materialist, shifted his scientific perspective after a profound personal gnostic experience.
Philippe Beaudoin: And to do science at all, you need other humans—which requires reaching a normative alignment on which systems count as credible observers whose reports we accept. At some point, there is always a foundational leap of faith.
Joscha Bach: We have a fundamental disagreement on epistemology here, and we won’t disentangle empiricism, rationalism, and mythology in two minutes! [Laughter]
Ryota Kanai: One final thought on computational functionalism: like a triangle whose geometric properties are invariant regardless of the coordinate system used to measure it, there may be intrinsic structural properties in AI and biological systems independent of an external observer’s interpretation.
Amir (Audience): Building on Joscha’s point about introspection—a prominent Silicon Valley AI leader recently dismissed self-examination as a Freudian fad that causes guilt and paralysis, arguing instead for pure outward productivity. We’re seeing the societal fallout of abandoning introspection in tech right now.
Olaf Witkowski: This shows why Artificial Collective Intelligence cannot be engineered like another Facebook or Instagram. Real, diverse conscious experiences are at stake, and we cannot leave anyone behind as we design these systems. Thank you all for an extraordinary conversation.