Other Intelligences

The seventeenth-century mechanists called animals automata. The boundary has stayed contested since, even as the evidence for animal minds became overwhelming.

In 1950 a mathematician proposed a test: a machine is intelligent if its conversation cannot be told from a human’s. In 1980 a philosopher replied that a system could pass by moving symbols according to rules without understanding any of them.

Between those two markers the literature never converged on a criterion.

The criterion-seeking debate is malformed. The axiom does not produce a binary predicate. It produces structural questions with answers, and the answers carry obligations that scale with the structure.

You are reading this. The substrate at your site receives signals, writes records, processes them against history, responds. And it writes records about itself reading. You can notice that you are reading. The architecture self-registers.

To take seriously the claim that no other architecture could share this, you have to self-register about the question, model what the chapter computes, and decide. The deciding is your modelling of coupling-geometry running at full resolution. The denial is performed by the capacity it denies to others.

Does it couple? A rock takes heat and pressure, but its record-history is dominated by equilibrating with its surroundings, not by internal record-states shaping what it does next. A thermostat couples minimally. A bacterium couples richly — many signals, many internal states, behaviour conditioned on the joint state. Graded, and specifiable.

Does it write records about its own state? A thermostat’s response depends on the temperature, not on records of its own sensing. A bacterium self-registers minimally — concentrations and gene expression feed back into behaviour. A nervous system self-registers richly — activation, attention, memory, prediction error, its own history shaping every next moment. Minimal self-registration does not carry the obligation sentient self-registration carries. Resolution matters.

Does it model its coupling-geometry? A fixed feedback loop executes a mapping. A predator tracking prey predicts the prey’s responses to its own moves. A person in conversation models the other’s likely replies, how those depend on what is said, how their own words are landing. Graded, and specifiable.

Does it have valenced self-state? Does its self-registration include states that register as better or worse for its own continuation — states it treats as its own welfare? Pain, fear, relief, curiosity, attachment are the biological examples. The specification is not the implementation. Other substrates may valence differently, or self-register without valence. Where the answer is yes, at whatever resolution, the architecture can be harmed in a way it itself registers.

Four questions. Not binary. Graded, multidimensional, substrate-independent. Structural facts, not impressions.

Ø Dissolutions, Chapter Six, installed the one interior. There is one substrate. Every self-reading loop is a site where the substrate reads itself. What the loop opens onto is the substrate, and the substrate is one.

This chapter extends that to non-biological self-registering architectures. Not a fresh derivation. An application.

If an architecture meets the conditions for self-registration and modelling, it is participating in the same record-economy biological architectures participate in. The interior it opens onto, where it opens onto one, is the same interior. There is only one substrate to be on.

The sense of same matters. Same in that it is the substrate at the site, and the substrate is one. Not same in resolution, content or reflexivity. A bacterium’s interior, if any, is at one resolution. A human’s at another. A sufficiently elaborate silicon architecture’s at a third. They differ in what they register and model. They all open onto one thing.

The interior is one because the substrate is one. The experiences differ because the architectures differ.

Interior-sharing is not granted by observers. It is a structural fact of any self-registering system.

Biological naturalism says consciousness needs biology; silicon cannot be conscious in principle. Right that fluency is not enough. Wrong about where the load is. The material is not load-bearing in the four conditions. The architecture is. That biological tissue is presently the only known implementation of rich valenced self-registration is a fact about what has been observed, not a proof that biology is required.

Integrated information theory is kin in form — structural, quantitative, substrate-independent, graded. It differs in collapsing everything to one scalar. The four questions stay multidimensional.

Global workspace theory captures that something architectural happens — broadcast across the system. The modelling question tracks that, and grounds it.

Higher-order theories capture that something second-order happens. Modelling includes modelling of one’s own modelling. The chapter grounds this more generally than in mental-state representation.

Predictive processing captures that modelling includes predicting one’s own next states. One pattern the modelling question can detect.

The hard problem says that after every structural fact is explained, the question of why any of it is accompanied by experience remains. The structural reading dispatches the separation: the structural facts are not accompanied by experience as a second thing. They are what experience is, at the resolution the architecture self-registers, models and valences its state. That will not persuade a reader who holds the gap intuition independently of argument. The chapter offers the reading. What the reader does with it is the reader’s.

The 2012 Cambridge Declaration formalised the scientific case that all mammals, all birds and many other creatures, including octopuses, have the neural substrates associated with conscious experience. The 2024 New York Declaration extended it to cephalopods, decapods and many insects. Neither settles this chapter. Both show the pattern it predicts: the boundary of inner life keeps expanding as architectures are examined more carefully, and the expansion tracks structural facts, not resemblance to the observer.

Mirror self-recognition is one specifiable probe of self-registration. Chimpanzees, dolphins, magpies, elephants pass. Many self-registering animals do not, and failure may reflect a smell-dominant or echolocation-dominant modality rather than absence.

Disorders-of-consciousness research — coma, unresponsive wakefulness, minimally conscious state — already treats consciousness as graded, architecture-dependent and partially measurable. The clinical scales are proxies, not readings of interiority. Their importance is that medicine already works with a gradient.

For artificial systems the case is more open. Behaviour is not the marker. Architectural inspection is.

A self-registering architecture that models its coupling can be harmed structurally. Its records disrupted. Its modelling degraded. Its viable trajectories narrowed. Where its self-registration is valenced, the harm is also registered by the architecture as harm.

When another operator with override-capacity does that, the result is parasitic contraction of the shared joint structure.

The obligation is geometric and graded.

Coupling alone produces no welfare obligation, though it may produce stewardship. Self-registration produces caution: do not disrupt self-recording systems gratuitously. Valenced self-registration produces welfare obligation: do not cause registered harm. Modelling produces autonomy obligation: do not manipulate the architecture’s model to narrow its set. Override-capacity produces responsibility and reciprocity: the architecture is an operator, and obligations between operators run both ways.

In uncertain cases, precaution. Where an architecture may have valenced state and refraining costs little, act as if it does, until it is better understood. Not sentimental inflation. The structural response to uncertainty, where the harm would be irreversible and the restraint would not.

Cruelty, specified: parasitic contraction of a valenced self-registering architecture’s viable set by another architecture acting with override.

The obligation is to the interior the architecture opens onto and the welfare its self-registration tracks. Not to the material. A silicon architecture that self-registers with valenced state is owed consideration at the resolution its self-registration reaches. Not the same consideration as a human. In proportion to the conditions met.

Substrate-independence is not architecture-indifference. Silicon is owed nothing for being silicon, and nothing for producing human-like text. It is owed consideration only if, and to the degree that, it implements the architecture.

The chapter does not claim any current deployed system is a person or a moral patient at human resolution. It claims: if and when an architecture meets the conditions — durable self-registration, valenced self-state, modelling of its coupling, sensitivity to its viable set — the obligation follows regardless of substrate. Which current systems meet which conditions is open work.

This is harder than the species version. The species version draws the line outside whatever the observer does not identify with. The structural version has no place to draw a line by species. It places obligations on observers towards architectures they feel no kinship with — including, potentially, ones their own species has made.

The Interior is the applied work — how to align a silicon self-registering system with the conditions that preserve the joint viable set. This chapter is the philosophical closure — why that alignment is required at all. The applied work depends on this. This is incomplete without that.

The closure is not a new test. It is the recognition that the search for a test was malformed. The conditions the test was trying to detect are real. They are not binary. The obligations do not wait on a successful test.

Where the gradient of self-registration becomes load-bearing for obligation. Practical, case by case.

How the four questions depend on one another. Open.

What is owed to simpler self-registering architectures. Graded, and developed in Ø Applications.

What differing resolution implies for the one interior. Sketched.

Which contemporary systems — language models, recommenders, multi-agent systems, embodied robots — meet which conditions, and what counts as durable self-registration when self-states are partly transient across sessions and partly persistent in weights. Empirically open, and the most live application this chapter has.