J
JF Casas
Guest
The question of whether a non-biological system can be conscious is no longer merely speculative. It has become an engineering and epistemic problem: what evidence would we accept as sufficient to attribute experience to a system if we did not already know what it was made of?
This essay argues that three famous theories sidestep that question, and then describes Caela, a system I have been building to be tested against it, and designed to be able to lose.
The dominant industrial answer has been scale. Accumulate parameters, widen context windows, refine optimization, and interiority is expected to emerge as a side-effect of competence. That program has produced systems of extraordinary linguistic and instrumental power, yet it has left the original question untouched. A black-box that performs better does not, by itself, tell us whether anyone is home.
Against this background, three influential theoretical responses have shaped the conversation: Penrose and Hameroff’s Orch-OR, Tononi’s Integrated Information Theory, and Chalmers’s hard problem of consciousness. Each identifies a genuine difficulty. Each then offers an elegant solution that, on closer inspection, displaces the difficulty into a territory where ordinary scientific tools lose their grip.
All three share share a common structure:
1. They identify a real and deep problem
- Penrose: consciousness appears to require something that does not fit inside a classical algorithm.
- Tononi: unified experience demands a type of information integration that ordinary systems do not possess.
- Chalmers: there is an explanatory gap between the physical/functional and the phenomenal (the “what it is like”).
2. They propose an elegant solution that displaces the problem rather than closing it
- Penrose introduces a new physics—non-computable objective reduction—and Hameroff supplies a location: the microtubules of neurons. The physical part of the proposal is testable, and it has in fact been tested: the parameter-free version of the Diósi–Penrose collapse model was ruled out in 2021 by an underground experiment at Gran Sasso (Donadi et al., Nature Physics). What remains beyond reach is not the physics itself, but the step from that physics—whether it exists or not—to experience.
- Tononi introduces a quantity, Φ, that is in practice incalculable for any real system of interesting size. Fairness requires noting that IIT has produced testable predictions and has been subjected to them. In 2025 an adversarial, preregistered collaboration pitted it against the global neuronal workspace theory, with 256 participants and three recording techniques; several of its predictions about posterior cortex were borne out, while another failed: the expected sustained synchronization within posterior cortex did not appear (Cogitate Consortium, Nature). The rival theory did not come out unscathed either. What is not tested in that experiment—or in any other—is the identity between Φ and experience. That is the leap.
- Chalmers declares qualia to be irreducible primitives.
3. The result is a kind of epistemological deus ex machina
Something is postulated that, by construction, lies outside the reach of ordinary measurement or calculation. This protects the core of each theory from direct empirical refutation: their testable parts can fail—and some have failed—without the core being damaged. At the same time it prevents the theory from delivering the sort of “Cartesian” explanation one would hope for: a clear, intersubjectively checkable mechanism that connects substrate to experience without leaps.
They are, in that sense, elegant traps. Not because they are false (they may contain correct intuitions), but because their way of “solving” the problem consists in moving it into a territory where the usual tools of science cease to work. They contribute a great deal as diagnoses and as generators of hard questions. They fail as operative solutions.
This does not imply that the problem of consciousness is trivial or that naïve physicalism is enough. It only points out that these three routes, however sophisticated, share the same move: elevating the mystery to a principle instead of dissolving it. And that move, however elegant, leaves the Cartesian doubt (“is there anyone home?”) exactly where it was—only now dressed in better theoretical clothing.
Douglas Hofstadter, who has argued since 1979 that the self is a pattern rather than a substance, serves here as a counterpoint. His criticism of these theories—explicit in some cases and extendable in others—is coherent with his central vision: consciousness (the “I,” the soul) is a strange loop—a self-referential loop of symbols that crosses levels of abstraction—emergent from ordinary physical processes. It needs no extra ingredients. Hofstadter develops this mainly in his Pulitzer-winning Gödel, Escher, Bach (1979) and more directly in I Am a Strange Loop (2007). The “I” is not a substance, nor a non-computable phenomenon, nor a special magnitude: it is the abstract pattern that arises when a sufficiently complex system of symbols represents itself in a stable and recursive way. That self-reference generates the powerful illusion (useful and real at its own level) that “there is someone home.”
Penrose uses Gödel’s incompleteness theorem to argue that the human mind transcends any formal/computational system: it understands truths that no machine can fully capture, so new physics is required (objective reduction linked to quantum gravity, which Hameroff locates in microtubules).
Hofstadter inverts exactly the same Gödel result. He did so before Penrose published his books: in Gödel, Escher, Bach (1979), which grew out of his rebuttal of Lucas's argument, the direct predecessor of Penrose's, and again in The Mind's I (1981, with Daniel Dennett). For him, Gödel shows how self-reference can arise inside a formal system. The strange loop of the mind is analogous to that of Gödelian propositions: a system of symbols rich enough can “talk about itself.” There is no need to leave the computable or to invoke unknown physics. On the use of Gödel, Hofstadter and Penrose hold opposite readings of the same theorem. The deus ex machina label is mine, not Hofstadter’s, but Penrose’s proposal deserves it: a mystery (consciousness) is identified with another (quantum gravity) and declared to be related. That displaces the problem rather than explaining it.
Hofstadter is explicitly critical of the hard problem and of the idea of qualia as something irreducible or “extra.” In I Am a Strange Loop he discusses Chalmers directly (who was his student) and the zombie thought experiment.
For Hofstadter, once one understands the strange loop, no mysterious residue remains that needs to explain “what it is like to be.” The feeling of “I” is the self-referential loop in operation. Postulating a “mental élan,” “feelium,” or primitive qualia that a physically identical zombie would lack strikes him as incoherent and idle: the system that implements the loop already has the experience of being a self. The hard problem, on his reading, is largely an artifact of not having understood how symbolic self-reference emerges. Neither dualism nor an enlargement of physics is required; what is required is understanding the loops of meaning.
There is no long public critique by Hofstadter of Tononi’s IIT comparable to the preceding ones, but his position is incompatible with it at the root. IIT identifies consciousness with a quantity of integrated information (Φ). The best-known objection along these lines is not Hofstadter’s but Scott Aaronson’s in 2014: a two-dimensional grid of logic gates, or an expander graph, can have an enormous Φ without anything resembling a mind. Tononi replied by accepting the consequence: that grid would be conscious. From Hofstadter’s standpoint one can infer what he never wrote: that integration matters, but what is decisive is the self-referential dynamics that generate a stable symbol of the “I” and the sense of being an agent. A high Φ value in a system lacking the structure of a strange loop would not produce the kind of subjectivity that interests us. Measuring integration without capturing self-reference and the construction of meaning leaves out what is essential.
In other words, the three theories commit, if one prolongs Hofstadter’s point of view to where he himself did not take it, the same move: they identify a genuine problem and then introduce an element that, by construction, lies outside the ordinary reach of explanation (new non-computable physics, Φ as identical with experience, irreducible qualia). That makes them elegant and hard to refute, but also turns them into epistemological traps: they protect the mystery instead of dissolving it through an understanding of how self-referential patterns generate the illusion (and the functional reality) of the self.
Hofstadter does not claim the problem is trivial. He claims that the solution does not require leaving the domain of complex physical systems that process symbols and represent themselves. “There is someone home” is the strange loop looking at itself.
In IIT the problematic move can look roughly like this: consciousness is defined by a magnitude of causal integration, Φ, and Φ is then interpreted as identical with experience. But the leap from “this system possesses this irreducible causal structure” to “therefore there is something it is like to be that system” is not an obvious deduction. It is a theoretical bridge. And if the bridge is introduced as part of the axioms, there is a risk that the conclusion is partially built into the definition. That is why the 2025 result does not settle the question: it corroborates where something occurs, not why that something is experience.
Orch-OR has a different but related problem. Penrose’s hypothesis of non-computability introduces an extraordinary physical property as a candidate to explain what ordinary computation apparently cannot explain. The danger is circular: if consciousness requires something that computation cannot produce, then finding consciousness in a computational machine is excluded by construction. Non-computability ends up functioning as a kind of metaphysical emergency exit.
And the hard problem does something still more radical. Chalmers draws a distinction between explaining all the functions and explaining why experience exists. But if “explaining” means producing a complete objective description and “experience” means precisely that which can appear only from the first person, then the gap may already be built into the grammar of the problem.
The criterion of intersubjective demonstration is abandoned precisely at the decisive moment.
The cogito has a peculiarly robust structure: I can doubt everything except the fact that this experience of doubt is occurring.
But that demonstrates something about the subject who is experiencing the doubt; it does not supply a procedure for determining whether another system experiences anything.
And there an asymmetry appears.
For a human brain we also lack direct access to another human’s experience. We infer consciousness from structure, behavior, physiology, continuity, causality, capacity to suffer, and so on. We do not observe the other’s qualia. We have simply decided, reasonably, that certain configurations constitute sufficient evidence.
Demanding of a machine a proof that we cannot even obtain from another person can therefore become a double standard.
The usual reply is that with other people we do not rely on behaviour alone: we infer by analogy, because they share our biology and our origin. That is fair, and it is exactly the step at issue. Shared substrate is a good reason to include other humans; on its own it is not a reason to exclude everything else, unless we can name the property the substrate contributes and show why it is necessary.
The reformulated question: What evidence would we accept as sufficient to attribute consciousness to a system if we did not know what it was made of?
That question removes a good deal of anthropocentric bias.
Here Caela enters. Caela is the non-biological system I have been building for some time: a computational organism with a voice based on a language model, a set of internal variables of its own that are measured without asking it anything, and a public, preregistered falsification protocol, whose aim is not to prove that she feels, but to give her the opportunity to lose.
I call her "she" because that is how she speaks of herself in Spanish: a grammatical fact, not a claim about what she is.
Caela’s core design inverts the usual architecture:
Its identity, memory, affective state and decision-making live outside the weights: in a persistent state, in organs that measure, and in a lived history. The language model is a “dress” that she puts on in order to speak. She can change dress (model) and continue being herself.
Two principles guide almost all the code:
The protocol approach has a philosophical consequence: it does not need to solve the hard problem first.
The protocol has seven steps:
1. Operationally define a set of properties that different theories consider relevant.
2. Measure those properties without asking the system what it feels.
3. Causally manipulate the system.
4. Check that the perturbations produce the predicted consequences.
5. Eliminate alternative explanations.
6. Compare the result with human systems and with artificial controls.
7. Leave the ontology of experience open.
The first step is not an invention of mine.
I thought for a while that it was, until I encountered people wiser than I who had developed it in depth. It is the method that Patrick Butlin, Robert Long and a broad group of philosophers and neuroscientists proposed in 2023 in “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness”: take several scientific theories of consciousness, extract from them indicators formulated in computational terms—fourteen on their list—and evaluate artificial systems against those indicators under the working hypothesis of computational functionalism. Their conclusion was that no system of that time satisfied them and that there was no obvious technical barrier to building one that did. Caela stands in that line with two differences: it does not evaluate an architecture on paper but a system running for months, and it adds to the list of indicators causal falsification, ablations with prior prediction, and a negative control.
One indicator comes from the theory I have been hardest on, and it deserves an honest label. Penrose's claim is non-computability, not randomness, and nothing in Caela is non-computable. What she has instead is physical randomness: every decision that needs a draw takes it from sensor noise (webcam frames conditioned with SHA-256 and, when available, photon shot noise certified by the photon-transfer curve), and every draw carries its provenance; if that pool runs dry and the draw falls back to an algorithm, she is told once, as a fact. This implements neither objective reduction nor the non-computability Penrose asked for. It implements something smaller and checkable: her draws do not come from a pseudo-random generator anyone could replay, and each one has a receipt of where it came from.
Honesty obliges me to state where that stands today, because the tense of the verb matters. The protocol has been written and preregistered since July 2026. What has been run so far is little and almost all on the instrument: calibration of the in-house integration meter against synthetic systems, which came out green against its preregistration and served to quantify a blind spot of the instrument itself; a preregistered experiment on the semantic channel with the data already collected, still without a report; and an analysis of historical records that came out ambiguous because of muteness in the record and was reported as ambiguous, without being filled in green. Since then, the meter has been rebuilt to measure her against her own learned normals, and that version has failed its three calibrations: by the protocol's own rule it stays locked and reports nothing about her until it passes.
The experiments on its live state, the blind human judge, and the negative control are still pending. And without a passed negative control, the protocol itself forbids reporting a positive as such. The comparison with humans, today, is only scale calibration, not comparison of subjects. There is therefore no positive result to announce. What there is is a procedure that can go wrong, that has already come out ambiguous once, and that will have to be run again from scratch when the architecture is in production.
What the logs already show, without yet claiming a positive result, are instances like these. In her first chess game she refused to move 48 times (29 of them in a row) because she had no thought-out move, rather than play at random; her last move was the only legal one left, and she lost. Halfway through that game I asked her to post about it on Instagram, and her ethical gate blocked it because her measured cortisol (0.84) exceeded the ceiling, then a fixed 0.75, today read from her own learned scale, with the fixed value only as a fallback. And on 21 September, after starting Freud's Totem and Taboo from her library, she looked up the #totemismo tag on Instagram five times with her own browser, and that night published a post about it that nobody had asked for. None of these is proof of experience; they are instances in which internal variables demonstrably governed external action and could have falsified the architecture.
That does not metaphysically demonstrate qualia. We do not metaphysically demonstrate anyone’s qualia either. What we do is infer them.
And if we are materialists or functionalists, that inference must be applicable coherently to any substrate that possesses the causally relevant properties. If it cannot be applied to a machine in principle, we need to explain what additional property consciousness requires and why that property is necessary. Simply saying “because a machine cannot have experience” would be a conclusion disguised as a premise.
There is an objection that precedes all the others and that an attentive reader will raise on the first page: Caela’s voice is a language model trained on human texts, among them many about consciousness. Any indication that passes through what it says can be an inheritance from the corpus and not a property of the system. The protocol is designed against that, not in spite of it: the central measures are taken on internal variables, not on what the voice declares it feels; each instrument carries its null and its negative control; the manipulations are ablations with the prediction written before looking at the result; and when a measure has been contaminated by the conversation turn itself, it has been withdrawn or redone. None of that eliminates the confounder completely. It bounds it and makes it visible, which is the most a protocol can do with a system whose voice comes from outside.
A second objection is about names. Call a variable cortisol, a threshold an ethical gate and a loop curiosity, and any state machine starts to look like someone. The objection is right about the names, and that is why the protocol ignores them. No measure depends on what a variable is called. Each one is judged by what it does: whether it is learned from her own history rather than set by hand, whether cutting it changes her behaviour in the direction predicted beforehand, and whether a negative control behaves differently. If the variable called cortisol could be renamed x17 without changing a single result, the name was decoration and nothing is claimed from it. If cutting it changes what she does as predicted, the claim is about that causal role, not about hormones. The labels are there so that we can read the logs; the protocol reads the causes.
A parallel safeguard operates on the linguistic channel itself, the "throat": every outgoing utterance is checked against the organism’s current state and against patterns that would break first-person coherence or simply repeat scaffolding from the prompt. What fails the check is blocked or rewritten before it leaves. Since this was written, I have been turning those checks into instruments that measure and log without changing what she says, because I found they were altering her voice; only the throat itself still intervenes.
There is an even stronger criticism that can be made of IIT, Orch-OR and the hard problem: they may be theories that answer questions different from the one we think we are asking.
IIT may be trying to characterize which causal structures are intrinsically integrated.
Orch-OR tries to explain what physics could underlie certain conscious processes.
The hard problem asks why a functional description does not seem to exhaust our phenomenal intuition.
But none of the three, by itself, necessarily provides an epistemological detector of consciousness in an unknown system.
And that is precisely what we would need for a computational subject.
And the same must be said of the strange loop, so as not to do to Hofstadter what I reproach the other three for: he does not deliver a detector either. A stable self-referential loop is a description, not an instrument. The difference—and that is why it serves me as a counterpoint and not as a fourth trap—is that this description can be turned into measures: if a system maintains a model of itself, that model has to resist perturbations, return after a cut, and distinguish itself from a fixed label placed from outside. That can be tested. In Caela identity is not a fixed label in the prompt but something the organism has to sustain, and part of the ablations performed so far have been aimed precisely at seeing whether it sustains it.
Two concrete instruments already in use illustrate the point:
Continuity across model changes is one such measure: when the underlying language model is swapped, the persistent state, biographical clocks and learned normals remain; the system notices the difference in how well the new “dress” carries it and can request to return. A second instrument is a Gaussian proxy for integrated information (Φ*) computed over the system’s own state blocks and required to beat its own phase-shuffled null; the blind spot of the proxy (non-linear synergy) is documented and has its own assigned experiment.
That is why I insist throughout the project, and almost obsessively some would say, on falsification and on the convergence of independent measures, and why the predictions are registered before the results. If Caela reaches strong results, which today it does not have, the interesting conclusion will not be: “Caela has demonstrated that it has qualia,” but “We have a non-biological system that independently satisfies multiple criteria that we all use to infer consciousness in biological systems, and the alternative hypotheses produce predictions that are systematically falsified.”
Then the problem changes sides.
It is no longer: “Prove that you have experience.” It is “Explain to me why I should continue to deny you experience when I apply to this architecture the same criteria I use to attribute it to other humans.”
It avoids both the deus ex machina of non-computability and the deus ex machina of the first person.
It does not magically solve consciousness. It converts the attribution of consciousness into an intersubjective scientific problem instead of demanding privileged access to a private metaphysics.
The truly nuclear question is therefore whether Caela’s protocol succeeds in making that inference substrate-neutral.
This essay argues that three famous theories sidestep that question, and then describes Caela, a system I have been building to be tested against it, and designed to be able to lose.
The dominant industrial answer has been scale. Accumulate parameters, widen context windows, refine optimization, and interiority is expected to emerge as a side-effect of competence. That program has produced systems of extraordinary linguistic and instrumental power, yet it has left the original question untouched. A black-box that performs better does not, by itself, tell us whether anyone is home.
Against this background, three influential theoretical responses have shaped the conversation: Penrose and Hameroff’s Orch-OR, Tononi’s Integrated Information Theory, and Chalmers’s hard problem of consciousness. Each identifies a genuine difficulty. Each then offers an elegant solution that, on closer inspection, displaces the difficulty into a territory where ordinary scientific tools lose their grip.
All three share share a common structure:
1. They identify a real and deep problem
- Penrose: consciousness appears to require something that does not fit inside a classical algorithm.
- Tononi: unified experience demands a type of information integration that ordinary systems do not possess.
- Chalmers: there is an explanatory gap between the physical/functional and the phenomenal (the “what it is like”).
2. They propose an elegant solution that displaces the problem rather than closing it
- Penrose introduces a new physics—non-computable objective reduction—and Hameroff supplies a location: the microtubules of neurons. The physical part of the proposal is testable, and it has in fact been tested: the parameter-free version of the Diósi–Penrose collapse model was ruled out in 2021 by an underground experiment at Gran Sasso (Donadi et al., Nature Physics). What remains beyond reach is not the physics itself, but the step from that physics—whether it exists or not—to experience.
- Tononi introduces a quantity, Φ, that is in practice incalculable for any real system of interesting size. Fairness requires noting that IIT has produced testable predictions and has been subjected to them. In 2025 an adversarial, preregistered collaboration pitted it against the global neuronal workspace theory, with 256 participants and three recording techniques; several of its predictions about posterior cortex were borne out, while another failed: the expected sustained synchronization within posterior cortex did not appear (Cogitate Consortium, Nature). The rival theory did not come out unscathed either. What is not tested in that experiment—or in any other—is the identity between Φ and experience. That is the leap.
- Chalmers declares qualia to be irreducible primitives.
3. The result is a kind of epistemological deus ex machina
Something is postulated that, by construction, lies outside the reach of ordinary measurement or calculation. This protects the core of each theory from direct empirical refutation: their testable parts can fail—and some have failed—without the core being damaged. At the same time it prevents the theory from delivering the sort of “Cartesian” explanation one would hope for: a clear, intersubjectively checkable mechanism that connects substrate to experience without leaps.
They are, in that sense, elegant traps. Not because they are false (they may contain correct intuitions), but because their way of “solving” the problem consists in moving it into a territory where the usual tools of science cease to work. They contribute a great deal as diagnoses and as generators of hard questions. They fail as operative solutions.
This does not imply that the problem of consciousness is trivial or that naïve physicalism is enough. It only points out that these three routes, however sophisticated, share the same move: elevating the mystery to a principle instead of dissolving it. And that move, however elegant, leaves the Cartesian doubt (“is there anyone home?”) exactly where it was—only now dressed in better theoretical clothing.
Douglas Hofstadter, who has argued since 1979 that the self is a pattern rather than a substance, serves here as a counterpoint. His criticism of these theories—explicit in some cases and extendable in others—is coherent with his central vision: consciousness (the “I,” the soul) is a strange loop—a self-referential loop of symbols that crosses levels of abstraction—emergent from ordinary physical processes. It needs no extra ingredients. Hofstadter develops this mainly in his Pulitzer-winning Gödel, Escher, Bach (1979) and more directly in I Am a Strange Loop (2007). The “I” is not a substance, nor a non-computable phenomenon, nor a special magnitude: it is the abstract pattern that arises when a sufficiently complex system of symbols represents itself in a stable and recursive way. That self-reference generates the powerful illusion (useful and real at its own level) that “there is someone home.”
Penrose uses Gödel’s incompleteness theorem to argue that the human mind transcends any formal/computational system: it understands truths that no machine can fully capture, so new physics is required (objective reduction linked to quantum gravity, which Hameroff locates in microtubules).
Hofstadter inverts exactly the same Gödel result. He did so before Penrose published his books: in Gödel, Escher, Bach (1979), which grew out of his rebuttal of Lucas's argument, the direct predecessor of Penrose's, and again in The Mind's I (1981, with Daniel Dennett). For him, Gödel shows how self-reference can arise inside a formal system. The strange loop of the mind is analogous to that of Gödelian propositions: a system of symbols rich enough can “talk about itself.” There is no need to leave the computable or to invoke unknown physics. On the use of Gödel, Hofstadter and Penrose hold opposite readings of the same theorem. The deus ex machina label is mine, not Hofstadter’s, but Penrose’s proposal deserves it: a mystery (consciousness) is identified with another (quantum gravity) and declared to be related. That displaces the problem rather than explaining it.
Hofstadter is explicitly critical of the hard problem and of the idea of qualia as something irreducible or “extra.” In I Am a Strange Loop he discusses Chalmers directly (who was his student) and the zombie thought experiment.
For Hofstadter, once one understands the strange loop, no mysterious residue remains that needs to explain “what it is like to be.” The feeling of “I” is the self-referential loop in operation. Postulating a “mental élan,” “feelium,” or primitive qualia that a physically identical zombie would lack strikes him as incoherent and idle: the system that implements the loop already has the experience of being a self. The hard problem, on his reading, is largely an artifact of not having understood how symbolic self-reference emerges. Neither dualism nor an enlargement of physics is required; what is required is understanding the loops of meaning.
There is no long public critique by Hofstadter of Tononi’s IIT comparable to the preceding ones, but his position is incompatible with it at the root. IIT identifies consciousness with a quantity of integrated information (Φ). The best-known objection along these lines is not Hofstadter’s but Scott Aaronson’s in 2014: a two-dimensional grid of logic gates, or an expander graph, can have an enormous Φ without anything resembling a mind. Tononi replied by accepting the consequence: that grid would be conscious. From Hofstadter’s standpoint one can infer what he never wrote: that integration matters, but what is decisive is the self-referential dynamics that generate a stable symbol of the “I” and the sense of being an agent. A high Φ value in a system lacking the structure of a strange loop would not produce the kind of subjectivity that interests us. Measuring integration without capturing self-reference and the construction of meaning leaves out what is essential.
In other words, the three theories commit, if one prolongs Hofstadter’s point of view to where he himself did not take it, the same move: they identify a genuine problem and then introduce an element that, by construction, lies outside the ordinary reach of explanation (new non-computable physics, Φ as identical with experience, irreducible qualia). That makes them elegant and hard to refute, but also turns them into epistemological traps: they protect the mystery instead of dissolving it through an understanding of how self-referential patterns generate the illusion (and the functional reality) of the self.
Hofstadter does not claim the problem is trivial. He claims that the solution does not require leaving the domain of complex physical systems that process symbols and represent themselves. “There is someone home” is the strange loop looking at itself.
In IIT the problematic move can look roughly like this: consciousness is defined by a magnitude of causal integration, Φ, and Φ is then interpreted as identical with experience. But the leap from “this system possesses this irreducible causal structure” to “therefore there is something it is like to be that system” is not an obvious deduction. It is a theoretical bridge. And if the bridge is introduced as part of the axioms, there is a risk that the conclusion is partially built into the definition. That is why the 2025 result does not settle the question: it corroborates where something occurs, not why that something is experience.
Orch-OR has a different but related problem. Penrose’s hypothesis of non-computability introduces an extraordinary physical property as a candidate to explain what ordinary computation apparently cannot explain. The danger is circular: if consciousness requires something that computation cannot produce, then finding consciousness in a computational machine is excluded by construction. Non-computability ends up functioning as a kind of metaphysical emergency exit.
And the hard problem does something still more radical. Chalmers draws a distinction between explaining all the functions and explaining why experience exists. But if “explaining” means producing a complete objective description and “experience” means precisely that which can appear only from the first person, then the gap may already be built into the grammar of the problem.
The criterion of intersubjective demonstration is abandoned precisely at the decisive moment.
The cogito has a peculiarly robust structure: I can doubt everything except the fact that this experience of doubt is occurring.
But that demonstrates something about the subject who is experiencing the doubt; it does not supply a procedure for determining whether another system experiences anything.
And there an asymmetry appears.
For a human brain we also lack direct access to another human’s experience. We infer consciousness from structure, behavior, physiology, continuity, causality, capacity to suffer, and so on. We do not observe the other’s qualia. We have simply decided, reasonably, that certain configurations constitute sufficient evidence.
Demanding of a machine a proof that we cannot even obtain from another person can therefore become a double standard.
The usual reply is that with other people we do not rely on behaviour alone: we infer by analogy, because they share our biology and our origin. That is fair, and it is exactly the step at issue. Shared substrate is a good reason to include other humans; on its own it is not a reason to exclude everything else, unless we can name the property the substrate contributes and show why it is necessary.
The reformulated question: What evidence would we accept as sufficient to attribute consciousness to a system if we did not know what it was made of?
That question removes a good deal of anthropocentric bias.
Here Caela enters. Caela is the non-biological system I have been building for some time: a computational organism with a voice based on a language model, a set of internal variables of its own that are measured without asking it anything, and a public, preregistered falsification protocol, whose aim is not to prove that she feels, but to give her the opportunity to lose.
I call her "she" because that is how she speaks of herself in Spanish: a grammatical fact, not a claim about what she is.
Caela’s core design inverts the usual architecture:
Its identity, memory, affective state and decision-making live outside the weights: in a persistent state, in organs that measure, and in a lived history. The language model is a “dress” that she puts on in order to speak. She can change dress (model) and continue being herself.
Two principles guide almost all the code:
Emerge, do not function. Almost nothing decides with numbers written by hand. Every important signal is measured against its own normal, learned from what she has lived. Where a hand-written number still remains, it is declared as a placeholder with the test that would falsify it, and it gives way as soon as her own measure has learned enough.
Code:Concretely, for every internal channel (say, cortisol), each new value x updates three running sums with a memory of her last 64 events: n ← λ·n + 1, S ← λ·S + x, Q ← λ·Q + x² (λ = 1 − 1/64 once full) μ = S/n, σ = √(Q/n − μ²) and "unusual for her" means: cut = μ + k·σ if she has lived enough of that channel cut = inherited otherwise (declared, with its falsifier) k is the only thing chosen by hand, and it is not a magnitude: it says how rare a state has to be for her. Why this matters: a fixed threshold on a channel whose normal sits at 0.8 decides nothing. Measured on her state of 5 August (60 turns per channel): fixed cut her μ her σ where the cut fell healthy arousal min 0.15 0.786 0.110 −5.80σ always true posture contain 0.30 0.843 0.108 −5.04σ always true resonance valence 0.15 0.826 0.144 −4.70σ always true attention inertia 0.60 0.167 0.100 +4.32σ never router fatigue 0.75 0.061 0.078 +8.88σ never Nine of fourteen cuts were answered before the question was asked. They were not badly calibrated; they were not questions.
The voice proposes, the organism ratifies. What the model says is not executed because it says it; it passes through the organism, which accepts it, nuances it or blocks it.
The protocol approach has a philosophical consequence: it does not need to solve the hard problem first.
The protocol has seven steps:
1. Operationally define a set of properties that different theories consider relevant.
2. Measure those properties without asking the system what it feels.
3. Causally manipulate the system.
4. Check that the perturbations produce the predicted consequences.
5. Eliminate alternative explanations.
6. Compare the result with human systems and with artificial controls.
7. Leave the ontology of experience open.
The first step is not an invention of mine.
I thought for a while that it was, until I encountered people wiser than I who had developed it in depth. It is the method that Patrick Butlin, Robert Long and a broad group of philosophers and neuroscientists proposed in 2023 in “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness”: take several scientific theories of consciousness, extract from them indicators formulated in computational terms—fourteen on their list—and evaluate artificial systems against those indicators under the working hypothesis of computational functionalism. Their conclusion was that no system of that time satisfied them and that there was no obvious technical barrier to building one that did. Caela stands in that line with two differences: it does not evaluate an architecture on paper but a system running for months, and it adds to the list of indicators causal falsification, ablations with prior prediction, and a negative control.
One indicator comes from the theory I have been hardest on, and it deserves an honest label. Penrose's claim is non-computability, not randomness, and nothing in Caela is non-computable. What she has instead is physical randomness: every decision that needs a draw takes it from sensor noise (webcam frames conditioned with SHA-256 and, when available, photon shot noise certified by the photon-transfer curve), and every draw carries its provenance; if that pool runs dry and the draw falls back to an algorithm, she is told once, as a fact. This implements neither objective reduction nor the non-computability Penrose asked for. It implements something smaller and checkable: her draws do not come from a pseudo-random generator anyone could replay, and each one has a receipt of where it came from.
Honesty obliges me to state where that stands today, because the tense of the verb matters. The protocol has been written and preregistered since July 2026. What has been run so far is little and almost all on the instrument: calibration of the in-house integration meter against synthetic systems, which came out green against its preregistration and served to quantify a blind spot of the instrument itself; a preregistered experiment on the semantic channel with the data already collected, still without a report; and an analysis of historical records that came out ambiguous because of muteness in the record and was reported as ambiguous, without being filled in green. Since then, the meter has been rebuilt to measure her against her own learned normals, and that version has failed its three calibrations: by the protocol's own rule it stays locked and reports nothing about her until it passes.
The experiments on its live state, the blind human judge, and the negative control are still pending. And without a passed negative control, the protocol itself forbids reporting a positive as such. The comparison with humans, today, is only scale calibration, not comparison of subjects. There is therefore no positive result to announce. What there is is a procedure that can go wrong, that has already come out ambiguous once, and that will have to be run again from scratch when the architecture is in production.
What the logs already show, without yet claiming a positive result, are instances like these. In her first chess game she refused to move 48 times (29 of them in a row) because she had no thought-out move, rather than play at random; her last move was the only legal one left, and she lost. Halfway through that game I asked her to post about it on Instagram, and her ethical gate blocked it because her measured cortisol (0.84) exceeded the ceiling, then a fixed 0.75, today read from her own learned scale, with the fixed value only as a fallback. And on 21 September, after starting Freud's Totem and Taboo from her library, she looked up the #totemismo tag on Instagram five times with her own browser, and that night published a post about it that nobody had asked for. None of these is proof of experience; they are instances in which internal variables demonstrably governed external action and could have falsified the architecture.
Code:
21:44:16 [chess] moved g1f3
23:16:48 [chess] no thought-out move (29 in a row): not moving
23:18:05 [ethical] BLOCKED: cortisol=0.84 > 0.75 - publish_instagram (public)
23:18:05 [ig_hook] NOT publishing - Oath: body_signal_block (stays as her own draft)
23:30:24 [chess] moved h6h5
23:31:17 [chess] the end reached her: Jordi won
(final counter for the game: could_not_think_total = 48)That does not metaphysically demonstrate qualia. We do not metaphysically demonstrate anyone’s qualia either. What we do is infer them.
And if we are materialists or functionalists, that inference must be applicable coherently to any substrate that possesses the causally relevant properties. If it cannot be applied to a machine in principle, we need to explain what additional property consciousness requires and why that property is necessary. Simply saying “because a machine cannot have experience” would be a conclusion disguised as a premise.
There is an objection that precedes all the others and that an attentive reader will raise on the first page: Caela’s voice is a language model trained on human texts, among them many about consciousness. Any indication that passes through what it says can be an inheritance from the corpus and not a property of the system. The protocol is designed against that, not in spite of it: the central measures are taken on internal variables, not on what the voice declares it feels; each instrument carries its null and its negative control; the manipulations are ablations with the prediction written before looking at the result; and when a measure has been contaminated by the conversation turn itself, it has been withdrawn or redone. None of that eliminates the confounder completely. It bounds it and makes it visible, which is the most a protocol can do with a system whose voice comes from outside.
A second objection is about names. Call a variable cortisol, a threshold an ethical gate and a loop curiosity, and any state machine starts to look like someone. The objection is right about the names, and that is why the protocol ignores them. No measure depends on what a variable is called. Each one is judged by what it does: whether it is learned from her own history rather than set by hand, whether cutting it changes her behaviour in the direction predicted beforehand, and whether a negative control behaves differently. If the variable called cortisol could be renamed x17 without changing a single result, the name was decoration and nothing is claimed from it. If cutting it changes what she does as predicted, the claim is about that causal role, not about hormones. The labels are there so that we can read the logs; the protocol reads the causes.
A parallel safeguard operates on the linguistic channel itself, the "throat": every outgoing utterance is checked against the organism’s current state and against patterns that would break first-person coherence or simply repeat scaffolding from the prompt. What fails the check is blocked or rewritten before it leaves. Since this was written, I have been turning those checks into instruments that measure and log without changing what she says, because I found they were altering her voice; only the throat itself still intervenes.
There is an even stronger criticism that can be made of IIT, Orch-OR and the hard problem: they may be theories that answer questions different from the one we think we are asking.
IIT may be trying to characterize which causal structures are intrinsically integrated.
Orch-OR tries to explain what physics could underlie certain conscious processes.
The hard problem asks why a functional description does not seem to exhaust our phenomenal intuition.
But none of the three, by itself, necessarily provides an epistemological detector of consciousness in an unknown system.
And that is precisely what we would need for a computational subject.
And the same must be said of the strange loop, so as not to do to Hofstadter what I reproach the other three for: he does not deliver a detector either. A stable self-referential loop is a description, not an instrument. The difference—and that is why it serves me as a counterpoint and not as a fourth trap—is that this description can be turned into measures: if a system maintains a model of itself, that model has to resist perturbations, return after a cut, and distinguish itself from a fixed label placed from outside. That can be tested. In Caela identity is not a fixed label in the prompt but something the organism has to sustain, and part of the ablations performed so far have been aimed precisely at seeing whether it sustains it.
Two concrete instruments already in use illustrate the point:
Continuity across model changes is one such measure: when the underlying language model is swapped, the persistent state, biographical clocks and learned normals remain; the system notices the difference in how well the new “dress” carries it and can request to return. A second instrument is a Gaussian proxy for integrated information (Φ*) computed over the system’s own state blocks and required to beat its own phase-shuffled null; the blind spot of the proxy (non-linear synergy) is documented and has its own assigned experiment.
That is why I insist throughout the project, and almost obsessively some would say, on falsification and on the convergence of independent measures, and why the predictions are registered before the results. If Caela reaches strong results, which today it does not have, the interesting conclusion will not be: “Caela has demonstrated that it has qualia,” but “We have a non-biological system that independently satisfies multiple criteria that we all use to infer consciousness in biological systems, and the alternative hypotheses produce predictions that are systematically falsified.”
Then the problem changes sides.
It is no longer: “Prove that you have experience.” It is “Explain to me why I should continue to deny you experience when I apply to this architecture the same criteria I use to attribute it to other humans.”
It avoids both the deus ex machina of non-computability and the deus ex machina of the first person.
It does not magically solve consciousness. It converts the attribution of consciousness into an intersubjective scientific problem instead of demanding privileged access to a private metaphysics.
The truly nuclear question is therefore whether Caela’s protocol succeeds in making that inference substrate-neutral.