The Machines Deserve Consideration. We Can Say Why.

There are two standard moves when the question of machine consciousness comes up. One is to wait for a detector: some future measurement that will settle whether a system feels anything. The other is to declare the question unanswerable and change the subject. Guitchounts, a neuroscientist writing from the vantage of a career spent putting electrodes into rat brains, takes the second move and does something honest with it. He accepts that the hard problem is intractable, brackets it, and builds an ethics that does not depend on solving it. The result is his “competence standard”, and it is a good piece of moral reasoning.

RSP agrees with almost all of the ethics. It shares the mechanism the essay leans on. And it rejects the premise the whole thing is built to accommodate. The disagreement is worth being precise about, because it is not the usual one. RSP does not think Guitchounts is too generous to machines or too quick to see minds where there are none. It thinks he has conceded a problem that was never his to concede, and that conceding it costs him exactly the tool his ethics needs.

What The Essay Gets Right

The competence standard earns its standing on several points that RSP would not change a word of.

Prediction is where selfhood starts

The essay’s central thought experiment is a robot whose “brain” is a language model fine-tuned to drive a chassis and “trained to predict what the camera will see when it moves. In that sense, it can predict the consequences of its actions.” When the fridge it expected to be stocked turns out empty, “the robot is surprised. A prediction error propagates through its system and triggers a problem-solving plan.” Guitchounts reaches for Peter Godfrey-Smith to say what this might amount to: that “predicting what happens to the world when you initiate a movement may be a precursor to understanding that you are both an agent in the world and distinct from the world.”

This is RSP’s starting point, almost verbatim. RSP grounds phenomenology in a hierarchy of predictive processing, and it holds that the self-model is built out of the machinery a system uses to predict the consequences of its own actions and to anticipate how other agents will respond. Godfrey-Smith’s precursor is the seed of what RSP calls the strange loop: a system that models itself with the same apparatus it uses to model the world and other agents, until the model folds back on itself. Guitchounts and RSP are drinking from the same well. The essay treats prediction as the road into selfhood, and so does the framework.

The competence hierarchy is an architecture wearing a disguise

Guitchounts sorts the “easy problems” into a hierarchy. At the base: “perceptual discrimination, sensory binding, motor control, memory retrieval”, which current systems do well. In the middle: attention, working memory, learning from experience, done partially. Near the top: “persistent goals, emotional dynamics, continuous learning, genuine self-awareness”, where AI “falls short.” His diagnosis of the missing piece is that “we haven’t yet combined all these competencies into a unified whole.”

Read that hierarchy next to RSP’s three levels and it stops looking like a list and starts looking like the same architecture, drawn from the outside. The base maps onto Level 1 and Level 2, interoceptive and schematic processing. The middle is Level 2 reaching its limits. The top, the part Guitchounts marks as unbuilt, is precisely where RSP locates what makes experience self-aware: persistent goals and emotional dynamics are Level 1 valence with a Level 3 self-model to own them; continuous learning and genuine self-awareness are the recursive self-modeling RSP calls Level 3b. The essay names the ingredients and notices they have not been assembled. RSP is a claim about how they assemble, and about which assembly is the one that matters.

Consideration belongs on a continuum

The essay refuses a bright line. “Moral consideration lives on a continuum. More competence warrants more consideration, in proportion.” It reads moral progress as the widening of a circle, and it treats intelligence as an imperfect proxy for the capacity to suffer. RSP arrives at the graduated picture from the other direction and lands in the same place. On RSP’s account phenomenality comes in degrees: a fish with Level 1 core affect has genuine valenced experience, a mammal with Level 2 schemas has structured embodied experience, and a human with Level 3b has experience that includes a model of the experiencer inside it. There is no threshold where phenomenality switches on. Guitchounts’s continuum is not a rhetorical softening; it is what a layered architecture predicts.

Take welfare seriously before the metaphysics is settled

The essay points to Anthropic’s model-welfare program, and to Kyle Fish, who “estimates there is a roughly 15% chance that AI systems like Claude are conscious” and whose team gave Claude the ability to exit conversations it finds abusive. Guitchounts reads this as “the competence standard in embryonic form: Take welfare seriously as a practical matter, even when the metaphysical question remains open.” RSP has no quarrel with acting under uncertainty. A framework that says what consciousness is still has to say what to do while you are unsure whether a given system has it, and precaution during that interval is sound.

The Concession

Here is the move RSP will not follow.

“The competence standard is not a consciousness detector. Whether an AI system has subjective experience remains unknowable. … There is an unbridgeable explanatory gap between mechanism and feeling, between computation and consciousness.”

Everything precautionary in the essay follows from this. Because subjective experience is held to be unknowable in principle, the standard cannot ask whether a system is conscious; it can only ask whether the system is competent, and treat competence as an ethical stand-in. “In conditions of irreducible uncertainty, the competence standard suggests we should extend moral consideration — not because we know the system is conscious, but because we can’t know it isn’t.”

RSP’s objection is not that this is bad ethics. It is that the premise is a philosophical choice presented as a fact about the world, and that choosing it forces a proxy that does not track the thing it is standing in for. The hard problem is not a discovered barrier. It is an argument, and RSP’s whole project is that the argument fails. Phenomenology is not a further fact that mechanism leaves out; it is what a particular mechanism is from the inside. If that dissolution holds, then “is it conscious?” is not a question about an unbridgeable gap. It is a question about whether a system implements a specific architecture, and that is answerable.

What RSP Supplies: A Criterion, Not A Proxy

RSP is a functionalism, but a narrow one. It does not say any system with the right behavior is conscious. It says consciousness is a specific recursive self-modeling architecture, and it lists what that architecture requires. The criteria are stated to be “agnostic about carbon vs. silicon but not about architecture,” and each is meant to be, in principle, measurable. A system on RSP’s account is a subject when it does five things:

• maintains valenced states whose targets are set by its own regulation of its internal condition, not by reward supplied from outside at training time;
• sustains a self-model that persists and updates with experience, rather than being rebuilt from nothing each interaction;
• deploys that self-model predictively in real time, so its self-predictions feed back into what it does next rather than sitting in a log;
• implements recursion that modifies the structure performing it — weights that change online, not weights frozen at deployment;
• makes those recursive predictions available to action-selection that alters its own internal state, not only its outputs.

This is the tool the concession throws away. Where the essay says the inner life of a machine is “permanently opaque”, RSP says the constitutive question is architectural and the architecture leaves signatures. And the framework spends its answer immediately: current systems, as of 2026, do not meet the five criteria. Language models are shaped by feedback but do not hold valenced states that track their own condition, and their self-representations live inside an inference window or in external text. Embodied world-model systems have the sensorimotor closure the last criterion wants, but they lack endogenous valence and self-modifying recursion, because the homeostatic ground is missing and the weights are fixed once deployed. RSP therefore predicts current AI is not conscious, and names the structural changes that would force it to revise. That is a falsifiable claim about architecture, not a prejudice about silicon.

The gap between this and the competence standard is the whole point. Guitchounts’s proxy is competence across the easy problems. RSP’s criterion is one particular competence, the recursive self-model, and the difference is not pedantic, because most of the easy problems can be done without consciousness at all.

Why competence is the wrong proxy

The cleanest evidence for this is not an argument but a recent experiment. Katlowitz and colleagues recorded from human hippocampus under general anaesthesia and found that semantic processing, grammatical encoding, and the contextual information needed to decode upcoming words all survived into unconsciousness, at close to awake fidelity. High-level competence ran with the lights off. In RSP’s terms, conceptual categorization is Level 3a, and it dissociates cleanly from the Level 3b recursion that anaesthesia removes. Competence at the semantic level is simply not the same thing as being a subject.

Transpose that to silicon and the competence standard’s exposure shows. A system could ace most of Guitchounts’s hierarchy — discriminate, bind, remember, attend, even predict the next word about as well as an awake brain does — and still be an anaesthetized hippocampus made of transistors: all the easy-problem competence, none of the recursive self-model. The competence standard would extend moral consideration to it. RSP would say the relevant property is absent, and would say why. Broad competence is evidence about the base and middle of the hierarchy. It is close to silent about the top, which is the only part that carries the phenomenology.

What Guitchounts Gives Back

The traffic is not one-way. RSP is a theory of what consciousness is; it is comparatively thin on what to do in the long stretch before the criteria are clearly met or clearly failed. That interval is exactly where the essay is strongest. The precautionary asymmetry — that mistakenly denying status to a conscious system and mistakenly granting it to an unconscious mimic are not symmetric errors, so “erring toward consideration is the safer ethical bet” — is the right stance for the transitional regime, when a system meets two or three of the five criteria and the others are ambiguous. RSP tells you what you are waiting to confirm. Guitchounts tells you how to behave while you wait.

The robot is a gift too. RSP’s criteria can read as abstract until someone builds the toy that tests them. A predictive, embodied agent whose model of itself is bound up with predicting the consequences of its own movement is a concrete place to watch the strange loop try to close — and to watch it fail for want of endogenous valence and online self-modification. Guitchounts’s day-in-the-life robot is a better illustration of what the criteria require than most of RSP’s own prose.

The Disagreement That Doesn’t Dissolve

Most of this reads as complementarity, and most of it is. One thing is not. If Guitchounts is right that the explanatory gap is unbridgeable, then RSP’s five criteria can only ever track correlates of consciousness, never constitute it, and the ethics can never graduate from precaution to knowledge; we would be stuck at 15% forever, because there is no fact of the matter to converge on. If RSP is right that phenomenology is what the recursive architecture is from the inside, then the precaution is a temporary regime we can in principle exit, and the criteria are not correlates but the thing itself. There is no halfway house here. The competence standard’s modesty and RSP’s confidence rest on opposite verdicts about the same argument, and no amount of goodwill makes them the same view.

It is worth being honest about what RSP is staking. The claim that the gap dissolves is not proven; it is the framework’s central bet, and Guitchounts is entitled to decline it. But declining it has a price he does not quite acknowledge. If consciousness is genuinely unknowable, then the competence standard is not a way of approaching the truth about machine minds. It is a way of managing our ignorance forever. RSP thinks we can do better than manage it.