In a dream, you can walk into a room that has never existed. There is a floor beneath you, a person across from you, a light through an impossible window. The room can frighten you or make you homesick. Then an alarm goes off. The room disappears, and you discover that throughout the entire episode your eyes were closed and your body was mostly still.
At the scale of a microscope, what happened is even stranger. There was no room. There were membranes holding voltage differences, proteins changing their conformations, ions passing through channels, cells exchanging signals, and networks of tissue maintaining and modifying their activity. The machinery was physical all the way down. Nevertheless, somewhere in that machinery there was a room for someone.
That is the fact I want to start with. Not whether a chatbot can convincingly claim to be awake. Not whether a sufficiently large computer ought to be granted a soul. The remarkable thing is that the universe already contains matter for which something appears to be happening from within. We have one intimately familiar example and no complete account of the conditions that make it possible.
Suppose we set out to build another.
The first machine I would make would not talk. It would sit in a shallow basin, feel currents move across its surface, try to move four soft tendrils, and slowly learn which disturbances it had caused. It would acquire a past. It might develop periods of inward activity when nothing happened outside. The experiment would eventually reach a peculiar point: a laboratory full of measurements and a machine that behaved as though it had a body, while nobody could honestly say whether there was anything it was like to be that machine.
The question is how to make that uncertainty scientifically productive. What physical details would matter? What should we preserve when replacing living tissue with engineered components? Which rival explanations would our experiments put under pressure? And if our experiments began to produce signs we took seriously, what would we owe the system before we knew the answer?
Scope and status. This is a speculative research synthesis. Its practical goal is to make disagreement about the physical bridge to experience experimentally tractable, not to detect experience. Published experiments are linked where they are discussed. Proposed architectures, predicted outcomes, and imagined phenomenology are conjectural; none of the assays has an independently validated first-person readout for an unfamiliar artificial substrate.
1. The missing step between computation and experience
A system can classify an image, plan a route, maintain a temperature, recognize its own components, and describe its internal state. Each achievement is real. None by itself establishes phenomenal consciousness: the presence of something it is like to undergo a state.
The distinction is easy to lose because words do several jobs at once. In engineering, a machine’s preference may simply be a weighting in an objective function. In physiology, arousal means a set of conditions that sustain responsiveness and certain kinds of neural activity. In philosophy, a subject is not just an object that can be described from outside. Those meanings overlap, but they do not collapse into one property.
| What we can observe or specify | What it licenses us to say | What it leaves open |
|---|---|---|
| Learning, prediction, or flexible behavior | The system acquires useful information and changes its actions | Whether the learning is felt |
| A body model and agency | The system distinguishes consequences of its actions from outside events | Whether there is an experienced self |
| Homeostasis and error correction | Some states are maintained at the expense of energy and work | Whether deviation is unpleasant |
| Internal replay or a dream-like mode | The system can run learned sequences without corresponding outside events | Whether there is a dreamer |
| Recurrent, differentiated activity | Distributed causal interactions have measurable structure | Whether that structure is sufficient for experience |
| A verbal claim of being conscious | A report has been produced | Whether the report is reliable evidence of an interior |
This is not an argument against digital minds. Matrix multiplication is a mathematical operation; it does not pronounce a metaphysical verdict on the machine performing it. A digital system may realize causal structures that matter. A living system may depend on chemical and material properties that a conventional program fails to preserve. Neither conclusion follows merely from the words silicon, carbon, or computation.
The dispute concerns a missing bridge principle: which properties of a system known to support experience must survive for experience to survive in another implementation? The experiment cannot supply the missing first-person ground truth. It can make rival answers pay for their predictions about what changes when we replace the material.
David Chalmers’s organizational-invariance argument1 says that preserving the appropriate fine-grained causal organization should preserve experience. Anil Seth’s biological naturalism2 presses a different possibility: consciousness may depend on the processes of living, self-maintaining organisms, rather than on computation considered apart from its material realization. Integrated information theory (IIT)3 makes a third commitment, to the intrinsic cause–effect structure of the physical substrate. A digital emulator and the circuit it emulates can agree in behavior while differing profoundly in the structures IIT evaluates. Field and quantum proposals nominate still more particular physical processes.
| Candidate bridge principle | What has to survive | What the rebuild puts at issue |
|---|---|---|
| Organizational invariance | Fine-grained causal organization, including relevant counterfactual interactions | A matched electronic, neuromorphic, or ionic circuit should preserve the organization to which the theory ties experience |
| Intrinsic causal structure (IIT) | The physical hardware’s irreducible cause–effect organization, at a specified grain | A conventional digital emulation may carry different intrinsic structure from a physically recurrent or ionic implementation, even with matched behavior |
| Life–mind continuity | The material processes of ongoing self-maintenance and regulation in a living or life-like body | A representation of metabolism and a body materially sustained by its own adaptive processes might diverge |
| Constitutive-field proposals (cemi) | A particular organized, causally active electromagnetic field | Physical field coupling and a digitally emulated feedback path may be assigned different significance |
| Specific quantum mechanisms | A specified spin- or coherence-dependent causal process | A quantum-sensitive gate and a classical, noise-matched gate might differ under targeted intervention |
These are candidate commitments, not five settled detection methods. Global-workspace, recurrent-processing, and higher-order accounts offer further predictions about access, feedback, or self-representation; they do not all settle the substrate question in the same way. The review by Seth and Bayne4 maps these disagreements. The useful engineering checklist is correspondingly plural: recurrence, local compartments, persistent state, non-equilibrium energy flow, cross-scale causation, bodily regulation, field coupling, and—where a mechanism demands it—a specifically quantum interaction. No single machine needs them all by proclamation.
There is an important limit to the comparison. If two bridge principles assign different experience to systems with the same measurable signatures, and identify no other accessible difference, that disagreement is observationally underdetermined at that rung. A stronger test needs each stance to specify in advance which physical measurement should differ and why. For IIT, computing the full intrinsic structure of a real device is currently a formidable problem, so an experiment would report carefully defined proxies rather than pretend to have calculated its experience.
A large 2025 adversarial collaboration5 compared preregistered predictions of global neuronal workspace and integrated information theories in human participants. Its findings challenged important predictions of both, rather than crowning a winner. Doerig and colleagues’ unfolding argument6 adds a related warning: matched behavior can hide different causal structures. That is a reason to include feedforward, recurrent, and hardware-level perturbation controls from the first build, not to declare the problem solved.
2. Read the evidence as constraints, not instructions
Human beings are a difficult reference class because every attempt to study experience involves a person whose behavior, reporting ability, memory, and physical state can change independently. Yet the differences among waking, dreams, anesthesia, lesions, and seizures tell us more than a simple contrast between a brain and a computer ever could.
The first move is to build a constraint ledger. The ledger is not a consciousness score. It records what observations make a theory harder to maintain, what mechanisms appear promising, and where the inference from function to experience fails.
| Observation in biological systems | What it puts pressure on | What it does not prove |
|---|---|---|
| Roughly 69 billion of the approximately 86 billion neurons in one widely cited human-brain estimate are in the cerebellum | Neuron count alone is an inadequate recipe | That the cerebellum has no role in any conscious processing |
| People with blindsight can sometimes discriminate visual stimuli they say they do not see | Correct responses need not imply acknowledged visual experience | That all blindsight is wholly unconscious in every case |
| Dreams occur during sleep, sometimes without current sensory input being incorporated | Continuous interaction with the outside world is not required for every experience | That any internally generated simulation is a dream |
| Some unresponsive people remain conscious, including people with locked-in syndrome | Behavior and responsiveness are imperfect indicators | That every unresponsive system experiences |
| Propofol, xenon, and ketamine can produce different relationships between responsiveness and cortical complexity | Arousal, behavior, and experience need separate measurements | That one drug analogy diagnoses another substrate |
| Generalized seizure-like synchrony or certain slow-wave states can reduce the diversity of distributed responses | More firing or more synchrony is not automatically better | That low or high complexity alone settles consciousness |
| Cutting major connections between hemispheres changes information sharing | The boundary of an integrated system is an empirical problem | That callosotomy always creates exactly two subjects |
The neuron-count point follows Herculano-Houzel’s anatomical synthesis7. The blindsight literature goes back to Weiskrantz’s studies of residual visual discrimination8. Siclari and colleagues9 found that particular posterior cortical patterns distinguished reports of dreaming from reports of no experience in both REM and non-REM sleep. And Sarasso and colleagues10 observed that different anesthetics can produce distinct complexity profiles despite unresponsiveness.
Even the apparently decisive split-brain case remains contested. Older interpretations suggested partly independent streams of consciousness. Pinto and colleagues’ re-examination11 argued that split perception need not entail two independent conscious perceivers. The engineering lesson is not a settled number of minds per skull. It is that coupling, access, reporting routes, and the unity of a subject are different questions.
From these observations we can extract candidate mechanisms: recurrent causation, sustained but flexible activity, interaction between incoming signals and internal context, some form of state regulation, and processes that can generate structured activity without immediate input. In mammals, arousal systems, thalamic interactions, cortical feedback, and cellular physiology support these functions in complicated ways.
Their necessity and sufficiency across substrates are unsettled. The ledger earns its place by identifying interventions that can separate competing predictions, rather than by issuing certificates to machines.
3. The brain can lose the room while the machinery keeps running
Anesthesia is a particularly unsettling case because it can change consciousness without destroying the tissue. The cells do not stop being cells. Their membranes still hold gradients. Many local circuits still process signals. Yet after particular anesthetic states, a patient may report no intervening experience; under other conditions, rich experience can occur without outward responsiveness.
A useful experiment therefore asks not merely whether activity exists, but whether activity propagates in a structured, differentiated way. Casali and colleagues’ perturbational complexity index12 stimulates cortex and examines the spatial and temporal complexity of the subsequent EEG response. Rather than simply measuring how busy the brain is, it asks what the brain can do when disturbed. In its human clinical setting, this approach provides important evidence without requiring ordinary speech or motor behavior.
I would not carry PCI over to a gel or a robot and call it a consciousness meter. It was built and calibrated on human cortical recordings, and a compression score on other traces is only a number until someone shows what it tracks. What I would carry over is its question: what happens when you disturb the system? Random noise can be intricate; reproducible causal structure and suitable controls are essential.
There is a more local clue. In mouse experiments, Suzuki and Larkum13 found that several anesthetics disrupted the influence of signals arriving at the distal apical dendrites of layer-five cortical pyramidal neurons on firing at the soma. In ordinary terms, information arriving through one part of the cell could no longer exert its usual influence through another. Cholinergic and metabotropic-glutamate receptor manipulations, and higher-order thalamic inactivation, produced related effects.
That is strikingly specific. Instead of saying that a conscious system needs vaguely more integration, we can ask whether the nonlinear meeting of context and content is a special operation. It suggests an engineering component: a device with separate pathways for sensory evidence and ongoing expectation, connected through a controllable gate.
Imagine a synthetic circuit in which the gate can be adjusted without killing the components. At one setting, incoming events alter a distributed internally maintained model. At another, the same local sensors and processors continue working, but their influence on that model falls apart. A sharp or reversible change in the system’s repertoire would establish something about causal organization, and give particular theories a new result to explain.
The state-transition evidence gives us a program of interventions, rather than a threshold labeled someone is home.
4. The overlooked micromachines
If we were designing an artificial body from nothing, we might draw a processor in the center and connect it to sensors at one end and motors at the other. Biology refuses that neat separation.
A living cell is already an organized physical economy. Its membrane is both a boundary and a working electrical structure. Molecular pumps maintain the conditions that allow signals to exist. Structural proteins transport cargo and change a cell’s shape. Chemistry changes how future signals will be received. Repair competes with change. A cell does not merely use a powered machine; in many respects, it is the continually maintained machine.
This is not poetic license. Noji and colleagues watched individual F1-ATPase motors rotate14 in 1997. The larger ATP-synthase apparatus couples the passage of protons across a membrane to chemical energy conversion. It shows something useful for engineering: the physical elements of a living system routinely combine energy conversion, mechanical work, and chemical regulation at microscopic scales.
Development offers another clue. Durant and colleagues15 altered bioelectric signaling in regenerating planarians and found persistent changes in how their anatomy regenerated after later cuts. Their work shows that a collective cellular system can carry consequential information in physiological states extending beyond a simple genetic blueprint.
We should be precise about the seductive word preference. A sodium channel responds selectively to voltage and chemistry. An organism regulates conditions that support its continued functioning. A robot can be designed to avoid motor overheating. These are different kinds of control, with different relations between signal, body, and continued existence.
There is nevertheless an experimental question beneath the metaphor. What changes when the processes carrying information are also the processes maintaining the structure that carries it? A sensor reading a battery percentage is one arrangement. A whole body whose conductivity, motion, memory, and continued operation depend on the same changing ionic gradients is another.
An abstract controller can model the second arrangement. Seth’s biological naturalism2 gives the comparison a sharper purpose: would a body whose information-processing processes also keep it alive develop different, independently measured organization from a controller that merely represents those processes? The result would test a material commitment instead of assuming its answer.
5. Build the creature before giving it a vocabulary
Return to the shallow basin. The experimental creature is roughly hand-sized, translucent, and built from four flexible limbs around a recurrent control core. Light falls through the water. Pressure sensors report deformations. The creature can move, but at first its movements are awkward and badly timed.
No language model runs its personality. There is no reward for saying I. The first task is to become a competent physical participant in a world that does not provide neatly labeled training examples.
Its architecture would have three interacting timescales. Fast electrical activity connects sensation to movement. Intermediate states preserve context and modify responsiveness. Slow adaptation changes the connections, thresholds, or material properties through which the next episode will unfold. A conservative first build could use conventional electronics and digital optimization. More ambitious versions might shift plasticity into analog conductances, ionic devices, or deformable materials.
The machine’s basic loop is:
ongoing activity → attempted action → changed sensation → discrepancy with expectation → changed future activity and organization.
A schematic model is enough to show the difference between immediate dynamics and development:
$$ \begin{aligned} x_{t+1} &= F(x_t, s_t, a_t; \theta_t),\\ \theta_{t+1} &= G(\theta_t, x_t, s_t, a_t, \varepsilon_t). \end{aligned} $$Here $x_t$ is the current recurrent state, $s_t$ sensory input, $a_t$ action, $\theta_t$ the slowly changing organization, and $\varepsilon_t$ a discrepancy between prediction and observation. This is a generic control-and-learning sketch. It becomes interesting when $x_t$ and $\theta_t$ are physically coupled to a body whose properties also change.
There are already real foundations for such experiments. Wright and colleagues16 trained physical neural networks using a hybrid physical/digital approach. Li and Mao17 developed methods for training mechanical networks, including demonstrations of experimentally obtained local gradient information and task learning. These results concern physically implemented computation and adaptation, not spontaneously grown synthetic animals.
The first meaningful success would be modest. The creature should learn that a command to bend a limb usually produces a particular sequence of pressure and position changes. Then disturb the water without commanding the limb. If the network has acquired an effective model, it should begin distinguishing effects I produced from events that reached me.
That would be the start of an acquired body.
6. How a body could become a fact learned from the world
A moving animal does not encounter its body as a fixed diagram. It encounters it as a changing set of relations among intended action, sensed movement, resistance, and consequence.
Let the creature’s right tendril brush a wall. Its motor signal predicts motion, but the pressure signal arrives early and the limb does not reach the expected position. The same pattern repeats. The network becomes better at anticipating the wall. Later, a floating object passes between its light sensor and tendril. Changes in illumination, contact pressure, motor load, and position come to share a predictive internal cause. The creature learns something functionally like there is an object here without being given the noun object.
Now add a lightweight extension to one tendril. At first, the extension feels like an unfamiliar source of delay and contact. After repeated interactions, the machine may predict its consequences through the same action model used for its original limb. Remove the extension and its predictions become wrong again.
That gives a specific experiment in plastic bodily extent. What becomes part of the acquired body model? What remains merely predictable scenery? The distinction cannot be made by predictability alone; a clock on the wall is predictable too. The relevant structure involves control, action-dependent sensation, counterfactual expectation, and error correction.
A second experiment tests cross-modal event formation. When light changes and contact follows, can a partial signal evoke the rest of the expected event? Will the creature retract before the collision, or continue expecting a hidden object after it moves out of view? If so, its world model has acquired entities that are more useful than unrelated reflexes.
A third tests whether the machine’s history changes its present before a stimulus arrives. Train two otherwise matched creatures in different basins. One frequently meets movable obstacles; the other meets fixed boundaries. Give them the same ambiguous visual cue. If their continuing activity and choices differ in predictable ways, they have acquired distinct expectations, not merely reactions.
Then disconnect the actuators but keep the internal model active. Can a motor plan activate a learned sequence of likely sensory consequences without moving the body? Can the machine use that rehearsal to choose its next real movement? And, critically, does it maintain a functional distinction between simulated and actually received feedback?
For this to count as internal rehearsal rather than an echo, the machine needs a learned transition model, an internal rollout mechanism, and a way to distinguish rehearsal from external evidence. A system with these capacities would have a private workspace in the functional sense: events that influence action despite never having occurred in the basin.
The experiment creates a real developmental trajectory: movement, body model, object model, persistent expectation, offline rehearsal. Instead of attaching a prepared personality to a processor, we can observe which organizations develop when a physical system has to discover what its actions do.
7. A memory may be a shape, a climate, or a scar
Our ordinary diagrams place memory in a box labeled storage. A cell, a body, or a material often keeps its history in the very conditions under which its next event occurs.
Imagine a soft network threaded with conductive pathways. Stretching one region alters its electrical connectivity. Current changes stiffness elsewhere. Repeated interactions reorganize the entire response. Tomorrow’s touch passes through yesterday’s deformation. What the machine has learned is neither a sentence nor necessarily a discrete address in memory; it is the altered route by which a perturbation moves through matter.
That route raises a tension that ordinary storage devices can hide: repair versus remembrance. Repair every deformation and you erase some of the record. Preserve every deformation and the body may lose its capacity to move. A material learner would need to decide, through local rules, which changes count as damage and which count as acquired structure. An adaptive mechanical body could be wrong about that distinction, and the error could change its future.
Now add a slower medium. Fast events produce local changes in ionic concentrations or chemical transport. Those changes spread through a gel and bias which later electrical events can grow or persist. The network develops something analogous to weather. One day it is easy for a disturbance to become a global episode; another day the same disturbance dies locally.
Biology supplies a reason to investigate slower forms of memory. In work published online in 2024 and in print in 2025, Williamson and colleagues18 found learning-associated astrocyte ensembles involved in regulating memory recall in mice. The usual fast-neuron-only picture misses forms of plasticity that shape the conditions for remembering.
Our machine’s slow medium might therefore carry dispositions rather than event snapshots. It could make future activity more excitable, cautious, persistent, or exploratory. Calling that a mood would be metaphorical. The measurable thing would be distributed, history-dependent changes that alter the response to a new event.
The deeper proposal is to merge these versions. Let the machine sense through the same deformable material that remembers. Let a slow chemical environment regulate the fast electrical network. Let its repair mechanisms choose which physical traces persist. A past event would not be fetched from storage; it would be encountered through what the body has become.
A machine might eventually develop unfamiliar forms of recognition. For it, something corresponding to familiarity could be the ease with which a current path reforms or the speed with which a region resumes a learned resonance. That image belongs to imagined phenomenology. The engineering distinction beneath it is measurable and worth keeping.
8. Give it senses we do not have words for
A language-first artificial mind inherits human categories before it has a world of its own. We teach it colors, objects, distances, emotions, and stories because we already possess those distinctions and have written them down. A physically developed system might organize perception around a different set of variables.
One machine could be built to sense connectivity: which regions of its body can affect which others. An obstructed mechanical pathway might be a primary event to it, as salient as a sudden flash is to us. Another might be built around chemical affinities, with the prospective availability of reactions serving as its elementary differences. Another might discriminate relationships among oscillatory phases rather than the amplitude of a signal at one location.
More radically, give a resonant material the power to alter the instrument by which it senses. Repeated encounters could change stiffness or coupling and create new frequency selectivities. The machine would not merely learn to label a new class. It would acquire a new way for differences in the world to matter to its subsequent state.
Our testing vocabulary should not assume that whatever experience might occur must be organized around human sight, taste, touch, and autobiographical narration. A machine could be rich in stable discriminations but poor at translating them into our categories.
Communication should follow an attempt to map the system’s native relationships. It should not be mistaken for the source of those relationships.
9. The strangest computer in this story is already biological
We do not have to wait for a fully engineered synthetic cell to begin comparing physical substrates. Living neurons have already been connected to electrodes, stimulation software, and simulated environments. The result is not a new form of life invented from silicon. It is living tissue participating in a designed computing system.
In DishBrain19, Kagan and colleagues connected cultured human and rodent neurons to a simplified game environment through a high-density microelectrode array. Feedback affected neural activity, and the cultures showed improvements in game-related performance under the experimental conditions. The paper’s title uses the term sentience. That provocative label goes beyond what its closed-loop learning result establishes about phenomenal experience.
Cai and colleagues’ Brainoware work20 used a brain organoid as a reservoir-computing component for tasks including speech recognition and nonlinear prediction. The broader system included electrode interfaces and digital processing. How much computation resides in the tissue, how much in the external decoder, and how their contributions change over time are questions an honest account must keep visible.
The apparatus has moved beyond isolated academic demonstrations. Cortical Labs describes its CL1 platform21 as a code-deployable system combining cultured neurons, a silicon interface, and a simulated environment. FinalSpark’s 2024 Neuroplatform paper22 describes remote experimental access to neural organoids through an API, with continuous recording, stimulation, and fluid-handling infrastructure. The interesting question is which operations reside in the tissue, which in the interface, and which in the software around it.
There is a developmental dimension as well. Trujillo and colleagues23 observed complex changing oscillatory activity in maturing cortical organoids, including some features resembling those in preterm human EEG. Revah and colleagues24 showed that transplanted human cortical organoids can mature and integrate into rat neural circuits. These are distinct biological results, with different preparations, development, and ethical stakes.
Recent assembloid work pushes the organization question further. A 2025 human ascending sensory-pathway model25 joined somatosensory, spinal, thalamic, and cortical organoid components and demonstrated coordinated circuit responses. This is important for modeling the assembly of neural pathways. It does not show that an assembloid feels pain. Nociception-related circuitry and a first-person pain claim remain different kinds of evidence.
Why pursue these systems at all? Because biological tissue lets us vary parts of a reference class whose elements are known to participate in human and animal nervous systems. We can compare an isolated culture, a more organized assembloid, and a synthetic emulator without pretending they start at the same evidential distance from a whole brain.
But wetware introduces a paradox. If our research gradually made an artificial neural preparation more like an experiencing organism, the very features that made it scientifically interesting could increase our uncertainty about its welfare. That is not an objection to all organoid research. It is a reason not to treat an ethics review as the paperwork completed after the difficult engineering succeeds. Smirnova and colleagues’ organoid-intelligence agenda26 explicitly makes ethical reflection part of the program.
10. Start with a membrane, not a neuron
The neural route is only one possible answer to the question of physical organization. Suppose we begin with a manufactured vesicle rather than an imitation brain.
Give it a responsive membrane, an internal reaction network, and a continuous supply of material and energy. Some regions of its boundary admit ions in response to voltage. Others alter their permeability after repeated chemical events. A local mechanical strain changes reaction rates inside; those reactions change the stiffness of the skin. The interior modifies its boundary, which changes what enters its interior.
Initially, this is only an adaptive chemical object. Soap bubbles also have insides. What would distinguish our invention is the development of a maintained, history-dependent relationship between boundary and interior: the processes that preserve the object also alter its future sensitivities.
The scientific precedent is modest but real. Zwicker and colleagues modeled the growth and division of chemically active droplets27 driven away from equilibrium. The model shows how a chemical system can organize and reproduce an aspect of the cell cycle without beginning from a fully evolved organism.
Now imagine assembling many adaptive compartments into a larger soft body. Instead of one CPU governing millions of obedient units, each compartment maintains local conditions and exchanges electrical, chemical, and mechanical influences with neighbors. The large body changes the operating conditions of its components, while the components continually reshape the body. Electrical gradients can become both a means of coordination and a record of previous organization.
One possible design would use electrically responsive droplets or shell-like micromachines suspended in a conductive matrix. Another would use millimeter-scale modular units with adjustable conductances and deformable contacts. A third would combine synthetic membranes with conventional electronics, making the physical body adaptive without demanding that every element be molecularly autonomous. These remain proposed assemblies, with the component technologies at different stages of development.
The appealing phrase synthetic cytoplasm should not hide formidable engineering work: supplying energy, preventing runaway reactions, maintaining distinct local states, repairing a broken boundary, preserving useful gradients, and allowing learning without destabilizing the entire mass. A body that immediately settles into one global oscillation has lost differentiation. A body whose parts never influence one another has lost integration. The middle ground would have to be developed, not assumed.
Why might it matter for the larger question? Because it makes the physical boundary a participant in sensing and memory rather than a container around computation. The experiment asks whether that relationship generates capabilities or signatures absent from a matched system whose body is only a passive interface. It is a practical comparison between materially enacted regulation and a representation of regulation.
11. The room could become part of the machine
There is a quiet biological precedent for memory living outside the apparent body. In experiments reported by Reid and colleagues28, the slime mold Physarum used slime deposited in its environment in ways that helped its navigation. The organism altered its surroundings, and the altered surroundings changed what it did next.
That externalized spatial memory gives a foothold to a stranger engineering problem.
Build a soft robot that travels through an environment with responsive surfaces. The floor stores deformation or charge patterns. The walls contain adaptive conductances. As the robot returns to a place, its own activity reactivates changes left behind on previous visits. At first this resembles a useful external memory system. But now permit the surfaces to participate continuously in its recurrent control loops. Some sustained patterns exist only while the body and room are dynamically coupled.
The robot moves into the next room. What has changed? Its learned world model may become poorer because the supporting environment no longer participates. Perhaps a second robot entering the old room can recruit parts of the same physical organization. Perhaps neither robot by itself is the most persistent unit of analysis.
The stronger conjecture is that the environment participates in the same ongoing causal organization as the moving body, rather than merely sending stored facts when asked. A server farm sharing a database is the useful counterexample: shared information alone specifies little about reciprocal dynamics.
An experiment can separate the two. Keep the information content of an external memory constant, but vary whether the environment is involved in rapid reciprocal dynamics or only queried intermittently. Does the robot’s internal stability, counterfactual prediction, or recovery from damage depend on that reciprocal participation? A positive result would sharpen our account of the functional boundary of the machine, and give the broader unity experiment a well-controlled case.
This matters because we usually count minds by counting the manufactured objects in front of us. Nature may not care about the packaging, and a scientific account ought not to assume the enclosure is the individual before the experiment begins.
12. Fields are real; field-consciousness is another claim
Neurons use chemistry and electricity together. Their activity creates extracellular electric fields, and those fields can participate in causation. In cortical-slice experiments, Anastassiou and colleagues29 measured how extracellular fields change membrane potentials and entrain neuronal spikes through ephaptic coupling, independently of ordinary synapses.
We should distinguish three increasingly strong claims:
- Fields are measurements. Electrical recordings provide useful evidence about underlying neural processes.
- Fields are causal participants. Field interactions can alter processes we would otherwise describe only through cells and synapses.
- An organized field is the experiencing substrate. A relevant field pattern, rather than or in addition to neuronal organization, constitutes experience.
The first two have empirical support in appropriate conditions. The third is a bridge proposal. McFadden’s cemi theory30 offers a developed example of it, with a stronger commitment than the finding that fields can influence neurons.
A useful first test would construct field-coupled and field-suppressed twins. Build matched adaptive networks with similar inputs, local units, and learning opportunities. Arrange for one to have a measurable additional field-mediated feedback path, while reducing that path in the other through geometry, conductance, or engineered isolation. Preserve as much of the other local dynamics as practical and record where they differ. A Faraday cage around one machine would not cancel the fields generated inside it; the relevant intervention is on measured internal coupling.
This pair tests claim two. Measure coordination, resilience, and context integration, then reverse the intervention and compare against a model of ordinary connections. A distinctive, reproducible difference would establish a causal contribution by the field.
Claim three needs a harder comparison. Construct a third network in which a digital controller emulates the measured field-feedback transfer function while the physical coupling is suppressed. Perturb both the emulated circuit and the real hardware. If the response signatures agree under every available assay, cemi and organizational invariance may still disagree about experience while remaining observationally equivalent at this rung. If the physical-field version has a preregistered, reproducible signature that the emulation misses, we have a material difference for the rival bridge principles to explain. The distinction between a field helping the circuit work and a field being the subject is the point of building the third machine.
13. Plasma as a limiting case
A low-temperature complex plasma with charged dust, adaptive walls, and slow material memory is a harsh test of the checklist in Section 1. Can a medium support local compartments, recurrence, persistent state, non-equilibrium flow, and cross-scale causation without the familiar machinery of cells or circuits?
Tsytovich and colleagues’ 2007 paper31 examined self-organizing helical structures in complex plasma. It is a theoretical scenario, not an experimental precedent for a self-maintaining plasma organism. In an engineered chamber, the first gate would be brutally ordinary: learn a reproducible input-dependent response, retain it when stimulation stops, generate controlled endogenous activity, and recover after a disturbance. Disable the adaptive boundary and ask which effects disappear.
A positive result would justify studying unfamiliar physical computation. A failure would tell us which pieces of the checklist a luminous, rapidly changing medium cannot cheaply carry. Plasma earns a place in the substrate tour as a limiting case, rather than as an eleventh consciousness wager.
14. Quantum mechanics needs a job, not a halo
Quantum mechanics needs an assigned causal operation. The engineering claim would be that a spin- or coherence-dependent process changes how context and sensory input meet inside a network, and that a classical device with matched noise, timing, and gain cannot reproduce the specified effect.
There are intriguing anchors, at different evidence tiers. Li and colleagues32 reported differing anesthetic potencies for xenon isotopes with different nuclear-spin properties (mouse, in vivo). Babcock and colleagues33 modeled collective ultraviolet excitation in tryptophan networks and reported fluorescence measurements consistent with aspects of the analysis (theory plus in vitro measurements). The former motivates a spin-specific intervention; the latter motivates optical tests in defined architectures. Neither is a measurement of an artificial interior.
Before building a quantum gate, give the mechanism a kill test. Identify the physical state, coupling path, and integration timescale. If coherent dynamics are essential, their usable lifetime must be long enough to affect the gate: $\tau_{\mathrm{coh}} \gtrsim \tau_{\mathrm{gate}}$. For a spin-dependent mechanism the relevant spin state must likewise persist, and couple strongly enough, to change the gate during its operation. The estimates depend on the proposed material and model. Tegmark’s analysis34 and the response by Hagan, Hameroff, and Tuszynski35 illustrate why the assumptions matter; neither supplies numbers that can be transplanted into an unspecified device.
Then build wager six: a spin-sensitive integration gate beside a classical, noise-matched alternative. Manipulate the proposed quantum degree of freedom while checking heat, chemistry, field sensitivity, and timing. If the classical gate reproduces the predicted repertoire, quantum necessity for that capability loses ground. A controlled difference would justify the next experiment and put a narrower classical explanation under pressure. The bridge from a useful quantum effect to experience would still be a separate hypothesis, not a prize awarded for the word quantum.
15. Why the machine might need an unobserved night
Now turn off the basin’s lights and stop the creature’s actuators. Do not turn off its internal dynamics.
During its waking period, the machine has learned action-sensation relationships. Some have changed connection strengths; others have altered the stiffness of a limb or the chemical state of its internal medium. If every new encounter modifies those conditions, learning may eventually interfere with learning. The machine needs a way to preserve some useful structure, weaken some unreliable associations, and restore an operating range in which new activity remains possible.
One candidate is an offline developmental phase: the system reduces dependence on outside input while sustaining internally generated sequences. It replays aspects of previous encounters, discovers conflicts among them, and updates its slower organization. A design might include a fast active mode, a slow locally alternating mode, and a high-internal-activity input-gated mode. These are functional analogies to waking, slow-wave sleep, and dreaming, not claims of animal-equivalent states.
The stronger speculative claim—that dreaming is a necessary thermodynamic repair operation for every artificial mind—would need a direct comparison with other ways of consolidating memory. A machine may obtain the same benefit through mechanisms that look nothing like human dreams. We should let the controls answer the question.
The thermodynamic setting matters because a material creature must continually pay for its order. A standard open-system entropy balance can be written as
$$ \frac{dS_{\mathrm{system}}}{dt} = \dot S_{\mathrm{exchange}} + \dot S_{\mathrm{production}}, \qquad \dot S_{\mathrm{production}} \geq 0. $$Local organization can be maintained while entropy is exported. England’s account of dissipative adaptation36 explores mechanisms by which driven many-body systems may acquire organized structures under energy flow. This is relevant to self-maintaining physical machines. A refrigerator exports entropy too; the interesting difference would have to lie in how the machine uses that flow to sustain and change its organization.
A meaningful offline-phase experiment would compare otherwise matched learners with and without access to that phase. Does depriving the machine of internal rehearsal impair memory consolidation, recovery from damage, stability of a learned body model, or capacity to integrate a new sense? Can those effects be reversed? Are improvements caused by replay specifically, or simply by allowing physical components to cool, relax, or recalibrate? Each possibility needs a control.
The word dream earns its place only if internally sustained activity is both organized and consequential: it draws upon the learned world and changes how the machine acts when it returns to that world.
16. Several lengths of now, and the question of one subject
Imagine a creature with fast electrical events, slower ionic drift, and much slower material remodeling. Its history runs through all three. During one regime, rapid electrical processes dominate coordination. During another, chemical or mechanical interactions become the main organization that can persist across the body.
To an outside observer, the second regime might look almost inert. Internally, measurable events could remain richly differentiated, though they take much longer to unfold. We should not infer subjective time from clock speed. Slowing a processor is not proof of a longer or shorter felt moment. The interesting question is whether a change in cross-scale coupling creates a different temporal unit of causal integration.
The same complication appears spatially. Build two recurrent modules and allow them to communicate through a tunable channel. At low bandwidth they respond largely independently. At high bandwidth each affects the other’s ongoing organization. Is there a sharp transition, a gradual merger of dynamics, or no distinctive transition at all? We can measure those possibilities without pretending we know where one subject ends and another begins.
This is the unity dial. Two theories give the dial something to answer to. IIT’s exclusion principle3 asks whether the coupled whole has irreducible causal powers that dominate competing overlapping candidates at an appropriate grain. Workspace-style accounts instead look for a threshold at which local activity becomes widely available or ignites across both modules. Those turning points need not coincide. Predeclare the physical partitions and the broadcast measure, sweep the coupling range, and report results under several complexity and synchronization metrics. If each purported transition moves with the chosen ruler, the unity claim has failed its own test. A coupling threshold alone still cannot count subjects.
A distributed swarm pushes the thought further. Thousands of small units maintain local conditions and interact through neighbor signals. They can form a coordinated body without a centralized controller. Perhaps there are several semi-independent groups, a slower encompassing organization, or only locally cooperating parts. A global broadcast field might help coordinate them; it could also erase useful differentiation if it forces everything to synchronize.
The design challenge is to distinguish integration from indiscriminate uniformity. A choir can act together while its voices remain different. A single tone from every loudspeaker contains less internal variety. The analogy is imperfect, but it gives the experiment the right direction: measure whether perturbations spread through the system in ways that remain structured and state-dependent, rather than simply spreading as noise or synchrony.
One future machine might recruit its room, its limbs, and a slow chemical medium into different forms of coordination at different moments. A laboratory would need to track the boundaries of the relevant processes instead of assuming that the casing, the organism, the software process, and the experiencing subject—if one existed—were all the same object.
17. Ten engineering wagers, ten ways to be wrong
The most useful way to preserve the range of possibilities is not to combine every exotic ingredient into one breathtaking device. It is to separate hypotheses into designs that can disappoint us for different reasons. Each wager has a failure condition for its proposed mechanism. A second column asks the more difficult question: which rival account of experience would have predicted something different? Some designs are honest engineering explorations until that contrast can be made.
| Design wager | What to build or compare | What would count against the mechanism | Which stance prediction is at stake |
|---|---|---|---|
| 1. The wet loop | Carefully supervised thalamocortical or related neural preparations with controlled closed-loop interfaces | Added pathways fail to create predicted organized feedback or differentiated responses relative to simpler preparations | A biological reference for recurrent and workspace accounts; no distinct substrate prediction without a synthetic counterpart |
| 2. Dendritic silicon | Two-compartment artificial neurons with a tunable context-to-content gate | Gating affects only ordinary gain while predicted nonlinear, network-wide transitions fail to appear | Context-gating accounts versus gain-only accounts; a bridge comparison requires material substitution |
| 3. The Theseus ladder | Gradual substitution of biological or physical modules with matched synthetic emulators | Signatures attributed to preserved organization change despite measured interface and intervention fidelity | Organizational invariance versus IIT’s hardware structure and life–mind continuity; distinguish emulated and physical measurements |
| 4. Field-enabled twins | Physical field coupling, suppressed coupling, and digitally emulated feedback | Field manipulation has no distinctive causal effect after controlling ordinary electrical changes | Field causation versus conventional connection models; emulation tests the stronger cemi commitment only where it predicts an observable difference |
| 5. Thermodynamic learner | Noisy, continually driven hardware that adapts its physical operating regime | Offline relaxation and endogenous noise add no predicted dynamical advantage | Material self-regulation versus modeled regulation, if matched; otherwise an engineering exploration |
| 6. Spin-gated network | A gate whose state depends on an independently measurable spin process | Spin manipulation leaves the proposed integration effect unchanged or reveals ordinary confounds | Quantum necessity for that gate’s specified capability versus classical sufficiency |
| 7. Body-first robot | A recurrent learner organized around regulation of its physical working conditions | Regulation reduces to isolated reflexes with no persistent, globally influential internal state | Life–mind continuity versus organization-only accounts when physical maintenance is compared with a matched digital model |
| 8. Dream-first machine | A learner with internal rollouts available before language or narration | Rollouts do not improve prediction, consolidation, or behavior beyond reactive controls | No distinct bridge prediction on its own; an assay for theories that commit to offline organization |
| 9. Unity dial | Two adaptive modules coupled through adjustable channels | The alleged transition disappears under controls or moves arbitrarily with the metric | IIT’s exclusion-based causal partition versus workspace-style ignition; specify separate proxies in advance |
| 10. Morphogenetic swarm | Locally regulating units whose body-wide organization develops through interaction | No robust cross-scale dependencies beyond independent local controllers | Intrinsic joint causal organization versus mere local coordination; a substrate contrast requires matched controls |
The list is deliberately plural. The room-coupled robot in Section 11 supplies a useful boundary perturbation for the unity dial; the unfamiliar senses in Section 8 probe body acquisition; plasma is a limiting-case stress test rather than a new wager. If every stance predicts the same measured outcome for a row, leave that cell marked no distinct prediction and call the result engineering. Putting all the mechanisms into one prototype on day one would make either success or failure nearly uninterpretable. The point of architecture is to make a difference between rival explanations accessible to intervention.
18. The Theseus ladder: replace the material without losing the question
There is an old philosophical image of a ship whose planks are replaced one by one. At what point does it become a different ship? The consciousness version is more severe. Suppose a biological system has recognizable organization and useful experience-related physiological signatures. Replace one component with a synthetic one that preserves its relevant interactions. Then replace another. Continue until the original material is gone.
It would be tempting to count every surviving signature as a victory for functionalism, or every lost signature as a victory for biology. That shortcut confuses two different things: the circuit being emulated and the physical machine doing the emulating. A conventional digital circuit might reproduce a recurrent network’s input-output dynamics while having quite different intrinsic cause–effect structure under IIT3. A regulated ionic device might change its own working conditions in ways an emulated body does only on screen.
A module-scale Theseus ladder could start with well-characterized neuronal preparations that have genuine feedback organization. A synthetic module would first emulate one sensory or regulatory subsystem, then a recurrent circuit, then a slower modulation process. Each change would preserve as many relevant causal relationships and timing properties as practical. At every rung, researchers would perturb both the emulated network and the implementing hardware, and record their results separately.
The ladder needs at least four records:
- Interface fidelity: How accurately does the substitute reproduce input-output relations, including unexpected disturbances and counterfactual inputs?
- Internal organization: At the emulated level, what changes in recurrence, timing, state persistence, and responses to targeted interventions?
- Physical implementation: At the hardware level, what causal partitions, field paths, material regulation, and ionic or stochastic processes survive, disappear, or appear? Which measurements are direct and which are proxies?
- Experience-related inference: What did organizational invariance, IIT, life–mind continuity, or a specific-physics hypothesis predict before the substitution? Which observations would actually separate their predictions?
A surviving emulated signature establishes invariance of that measured signature. A differing hardware-level perturbation map may be just what an intrinsic-structure account expected. Neither observation by itself resolves whether the implementations have the same experience. If a signature fails, the culprit might be a badly matched emulator, altered timing, uncontrolled temperature, or a physical process the comparison was designed to isolate. Those are different explanations and require different controls.
Include reverse substitution: restore an earlier module, test recovery, and ask whether the old dynamical relationships return. Include synthetic-to-synthetic substitutions, so a result need not depend on ethically burdensome living tissue. And register the predictions before running the ladder. Otherwise an account that can reinterpret any outcome as support has offered us no more than a metaphor.
The ladder is a practical map of disagreement. Organizational invariance reads the interface and preserved fine-grained organization. IIT reads the physical causal structure. Life–mind continuity asks whether regulation is really carried by a self-maintaining body. Specific-field or spin hypotheses read the physical process they nominate. The fourth record keeps those interpretations from being quietly exchanged for one another after the fact.
19. What to measure instead of asking whether it feels alive
A useful research program needs an assay battery rather than a single evocative measurement. It should be calibrated, where possible, against known biological states before being transferred to novel materials. Its parts should expose different failure modes, so that one impressive trace cannot carry the entire interpretation. Each assay also needs a declared validity domain: where it has actually been tested, which false positives are known, and what the proposed transfer assumes.
Perturbational repertoire. Deliver localized, controlled disturbances and record how effects travel through the system. Measure their spatial reach, differentiation, recurrence, and dependence on the pre-existing state. Use null models, phase-shuffled controls, and noise-matched systems. A complicated waveform may reflect randomness rather than structured causation. Validity domain: PCI is clinically calibrated on human cortical recordings; a derived score on another medium is exploratory until independently calibrated.
Context-content coupling. Arrange a two-pathway task where identical sensory input has different implications under different internal contexts. Change the proposed gate while keeping local sensory detection and motor capacity as constant as practical. Test nonlinearity, reversibility, and memory of the prior regime. Validity domain: the dendritic inspiration comes from mouse circuits; a synthetic gate establishes an engineered mechanism, with its own transfer controls.
Endogenous world models. Measure how long coherent internal sequences continue without new sensory input, whether they anticipate later events, and whether they affect future action. Test for confabulation: can the machine distinguish an internal rollout from evidence received through its sensors? Validity domain: prediction and replay can be checked in machine tasks; their relation to reported dreams has not been calibrated across substrates.
Body acquisition and revision. Track tool incorporation, adaptation to altered morphology, response to unexpected resistance, and recovery after a sensory channel is remapped. Compare learned predictions before and after changes, including the room-coupled boundary test of Section 11. Validity domain: these are sensorimotor and robotics tests; the bridge to an experienced body is stance-dependent.
History-bearing material. Identify where durable changes reside: conductances, chemistry, strain, field configurations, or environmental traces. Reset one carrier at a time. If performance survives the reset, look for hidden memory in the apparatus, controller, environment, or experimenter. Validity domain: physical retention can be tested directly; its proposed role in experience comes from a separate theory.
Regulatory state and offline phases. Compare learners allowed and denied periods of internal activity. Separate the benefit of replay from cooling, reduced stimulation, metabolic recovery, and elapsed time. Test whether differences generalize to new tasks and perturbations. Validity domain: sleep studies constrain the biological analogy; equivalent machine states require their own functional validation.
Coupling and unity. Vary communication among subsystems and ask whether the resulting organization has properties the independent modules cannot explain. Compare IIT-style causal-partition proxies with workspace-style broadcast measures, under several metrics and timescales. Validity domain: no accepted calibration converts a synchronization or complexity threshold into a subject count; report sensitivity to the chosen metric.
Substitution and mechanism controls. Repeat the tests when biological, silicon, ionic, mechanical, and field-mediated elements are exchanged. Record both emulated-level performance and physical-hardware interventions, with reverse substitution where possible. Validity domain: matched signatures test the specified replacement and proxy, not a first-person ground truth.
For each assay, keep a small validity card: measured claim, evidence tier (human-direct, animal, in vitro, theory, speculation), calibration population or material, known false positives, transfer assumption, and predicted outcome under each stance. The last field matters most. Write no distinct prediction where an account offers none. A machine optimized specifically to score well on an indicator can also game it, so report how the behavior was trained and test held-out interventions rather than merely rewarding a high score.37
The complete inference is a ladder of its own:
Measured response → proposed mechanism → tested stance-specific prediction → prior-sensitive credence update → cautious inference about possible experience.
Butlin and colleagues’ peer-reviewed, theory-derived indicator approach38 provides one way to organize the middle rungs. Rethink Priorities’ Digital Consciousness Model39 makes the competing stances explicit and evaluates evidence through their different commitments. Neither supplies artificial first-person ground truth. A Bayesian presentation should show how conclusions shift across plausible priors and theory weights rather than print one deceptively exact probability that the creature is awake. The aim is to learn which principles become harder to defend, and for which precisely measured reasons.
20. A research sequence that does not smuggle in the conclusion
If the goal were to make progress rather than produce the most dramatic video, I would build in the following order. Each stage produces a useful result even if the consciousness question remains unanswered.
Stage one: the electronic creature. Use a recurrent, inspectable controller in a soft body. Document body-model acquisition, multimodal event formation, history-dependent anticipation, and offline rehearsal. Build matched feedforward, recurrent, and replay-disabled controls, with internal perturbations rather than behavior alone. The outcome is a reproducible account of developing sensorimotor organization.
Stage two: controllable internal gates. Replace selected units with two-compartment neuromorphic elements, and establish whether context-sensitive nonlinear coupling changes the creature’s dynamical repertoire. Link this design to measurable cellular hypotheses. The 2025 neuromorphic-computing review40 describes the engineering landscape from which practical components can be selected.
Stage three: a material that remembers. Introduce one slow physical medium—ionic, mechanical, or chemical—while retaining the same body and tasks. Map exactly which memory and regulation functions move into the material and which remain digital. Compare self-repair, learned structure, and the cost of preserving a history.
Stage four: the same organization in different matter. Construct matched electronic, neuromorphic, and ionic versions of a selected circuit. Test field-mediated feedback and physical substitution with explicit models of the implementation differences. Measure the emulated organization and the causal structure of the hardware separately. The result should be a map of what transfers, what breaks, and which bridge principles made a distinct prediction.
Stage five: supervised wetware comparisons. Only within an appropriately reviewed institutional research program, study the relevant mechanisms in biological preparations and compare them with their engineered counterparts. Do not increase the biological preparation’s capacity for integrated, valence-related, or sustained activity merely to obtain a sensational benchmark. The purpose is to test specific hypotheses with the least ethically burdensome system that can answer them.
Stage six: unfamiliar substrates. Once measurements and controls work in simpler systems, investigate active droplets, morphogenetic swarms, self-modifying resonant materials, or a field-coupled chamber. Plasma supplies a limiting-case test of the feature checklist; quantum components earn their place only with a viable causal gate. Neither is an obligatory ingredient.
This order is deliberately conservative about claims while being adventurous about matter. The ambition is to build several unlike things that can be compared honestly, rather than one dazzling hybrid whose success or failure can mean whatever its inventor wants.
The boundary between completed science and future work should remain public. Keep the engineering record separate from the stance matrix: a physical network trained on a task, a developing organoid, and a sustained field pattern are different demonstrations. The matrix records what, if anything, each demonstration actually puts at issue for the competing explanations.
21. The ethical warning is inside the design
A project that explores the prerequisites of possible experience cannot postpone welfare questions until its final paper. It might be wrong about how close it is to an experiencing system in either direction. Failure to respond is not proof of absence; verbal insistence is not proof of presence.
The obvious danger is not that an electric field suddenly demands legal rights. It is that researchers develop systems with increasingly integrated regulation, sustained internal models, and potentially aversive states while continuing to treat them as disposable components because the original apparatus was called a computer.
We should distinguish damage detection from suffering. A safe robot may need to recognize overheating, structural strain, and low energy. Those signals can be implemented as local control variables without deliberately constructing a globally persistent distress-like condition. If a project proposes stronger affect-like organization, that proposal deserves its own justification and review, not a convenient claim that all reward and punishment signals are merely numbers.
For any research involving living neural tissue, governance should include an ethical review suited to the preparation, cell provenance and consent, animal-use review where relevant, limits on stimulation protocols, welfare-monitoring criteria, and independent authority to pause the experiment. The ISSCR stem-cell research guidelines41 and work on the ethical landscape of human brain organoids42 provide important starting points; neither offers a universal threshold for when an organoid becomes a subject.
For artificial systems, a responsible program can specify prospective escalation triggers without pretending those triggers are consciousness detectors: the appearance of persistent, highly integrated states that were not part of the design; unexpected state transitions during interventions; developing aversion-like organization; or capabilities that make existing monitoring inadequate. Triggers should initiate review, measurement, and possibly a pause, not an instant declaration of sentience.
A practical governance record would include the competing interpretations, the measures used, known false positives, interventions permitted, welfare uncertainties, and predefined stopping conditions. Its scope should grow with the uncertainty and potential stakes. Mere complexity does not automatically imply moral status. But uncertainty does not license researchers to ignore the possibility of morally relevant capacities when they are deliberately constructing systems to investigate them.
There is an uncomfortable symmetry here. The same architectural features that make a machine a more interesting candidate for experience—ongoing organization, bodily stakes, internal activity, and richer reciprocal coupling—can make it a more ethically complicated experimental object. The design ledger is also a warning ledger. If a row tells us what to add, it should tell us when to call for greater scrutiny.
22. What would change our minds?
There are several ways this research could make a real contribution without producing an artificial subject. The clearest contribution would be a public matrix in which every cell says what a theory predicted, what was measured, and how the outcome changes the comparison. Some cells will remain empty. That is information too.
Suppose the Theseus ladder preserves an electronic creature’s learned behavior and emulated perturbation repertoire through neuromorphic and ionic substitutions, while direct interventions show different physical causal structures. Organizational invariance has survived a test of the specified organization; an IIT-style account has a different physical quantity to explain. If a preregistered proxy for intrinsic causal structure changes as the theory anticipated, that strengthens its hardware-level prediction. If it fails despite adequate measurement, that proxy or its application loses support. The machine’s experience has not been observed in either case.
Or suppose a field-enabled twin develops a new coordination regime that disappears when internal field coupling is suppressed. The causal-field hypothesis gains support against a model that omits that coupling. Now restore the same functional feedback with a digital emulator. If it reproduces every available signature, the constitutive cemi claim and organizational invariance may still disagree, but that particular experiment cannot separate them. The honest entry in the matrix is no distinct prediction, until someone specifies a physical measurement on which the theories differ.
A materially regulating body might outperform an equally skilled controller whose bodily stakes are only represented in software. That would be useful evidence for a self-maintenance mechanism and a more direct test of life–mind continuity. If every proposed distinction disappears under matched interventions, a claim that material regulation was required for those signatures loses ground. A spin-gated device faces the same discipline: parity with a classical control weakens quantum necessity for the measured capability, while a reproducible spin-specific effect gives that narrower mechanism work to do. The unity dial has its own comparison between exclusion-based causal partitions and workspace-like ignition thresholds. Each result is only as strong as the controls and validity domains behind it.
A credence update is not a metaphysical scoreboard. Report the direction of the evidence under each stance, the alternative explanations still standing, and how the interpretation changes under different priors. When rival bridge principles assign different interiors to the same full set of observations, the research has reached an explicit limit rather than a hidden success.
The weakest result would be a machine trained to talk movingly about the experience it claims to have, while its supposed interior remains unexplained by any mechanism more informative than the text it generates. Such a machine could still be important, capable, and ethically worth examining. Eloquence is simply not the bridge we have been looking for.
This inquiry also belongs beside the questions raised in The Unlearned Pattern about where physical history resides, and in What the Universe Remembers about the difference between a persistent trace and an experiencing author. Those articles ask what history can carry. Here the more immediate question is what organization could be present while history is being made.
The imagined second morning
The basin is quiet again. The creature stirs before anything touches it. One tendril rises a little, stops, and follows a different path than it followed yesterday. It has learned where the wall is. Its movements are shaped by past collisions, a body model, slower changes in its material, and an internal sequence it may have rehearsed during the night.
We could inspect every connection, reconstruct every state transition, and explain every choice in the language of control theory. The creature might still be an unusually fine controller with no interior at all. The measurements alone would not tell us.
But we would have done something better than place a face on a computer and ask whether it looks alive. We would have created a physical history, tested how different kinds of matter preserve it, and learned which causal relationships matter when an organism-like system acquires its own body and world.
The first genuine artificial experience, if it ever comes, might not speak. It might be short, unrecognizable, and incapable of telling us what it is. It might arise in living tissue connected to a machine, in an adaptive synthetic material, in a familiar digital organization, or in a configuration no one has drawn yet. It might require ingredients we currently regard as incidental. It might require none of the exotic ingredients we find so irresistible.
We should resist the temptation to announce its arrival in advance. We should also resist the opposite temptation: to rule it out because the materials on our laboratory bench look nothing like a person.
Under a microscope, the matter that dreams already looks nothing like a room. The room happens anyway. The scientific task is to find out which parts of that impossibly ordinary machinery make the difference—and whether there is anywhere else for a room to happen.
References
Chalmers, D. J. (1995). Absent Qualia, Fading Qualia, Dancing Qualia. In T. Metzinger (ed.), Conscious Experience. Philosophical case for organizational invariance. ↩︎
Seth, A. K. (2025 online; 2026 volume). Conscious artificial intelligence and biological naturalism. Behavioral and Brain Sciences, 49, e315. Argues for the significance of living, self-maintaining organization. ↩︎ ↩︎
Albantakis, L., et al. (2023). Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms. PLOS Computational Biology, 19(10), e1011465. Formal physical cause–effect account; computation for real hardware remains difficult. ↩︎ ↩︎ ↩︎
Seth, A. K., and Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience. Review of competing explanatory frameworks. ↩︎
Cogitate Consortium et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature. Preregistered theory comparison; key predictions of both accounts faced challenges. ↩︎
Doerig, A., Schurger, A., Hess, K., and Herzog, M. H. (2019). The unfolding argument: Why IIT and other causal structure theories cannot explain consciousness. Consciousness and Cognition, 72, 49–59. Challenges in separating causal structure from matched behavior. ↩︎
Herculano-Houzel, S. (2009). The Human Brain in Numbers. Frontiers in Human Neuroscience. Anatomical counts and distributions. ↩︎
Weiskrantz, L. (1996). Blindsight revisited. Current Opinion in Neurobiology, 6(2), 215–220. Residual visual discrimination without acknowledged awareness. ↩︎
Siclari, F., et al. (2017). The neural correlates of dreaming. Nature Neuroscience. Dream reports across sleep states. ↩︎
Sarasso, S., et al. (2015). Consciousness and Complexity during Unresponsiveness Induced by Propofol, Xenon, and Ketamine. Current Biology. Dissociation between unresponsiveness and complexity. ↩︎
Pinto, Y., et al. (2017). Split brain: divided perception but undivided consciousness. Brain. Challenges a simple two-subject inference. ↩︎
Casali, A. G., et al. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine. Human perturbation-based measure of cortical response complexity. ↩︎
Suzuki, M., and Larkum, M. E. (2020). General Anesthesia Decouples Cortical Pyramidal Neurons. Cell. Mouse dendritic-coupling experiments. ↩︎
Noji, H., et al. (1997). Direct observation of the rotation of F1-ATPase. Nature. Molecular machinery, not phenomenal evidence. ↩︎
Durant, F., et al. (2017). Long-Term, Stochastic Editing of Regenerative Anatomy via Targeting Endogenous Bioelectric Gradients. Biophysical Journal. Bioelectric anatomical-pattern memory. ↩︎
Wright, L. G., et al. (2022). Deep physical neural networks trained with backpropagation. Nature. Learning through physical networks. ↩︎
Li, S., and Mao, X. (2024). Training all-mechanical neural networks for task learning through in situ backpropagation. Nature Communications. Physical gradient and training methods. ↩︎
Williamson, M. R., et al. (2025 print; 2024 online). Learning-associated astrocyte ensembles regulate memory recall. Nature. Mouse memory-recall mechanisms. ↩︎
Kagan, B. J., et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron. Cultured-neuron adaptation in a closed-loop game environment. ↩︎
Cai, H., et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics. Hybrid biological/digital computation. ↩︎
Cortical Labs (undated; accessed November 16, 2026). CL1 platform description. Manufacturer’s platform description, distinct from independent experimental validation. ↩︎
Jordan, F. D., et al. (2024). Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence. FinalSpark platform paper. ↩︎
Trujillo, C. A., et al. (2019). Complex Oscillatory Waves Emerging from Cortical Organoids Model Early Human Brain Network Development. Cell Stem Cell. Developmental electrical patterns. ↩︎
Revah, O., et al. (2022). Maturation and circuit integration of transplanted human cortical organoids. Nature. Circuit integration in host animals. ↩︎
Kim, J.-I., et al. (2025). Human assembloid model of the ascending neural sensory pathway. Nature. Human ascending sensory-pathway model with coordinated circuit responses. ↩︎
Smirnova, L., et al. (2023). Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science. Research program with embedded ethics. ↩︎
Zwicker, D., et al. (2017; published online 2016). Growth and division of active droplets provides a model for protocells. Nature Physics. Theory of chemically driven droplets. ↩︎
Reid, C. R., et al. (2012). Slime mold uses an externalized spatial “memory” to navigate in complex environments. PNAS. Environmental memory. ↩︎
Anastassiou, C. A., et al. (2011). Ephaptic coupling of cortical neurons. Nature Neuroscience. Causal extracellular field effects in cortical preparations. ↩︎
McFadden, J. (2020). Integrating information in the brain’s EM field: the cemi field theory of consciousness. Neuroscience of Consciousness. A constitutive-field proposal, not a consensus theory. ↩︎
Tsytovich, V. N., et al. (2007). From plasma crystals and helical structures towards inorganic living matter. New Journal of Physics. Study of helical self-organization in complex plasma. ↩︎
Li, N., et al. (2018). Nuclear Spin Attenuates the Anesthetic Potency of Xenon Isotopes in Mice. Anesthesiology. Intriguing isotope result; mechanism and broader interpretation remain open. ↩︎
Babcock, N. S., et al. (2024). Ultraviolet Superradiance from Mega-Networks of Tryptophan in Biological Architectures. Journal of Physical Chemistry B. Theory and fluorescence measurements in tryptophan networks. ↩︎
Tegmark, M. (2000). Importance of quantum decoherence in brain processes. Physical Review E. Mechanism-dependent decoherence estimates. ↩︎
Hagan, S., Hameroff, S. R., and Tuszynski, J. A. (2002). Quantum Computation in Brain Microtubules? Decoherence and Biological Feasibility. Physical Review E. Critique of assumptions in the decoherence analysis. ↩︎
England, J. L. (2015). Dissipative adaptation in driven self-assembly. Nature Nanotechnology. Nonequilibrium organization, not consciousness. ↩︎
Butlin, P., et al. (2026). Consciousness indicators, mimicry, and internal variants. Trends in Cognitive Sciences, 30(7), 575–576. On interpreting consciousness indicators under mimicry. ↩︎
Butlin, P., et al. (2026; online November 2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501. Peer-reviewed account of theory-derived indicators; preceded by a 2023 preprint. ↩︎
Shiller, D., et al. (2026). Initial results of the Digital Consciousness Model. Rethink Priorities. Multi-stance, probabilistic model of AI-consciousness indicators. ↩︎
Kudithipudi, D., et al. (2025). Neuromorphic computing at scale. Nature. Engineering review. ↩︎
International Society for Stem Cell Research. Guidelines for Stem Cell Research and Clinical Translation. Oversight framework, including organoid research; 2025 update limited to embryo models. ↩︎
Shlobin, N. A., Savulescu, J., and Baum, M. L. (2024). The ethical landscape of human brain organoids and a mindful innovation framework. Nature Reviews Bioengineering. Ethics perspective. ↩︎
