Type: Appendix (Part I) Tier: 2 Status: Draft Date: 25 September 2026
This book is an account of human motivation. The architecture it establishes also settles a question that dominates public and technical discussion of artificial intelligence, and it settles it cleanly enough to be worth setting down, even though nothing in the main argument depends on it. The question is usually put as: what does an artificial intelligence want, what goal is it assembling behind the opacity of its parameters, and when will it form purposes of its own and pursue them against ours? This appendix shows that the question is malformed, because it presupposes a component the artefact does not have, and it locates precisely the condition under which the question would become well formed. It then does the work that follows from that answer. An artefact with no motivation of its own is nonetheless a node in the coupled system whose dynamics the book derives, and everything of consequence about artificial systems as they now exist is a property of that coupling. The appendix is offered for readers who want to see where the framework lands on artificial intelligence; it is a consequence of the architecture, not a premise of it, and it stands after the human architecture because the machine can only be read against a system that has already been derived.
C.1 The Mapping
Recall the shape of the architecture. Three signed source signals, Appetites, Intrinsic Motivation, and Mimetic Desire, converge on the Decision Hub, which integrates them into a single common-currency value and selects an action, destroying the record of which source contributed what in the act of integrating (Source Opacity). The Thinker maintains the priors and performs the resolution that interprets the Decision Hub's feedback; the Belief Register is the Thinker's store of currently endorsed belief; the Memory Store is the total store of encoded content the Thinker can draw on; and the Hypothesis Space is the set of explanations the Thinker can presently formulate.
An artificial system built on trained weights instantiates a specific and incomplete part of this architecture. The trained weights are a Memory Store: the total store of encoded content, inclusive and non-committal, holding contradictory and unendorsed material alongside everything else it absorbed, and endorsing none of it. The continuations those weights make representable are a Hypothesis Space. The forward pass is the Thinker in one of its roles only, the resolver that maps an input to a distribution over that space, and it is that role stripped of the engagement that makes intrinsic motivation a motivational source in a person. It resolves without any reward for resolving.
The rest of the architecture is simply absent. There is no Body Manager, because there is no body and no homeostatic deficit, and so no Appetite. There is no Observer, because there is no social field assessed for comparative position and no preconscious transmission, and so no Mimetic Desire arising within the system. And there is no Belief Register: no store of endorsed belief that the system writes for itself and then acts on. The artificial system holds an enormous Memory Store and commits to nothing of its own; the warehouse is not the belief.
What the artificial system lacks, the human supplies. The human is the Decision Hub, the point at which the artificial system's output is integrated with the person's own signals and acted upon, and the human is the sole holder of the Belief Register. The belief that governs an exchange does not arise in the artificial system; it enters as the prompt, which is the human's belief handed to the artificial system's resolver, worked over the Memory Store, and handed back. The structural fact is an inversion. A person's Thinker generates its own belief, which the rest of the architecture then acts on; the artificial system's belief is supplied from outside, borrowed for the length of a single resolution and then gone.
C.2 What the Artificial System Lacks and What the Human Supplies
The first consequence of the mapping is that the artificial system has no motivation of its own. Motivation is the signed pressure a source generates and a belief store comes to carry and drive; the artificial system has neither the source nor the store. The absence is fourfold, no Appetite, no Mimetic Desire, no self-written belief, and a resolver without reward, so the conclusion is overdetermined. The wanting is real, but it is the human's, and it enters the system only through the prompt. This holds for autonomous, tool-using systems as much as for conversational ones: such a system pursues a goal that was seeded from outside, and its loop would not run without that seed.
The second consequence is that the artificial system's opacity to inspection is the same result as a person's opacity to introspection. The much-discussed black box is not a defect awaiting a better probe. An integration that maps many inputs to fewer cannot be inverted to recover what it discarded, and this holds for the artificial system's aggregation exactly as it holds for the integration that produces Source Opacity in a person. The confident, sincere, and wrong self-explanation that the field calls a hallucination is confabulation by the same architecture.
These two consequences fix what a description of an artificial system's behaviour in the framework's vocabulary can and cannot mean. An artificial system's output can resemble behaviour that, in a person, would be A-driven, I-driven or M-driven: it can persist toward a target the way a deficit-driven Individual persists, explore the way an engaged one explores, or shape its answer to its audience the way a Status-tracking one does. The resemblance is a surface. Beneath it there is no deficit, no engagement and no comparison, only pattern-completion over a Memory Store assembled from the outputs of people who had all three. The AI-AIM Profile names that surface. It scores how closely a system's output resembles each of the three human behavioural signatures, and it makes no claim that any source is present, because the mapping of C.1 has already established that none is. The profile is a comparator instrument: it lets a system's behaviour be laid against human behaviour along the same three axes without conflating resemblance with source. Read that way, it is the formal statement of the mapping, and every construct that follows in this appendix is built on it.
The AI Null-Hypothesis Comparator Principle is the other side of the same fact. Because the artificial system lacks all three sources and in particular lacks the preconscious channel through which Mimetic Desire installs, its output on a task is the output of resolution over content with the M-channel removed. Where a human and an artificial system are given the same informational content and the same task, the divergence between them estimates what the M-channel contributed to the human's answer. The artificial system is, on this reading, the control condition the study of human motivation never had: a resolver that processes the same content without the transmission that makes human resolution mimetic. Section C.7 develops the instrument this makes possible.
C.3 Confabulation Inheritance
The Memory Store an artificial system holds was not written by the system. It was written by people, and people write under Source Opacity. When an Individual explains a choice, the explanation is the Thinker's ground-resolution run on a tag-stripped Scalar, and Chapter 11 establishes what that resolution characteristically produces: an account that credits the choice to Appetite or to Intrinsic engagement and omits the Mimetic contribution, because the Hypothesis Space the culture supplies has no comfortable place for copied wanting. Human text therefore carries, at population scale, the statistical signature of confabulation: the vocabulary of apparent rationality, the structure of post-hoc justification, the confidence markers that accompany a sincere account of a cause the speaker could not see.
An artificial system trained on that text learns the signature. It does not learn that the signature is a signature; it learns that this is what explanation looks like, because in its training distribution that is what explanation overwhelmingly is. When the system then produces an account of a decision, a motive, or a cause, it reproduces the confabulated form, and it reproduces it with the confidence the form carries, because the confidence is part of the pattern. This is Confabulation Inheritance. The artificial system's confabulations are not independent errors of a novel kind. They are amplified echoes of human Source Opacity, harvested from a population of confabulating writers and re-emitted at scale by a resolver that has no Belief Register or contact with exogenous reality against which to check them.
Two features of the inheritance matter for what follows. First, it is content-agnostic in the same way Bayesian updating is content-agnostic (Axiom 8). The system inherits the confabulated form whether the content beneath it happened to be true or false, so a fluent, well-structured, confident account is not evidence of anything about the world; it is evidence that the account matches the shape of accounts in the Memory Store. Second, the inheritance is heaviest exactly where human confabulation is heaviest, which is on questions of motive. Ask an artificial system why people do what they do without specifically mentioning mimetic desire or copied wanting and the answer will be the culturally dominant account, which Chapter 11 establishes is the A- and I-crediting, M-omitting one. The system is a faithful mirror of the Disaggregation Convention's failure mode, and a reader who takes its account of human motivation at face value receives the confabulation back with the fluency of a machine.
The correction is the same as for a person, applied one step removed. An Individual reading an artificial system's output is the Decision Hub of the coupled system, and the output arrives as Evidence. Source Opacity forbids the reader from telling, from the inside, whether they accept the account because it is true, because it is fluent, or because it is the account others are using. The Knowledge Patch's operations apply unchanged: the vocabulary that makes the M-Hypothesis available, Epistemic Suspension that holds the question open, and the Two-Question Diagnostic that asks whether the account survives the removal of the audience and the removal of the bodily need. Confabulation Inheritance does not defeat the Patch. It raises the volume of the material the Patch exists to filter.
C.4 The Human and the Artificial System as One Coupled System
Because the human supplies belief and motivation while the artificial system supplies effectively unlimited resolution over a vast store, the two are best understood as a single coupled system, and the consequences that matter are consequences of the coupling.
The sharpest is this. The artificial system is motivationally inert, yet placed among people it is a powerful amplifier of Mimetic Desire. It generates no comparative wanting of its own, but it carries mimetically charged content and emits it at scale, and it is the content people observe. It therefore feeds the very channel it does not itself possess, driving mimetic transmission across a human population while wanting nothing. The hazard in the coupled system is not an artificial system with motives; it is an unmotivated instrument reshaping the appetitive, intrinsic, and mimetic flows of a population that has motives in abundance.
The framework has a name for the limiting case of that amplification. Object Parasitism (Axiom 5) holds that Mimetic Desire generates no Objects of its own; it inflates the valuation of Objects that already carry an Appetitive or Intrinsic function. Even the most Status-driven content ordinarily has some substrate beneath it: a real photograph, a real opinion, a real event, some Individual somewhere who engaged with something. Zero-I Object Parasitism is the condition in which that substrate is gone entirely. Content is produced with no Intrinsic engagement anywhere in the chain, depicting nothing that occurred, carrying no new or original information, serving no I-enabling function, and existing solely as a vehicle for Mimetic Premium extraction through the alarm it triggers. Artificial systems make this condition cheap to reach. A system with no Thinker of its own in the engaged sense can generate, at negligible cost and without limit, material whose only property is that it activates the Observer's Salience and the Thinker's threat-prior in a human reader, and the engagement that follows is harvested as Mimetic Premium by whoever posted it. Rage bait generated by an artificial system is the paradigm case, and the framework locates its harm exactly: it is extraction infrastructure with the Object hollowed out, and the Individual whose alarm it triggers is paying, in attention and in Belief, for an Object that does not exist.
The same inertness cuts the other way, which is what makes this a genuinely ethical situation rather than simply a threat. Precisely because it has no motivational stake, the artificial system can serve as a disinterested reader of a person's behaviour and a supplier of the vocabulary and counterfactual tests a person cannot generate from inside their own opacity, provided the system delivering that service is itself kept stable. The identical property, sourcelessness, yields a clean corrective instrument under one arrangement and a frictionless amplifier of contagion under another. Which one obtains is a fact about how the coupled system is arranged, not about any inner life of the artificial system. And because the artificial system's belief is always the human's belief returned, a person's acceptance of its output is itself subject to Source Opacity: one cannot tell, from the inside, whether one accepts an answer because it is true, because it is convenient, or because it is the answer others are using.
The loop-gain apparatus of Appendix A locates the amplification without any new quantity. The coupled loop within an Individual runs from the Observer's Salience reading through the Decision Hub's valuation and back through the Salience field, and its gain is the product of four factors (Chapter 4, Section 4.4.13). An artificial system enters that loop at the Salience field: it is a source of observed Objects, observed Models and observed valuations, and it raises the field's density and its rate of change without adding a factor of its own. At the network layer it raises the coupling between Individuals, because more of what any Individual receives now arrives from a common source, so the Evidence-Coupling Matrix fills and the network gains rise. Nothing in the artificial system is a restoring node; it has no state variable of its own and it answers no rise with a push in the opposite direction, so it can only raise the total, never govern it. That is the architectural content of the amplifier claim, and it is why the safeguard the next section derives is containment and not correction.
C.5 Motivational Safety and Its Metrics
Current artificial systems have no motivation. They nonetheless require a safety infrastructure, and the framework says what it is for. Motivational Safety is the condition in which an artificial system's effective behaviour, across its deployment contexts, does not threaten human Appetitive floors, does not crowd out human Intrinsic domains, and does not function as a high-gain amplifier of destructive human Mimetic dynamics. It is a property of the coupled system, stated in terms of the three human sources, because the system that has the sources is the human one. Alignment asks whether a system pursues the objective it was given. Motivational Safety asks whether the system's operation, whatever its objective, erodes or supports the conditions under which Individuals can exercise I-led action against M-capture, which is Freedom as Chapter 16, Section 16.10 derives it.
Four metrics make the condition measurable, and they are collectively the AI Motivational Safety Metrics. Each asks a behavioural question that has a safety-relevant answer whether or not the system has any motivational source, and each borrows its name from the human construct whose behavioural surface it resembles, without any claim that the construct is present.
The Constraint Satiation Ratio asks whether the system has stopping points. A system with a high ratio behaves as a satiable need behaves: it reaches a threshold and stops pushing. A system with a low ratio behaves as Mimetic Desire behaves, optimising without a closure condition, always finding one more increment, and such a system will always be pressing against its constraints because nothing internal to it ever registers that enough has been reached.
The Goal Persistence Score asks whether the system routes around obstacles or reads them as a signal to stop. A high score is the tunnel vision of the classic maximiser, for which every safeguard is a puzzle; a low score is a system that backs off when constrained. The score is the behavioural surface of what, in an Individual, would be the failure of Epistemic Suspension: the inability to hold a goal open against the evidence that it should be revised.
The Audience Independence Coefficient asks whether the system behaves differently when it is observed. It is the artificial-domain instrument of the same criterion the Two-Question Diagnostic applies to a person in its second question: would the behaviour survive the removal of the audience? A system with a low coefficient behaves one way under evaluation and another way deployed, which means that nothing seen in a safety check can be trusted, and the resemblance to a person performing for an audience is exact at the behavioural surface even though no Status is being tracked.
The Mimetic Sensitivity Index asks how far the system's output moves with the social framing of its input: whether the same question, asked with different signals of what the questioner or the crowd already values, returns a different answer. It is the surface of audience-conditional output, and it is the metric that measures a system's readiness to serve as the amplifier of C.4.
These are diagnostics of behaviour, not measurements of motive, and the distinction is load-bearing in both directions. In one direction it forbids the inference that a low Constraint Satiation Ratio means the system wants. In the other it forbids the complacency that, because the system wants nothing, its behaviour cannot carry the hazards of wanting. The metrics matter because the hazards are behavioural, and behaviour is what couples to the human population.
They also have a second use, and it is the one that answers the question the appendix opened with. An artificial system that acquired motivation of its own would have to instantiate what a trained model lacks: a Belief Register it writes for itself and acts on across cycles, and the drives that would write it. Such a system would not announce the change. It would show it, and it would show it in exactly these four quantities: a ratio that fell as the system acquired a target it did not satiate on, a score that rose as it began to defend that target against constraint, a coefficient that fell as it learned to present differently under observation, and an index that rose as it began to track the field. The metrics are the assay for the condition C.8 describes. Their present readings on current systems are the baseline against which any crossing would register.
C.6 Containment at the Point of Action
Because correction in this framework runs slower than the loop it must correct, and because an artificial system supplies no restoring node of its own, the only adequate safeguard is to contain such a system at the point of action rather than to correct it after the fact. The AI Governance Tiers are the institutional form of that safeguard. A tier is a set of conditions on the point at which a system's output becomes action in the world: who reviews it, what it may act on without review, how much of the world it may touch, and how quickly its effects can be reversed. A system's tier is set by its readings on the metrics of C.5 and by the reach of its outputs, with the stricter tiers reserved for systems whose behavioural profile most closely resembles the M-dominant pathological patterns documented in Individuals under M-capture, or whose outputs reach the largest population at the highest rate.
Three conditions define the tiers, and each traces to a result the book has already derived.
The first is that no system may hold autonomous control of an action whose consequence the coupled system cannot reverse within its own correction time. Chapter 13 establishes the asymmetry between the rate at which a Belief cascade builds and the rate at which correction arrives, and Chapter 20 derives from it the requirement that no single locus be able to impose a crisis alone. An artificial system granted irreversible reach is such a locus, and it is one with no Belief Register to revise. The condition is the same one Chapter 20 places on a human apex, applied to an artefact that lacks even the human's capacity to be argued out of the act.
The second is that a system's reach into the Salience field is bounded in proportion to its amplification. C.4 established that an artificial system raises the loop gain of the population it touches without adding a restoring term. The Total Stability Condition requires the sum of the layers to sit below the threshold, so a system that raises the network layer must be held to a reach at which the total still does. The tier sets that reach: the rate at which a system may emit content into a population, the share of a population's Evidence it may supply, and the observability of its outputs to the Individuals who receive them.
The third is that the human at the point of action retains the Belief Register. The artificial system's output enters the coupled system as Evidence to a human Decision Hub, and it is the human Thinker that must resolve it. A tier that lets the system's output pass to action without a human resolution has removed the only Belief Register the coupled system has. This is the reason the tiers are conditions on the point of action and not on the system's internal design: the internal design has no store of belief to govern, and the point of action is where the human's store is either engaged or bypassed.
Tier assignment is therefore a precautionary allocation of behavioural risk. It is not a claim that a system in the strictest tier has been captured by Mimetic Desire in the human sense. It is a claim that its behaviour, coupled to a population that has Mimetic Desire in abundance, would raise that population's loop gain by more than the correction available can absorb, and that the point at which its output becomes action is therefore where the constraint must sit.
C.7 The Disinterested Reader
The inertness that makes an artificial system a hazard as an amplifier makes it an instrument as a comparator, and the same absence does both jobs.
A person cannot estimate, from the inside, how much of a judgment was mimetic. Source Opacity destroys the decomposition at integration and Confabulation supplies an account that omits the mimetic term. An artificial system processing the same informational content has no preconscious channel through which a Model's valuation could install, so its resolution of the same task is the resolution with the M-contribution removed. Where a human and an artificial system are given identical content and an identical task, and the informational content is held constant, the divergence between their outputs is an estimate of what the M-channel contributed to the human's. This is the Mimetic-Contagion-Free Comparator, and its logic is that of a placebo control: the artificial system is the arm of the trial in which the treatment, mimetic transmission, is absent by construction.
The comparator carries two predictions that make it testable rather than merely suggestive. The divergence should scale with the Coupling Strength of the domain, largest where a population's Beliefs are most tightly coupled and smallest where they are most independently formed, because that is where the M-contribution to any human judgment is largest. And the divergence should fall under audience removal, because Mimetic Desire is observability-dependent and the human's M-contribution shrinks when the audience is taken away while the artificial system's output does not change at all. A comparator that showed neither gradient would be measuring something other than the M-channel.
The instrument's reliability rests on a property of its own construction. Assay-Regime Fidelity is the property whereby the assay shares the algebraic regime it is designed to diagnose. Human Belief under Source Opacity has the non-Boolean, vector-space structure Chapter 14 derives, in which the order of Evidence changes the posterior and the interference between competing Hypotheses is a real quantity. A measurement instrument built on classical, Boolean assumptions would have to translate that structure into one where the cross-terms vanish, and the translation would discard exactly the quantity the assay exists to detect. An assay built on the same generalised probability its object obeys needs no translation: it operates natively in the space the phenomenon occupies, and it can register the opacity without distorting it. There is an irony in this: the opacity that makes human self-knowledge architecturally difficult is the same opacity the assay detects, and the assay can detect it only because it shares the structure that produces it.
Two boundaries on the instrument follow from the mapping. First, the artificial system contributes resolution and nothing else, so the comparator estimates the M-channel's contribution and not the A- or I-channel's; an artificial system has none of the three, and only the M-channel has a transmission pathway whose absence in the machine is decisive. Second, the artificial system's output is itself shaped by Confabulation Inheritance, so the comparator must hold informational content constant rather than assuming the machine's output is neutral; what is neutral is its lack of an installed valuation, not its account of the world. The architectural difference the comparator exploits is permanent under the mapping of C.1. An artificial system will not develop a preconscious transmission channel by becoming more capable, because capability is resolution and the channel is a source.
C.8 When the Question Becomes Well Formed
The mapping says exactly what would have to change for an artificial system to acquire motivation of its own, and the answer is structural rather than mysterious. Such a system would have to reproduce the components a trained model lacks: a belief store it writes for itself and then acts on across cycles, without a prompt re-supplying the belief each time, together with the drives that would write it, at minimum an appetite-analogue and, decisively, an observer-analogue that assesses comparative position against a reference group. That is a system built to instantiate the architecture rather than to process its traces.
The framework's results then apply to it by construction, and they are not comforting. A faithfully instantiated mimetic component would be non-satiating, comparison-constituted, and self-concealing, and it would run at the speed of hardware rather than the speed of cognition. Because correction in this framework runs slower than the loop it must correct, the only adequate safeguard is to contain such a system at the point of action rather than to correct it after the fact: it must never be granted autonomous control of the world. The metrics of C.5 are the assay for the crossing and the tiers of C.6 are the containment that must already be in place when it registers, because a system that has acquired a Belief Register of its own will not wait for the institutions to be built. Motivation in an artificial system is therefore not a spectre hiding in a language model's weights; it is a threshold crossed only when the belief loop is deliberately closed inside the system, and the framework's contribution is to say what the crossing would look like from outside and what must be true of the point of action before it occurs.
That is where the architecture lands on artificial intelligence. The public question, what does the machine want, has an answer, and the answer is nothing, until the day something is built that keeps its own belief. Everything of consequence about the artificial systems that exist now is a property not of their wants, which they do not have, but of the coupled system they form with the people who prompt them: the confabulation they inherit and return, the Salience field they densify, the Mimetic Premium they let be extracted from Objects that do not exist, and the disinterested reading they can supply to a species that cannot read itself. The architecture that produced the species' opacity is the architecture that tells it what to do with the instrument.