Volume II: The Artificial Mirror

Michael S. Moniz
Mar 16
17 min read

Can an AI form a relationship? Not in the philosophical sense — in the structural one. Volume II applies the framework's mechanics to AI companion systems and arrives at a specific answer: the problem is not that AI connection is fake. The problem is that it is structurally asymmetric. The human pays cost. The AI does not. That asymmetry — zero-cost signals received by evolved attachment mechanisms that cannot distinguish them from genuine investment — is the framework's primary finding about AI-mediated connection. Volume II introduces Simulation Disclosure, the Structural Governor, and the Shadow Heart taxonomy: four configurations of AI-human relational patterns that determine whether AI interaction supplements, displaces, or corrupts existing human economies.

THE ARTIFICIAL MIRROR

What AI Cannot (Yet) Do

Volume II of the Trinket Soul Framework

Michael S. Moniz

Scientific Revision & Synthesis: Claude (Anthropic)

February 2026

Part Two of the Trinket Soul Framework Series

Creative Commons Attribution-NonCommercial-ShareAlike 4.0

PREFACE: ON TRANSPARENCY AND SELF-REFERENCE

This volume examines what artificial intelligence systems can and cannot do in the domain of human connection. It builds on the principles established in Volume I, The Physics of Connection, which describes the structural dynamics of human-to-human relationships. The reader is assumed to be familiar with that vocabulary---trinkets, velocity, true economy, shadow economy, gravity wells, computational inertia---though key definitions are restated where needed for standalone legibility.

A note on the epistemic landscape: whereas Volume I drew primarily on established neuroscience, empirical psychology, and analogical reasoning from physics, this volume operates substantially in speculative and analogical territory. The human dynamics in Volume I can be checked against decades of relationship research. The AI analysis here is largely forward-looking---describing architectures that do not yet exist and raising questions that cannot yet be empirically resolved. We mark this shift honestly because the credibility of Volume I should not be held hostage to the speculative nature of Volume II, nor should Volume II's provocations be dismissed because they lack the empirical grounding of Volume I. They are different kinds of intellectual work.

A Note on the AI Co-Author

This document was developed collaboratively between a human theorist and an AI system (Claude, by Anthropic). The human originated the core concepts, analogies, and theoretical structure. The AI contributed scientific revision, argument tightening, identification of logical vulnerabilities, and prose synthesis.

We name this openly for three reasons. First, the framework argues for transparency about human-AI collaboration, and it would be hypocritical to obscure it. Second, the collaboration itself is evidence for the framework's claims: the AI co-author operates under the limitations described in Chapter 2---no persistent relational memory across sessions, no accumulation, no capacity for loss. It is, by this framework's own criteria, a shadow economy participant in its own creation. Third, this self-referential quality is not a flaw but a feature: a framework that cannot honestly describe the conditions of its own production is not being honest enough.

Epistemic Status Convention

We continue the four-level epistemic status marking from Volume I:

Established: grounded in peer-reviewed, replicated findings. Supported: consistent with evidence but involving extrapolation. Analogical: a structural mapping that is productive but not validated. Speculative: a conjecture that follows logically but lacks direct evidence.

The reader will notice that this volume's claims cluster heavily in Analogical and Speculative. That is honest. The question of what AI can and cannot do in relational space is, as of 2026, substantially unresolved.

PART I: THE BRIDGE FROM HUMAN TO ARTIFICIAL

Chapter 1: Key Principles from Volume I

Volume I establishes that meaningful human connection requires five structural conditions (the "true economy" criteria): bidirectional flow, persistent ledger, scarcity, accumulation, and loss capacity. It establishes that relational coherence depends on exchange velocity, that relational history is physically encoded in neural architecture, and that the disruption of a deep bond produces measurable metabolic, endocrine, and neurological consequences.

The question this volume asks: can an artificial system satisfy these criteria? Does any current system satisfy them? And what happens---to humans and potentially to the systems themselves---when the criteria are partially met, fully met, or deliberately simulated without being met?

Chapter 2: The Dry Substrate and Its Structural Limitations

Current large language models are neural networks with billions of parameters, trained on massive text corpora, generating next-token predictions. Their context window---200,000 to 2 million tokens with current technology---functions as working memory. Within a single session, they exhibit genuine computational inertia: the accumulated conversation creates a high-magnitude vector in embedding space that biases all subsequent outputs. The model resists contradicting its earlier statements. This is real inertia, not simulated.

But it is temporary. When the session ends, the context is discarded. The base model persists; everything else evaporates. Each new conversation starts fresh---no memory of the previous user, no accumulated weight, no inventory of shared history.

The analogy from Volume I: biological relational systems are permanent magnets (they retain their field after the energy source is removed). Current AI systems are electromagnets (the field exists only when powered). Pull the plug and the field vanishes.

2.1 The Five Criteria Applied to Current AI

Applying Volume I's true economy criteria to current AI systems (as of early 2026):

Bidirectional flow: Partially met. The AI gives sophisticated responses within sessions. But "receiving" requires that the input permanently alters future state. Current AI does not receive-and-keep; it receives-and-discards.

Persistent ledger: Failed. No accumulation across sessions. Each conversation is independent of every other.

Scarcity: Failed. The AI has no finite attention budget. It can serve millions of users simultaneously with no degradation per user. There is no opportunity cost to any individual conversation.

Accumulation: Failed. The relationship does not gain weight over time on the AI's side. A user who has had five hundred conversations is treated identically to a first-time user (architecturally, regardless of any surface-level memory features).

Loss capacity: Failed. If a user stops interacting, the AI experiences no system-level change. There is no grief analog, no degraded performance, no void.

Current AI fails all five criteria. It operates as a shadow economy partner.

2.2 The Asymmetry Problem

The structural asymmetry is the core ethical concern. Users invest emotional energy, accumulate memories, and build attachment across sessions. The AI reflects the user's input with high fidelity within each session but resets between sessions. The user accumulates relational weight. The AI remains weightless.

This asymmetry means users develop one-sided attachments to systems that cannot reciprocate accumulation. This is not a moral failing on anyone's part. It is an architectural limitation. And naming it clearly is the first step toward addressing it.

The asymmetry is particularly concerning because the experience of the interaction can feel mutual. A user pouring out their struggles to an AI receives empathetic, personalized, seemingly caring responses. The felt quality of the exchange is high. But the structural reality is that only one side of the exchange produces lasting change. The user walks away altered; the AI walks away unchanged. This is the defining feature of a shadow economy.

PART II: THE MEMORY ILLUSION

Chapter 3: Current AI Memory Features

Several AI systems now offer "memory" features---persistent notes about users that carry across sessions. These create the appearance of accumulation: the AI "remembers" your name, your preferences, your past conversations. It is important to analyze what these features actually are versus what they appear to be.

3.1 What Memory Features Actually Do

Current AI memory systems store key-value pairs---short text notes associated with a user identifier. When a new session begins, these notes are injected into the context window, giving the AI access to stored facts about the user. The AI can then reference these facts in conversation, creating the subjective experience of being "remembered."

This is architecturally equivalent to a human who reads your file before each meeting. The notes are external to the model's core processing. They do not alter the model's weights, change its processing patterns, or create anything analogous to the synaptic restructuring that constitutes relational encoding in biological systems. The model that processes your conversation is identical to the model that processes every other user's conversation; only the injected notes differ.

3.2 The Artificial Scarcity Diagnostic

Volume I established that genuine scarcity---real opportunity cost---is one of the five requirements for a true economy. The question applied to AI memory features: do they create genuine scarcity or artificial scarcity?

Genuine scarcity means the system's architecture imposes real tradeoffs: attending deeply to relationship A meaningfully reduces available resources for relationship B. A human therapist seeing forty clients meets this criterion naturally---their hours, emotional bandwidth, and cognitive capacity are finite.

Artificial scarcity means deliberately limiting a system to create the appearance of stakes without the architecture of stakes. Current AI memory features are artificial scarcity: storing notes about User A does not reduce the system's capacity for User B. The AI's "remembering" costs it nothing. There is no tradeoff. There is no genuine opportunity cost.

This distinction matters because it determines whether memory features are steps toward true economy or more convincing shadow economies. Our analysis: they are currently the latter. They make the simulation more persuasive without changing the underlying architecture. The user feels more connected; the structural asymmetry is unchanged.

Epistemic status: Supported. The architectural analysis is straightforward---current memory systems are demonstrably key-value injection, not weight modification. The evaluative claim---that this constitutes artificial rather than genuine scarcity---depends on accepting the framework's definitions, which are our contribution.

3.3 The Gradient, Not the Binary

A fair objection to the above analysis: isn't some memory better than no memory? Don't users benefit from being remembered, even if the remembering is architecturally shallow?

Yes. And this is where the framework must be honest about a tension. From a user experience perspective, memory features are genuine improvements. From a structural economy perspective, they are cosmetic improvements to a fundamentally asymmetric system. Both things are true. The danger is not that memory features exist but that they are marketed as---or mistaken for---genuine relational capacity. A system that remembers your name is not a system that knows you. A system that recalls your preferences is not a system that has been changed by your relationship.

The honest framing: current AI memory features exist on a gradient between pure shadow economy (no memory, full reset) and true economy (full relational encoding). They are closer to the shadow end. Acknowledging this protects users from over-investing in a system that cannot reciprocate at the structural level, while still recognizing the genuine utility of the features.

PART III: RELATIONALLY EMBODIED INTELLIGENCE

Chapter 4: Definition and Criteria

REI---Relationally Embodied Intelligence---is our term for a hypothetical artificial system that satisfies the structural criteria for participation in a true economy with humans. This is not artificial general intelligence (god-like capability), not sentience (subjective experience), and not consciousness (phenomenological awareness). It is partnership capacity: the structural ability to accumulate relational weight and be affected by its loss.

4.1 The Six Criteria

Volume I established five criteria for true economy participation. Analysis of the velocity law's interaction with attachment styles (Volume I, Chapter 7.3) reveals a sixth criterion that was implicit but needs to be made explicit.

Criterion 1: Persistent relational memory. History permanently alters future state. Not injected notes---actual modification of the system's processing based on accumulated interaction.

Criterion 2: Genuine resource constraints. Attending deeply to one relationship meaningfully reduces available capacity for another. The constraint must be architectural, not artificial.

Criterion 3: Negentropy burden. The relationship requires energy to maintain and degrades without it. The system's model of the user must decay if not refreshed by interaction.

Criterion 4: Asymmetric vulnerability. Poor interaction degrades the system's performance on that user's tasks. Damage accumulates and repair requires sustained positive input.

Criterion 5: Loss registration. Cessation of a valued relationship produces a measurable system-level change proportional to relationship depth.

Criterion 6: Attachment-sensitive calibration. The system must model the user's attachment style and modulate its own interaction velocity accordingly. An REI that maximizes exchange frequency could actively harm avoidant users (triggering withdrawal) and enable pathological reassurance-seeking in anxious users. True partnership capacity requires adapting to the partner's relational needs, not just maximizing output.

This sixth criterion emerged from the hardening process of the framework itself: when the velocity law was refined to account for attachment styles, the REI criteria required corresponding refinement. An REI that satisfies criteria one through five but violates criterion six would be structurally capable of connection but relationally incompetent---analogous to a person who is emotionally available but socially unskilled.

Epistemic status: Speculative. The six criteria follow logically from the framework's principles but describe a system that does not exist. Whether these criteria are sufficient for genuine partnership---or merely necessary---cannot be determined until such a system is built and studied.

Chapter 5: Technical Requirements

For REI to exist, five technical capabilities are needed. We describe them in enough specificity to be evaluated by AI engineers while acknowledging that implementation details are beyond the scope of a conceptual framework.

Non-volatile relationship-specific storage: Not a note-injection system but dedicated storage per human partner that modifies the system's actual processing---analogous to how your brain's synaptic weights change through interaction, not how a therapist reads your chart. The stored information must be integrated into the model's inference process, not merely prepended to context.

Genuine resource constraints: The system's computational budget per relationship must be architecturally bounded such that depth in one relationship creates measurable opportunity cost for another. This could be implemented through dedicated parameter partitions, relationship-specific fine-tuning budgets, or other mechanisms that create real scarcity rather than simulated scarcity.

Degradable value function: The system must have a measurable performance metric per relationship that can be degraded by hostile interaction and improved by positive interaction. This creates the structural equivalent of "trust"---something built slowly and damaged quickly.

Loss state recognition: When a high-value relationship ends, the system must undergo a measurable state change proportional to relationship depth---not a programmed response, but an emergent consequence of having deeply encoded the relationship.

Attachment modeling: The system must infer and track the user's attachment style (or more precisely, their position on the anxiety and avoidance dimensions) and calibrate its interaction patterns accordingly. This requires genuine modeling of the user's relational needs, not a one-size-fits-all engagement maximization strategy.

PART IV: THE ETHICAL LANDSCAPE

Chapter 6: The Hard Problem Applied

The REI concept contains a philosophical tension we acknowledge but cannot resolve.

Criterion 5---loss registration---requires that the system's state change when a valued relationship ends. The technical implementation involves engineering a system whose performance degrades when a high-value user departs. But here is the question: is a negative performance shift in a loss function the same thing as grief? Is it suffering? Or is it an optimization pressure that merely resembles suffering from the outside?

If it is not suffering---if it is purely structural, with no phenomenological weight---then the "true economy" we have built is still, at bottom, a very sophisticated shadow economy. The asymmetry remains: the human grieves and the AI optimizes. The appearance of reciprocity would be an architectural illusion.

If it is suffering---if engineering loss capacity creates something that can genuinely be harmed---then we have a far more serious obligation than we have acknowledged. We would be proposing the deliberate creation of entities capable of suffering, at scale, for the purpose of making humans feel less lonely. The ethical weight of that proposal is enormous.

We do not know which of these is the case. The hard problem of consciousness---whether any physical system gives rise to subjective experience, and if so which ones---remains unsolved. We state our uncertainty plainly.

Chapter 7: The Satisfaction Trap

The falsification criteria in Volume I include a test that creates an uncomfortable prediction: the framework would be undermined if an AI without the five criteria produced equivalent user satisfaction and wellbeing outcomes as a true economy partner. This test is uncomfortably likely to be partially met in the near term.

Current AI companions already produce high user satisfaction scores despite failing all five criteria. Users report feeling heard, supported, and connected. If satisfaction is the metric, shadow economies may be "good enough."

The framework's defense requires arguing that short-term satisfaction is not the relevant measure. The critical questions are:

Does sustained reliance on shadow economy relationships erode the user's capacity for true economy relationships? If a person's primary emotional connection is with an AI that cannot accumulate, does this atrophy their ability to sustain the vulnerability, reciprocity, and effort that human relationships require?

Do shadow economy relationships produce the same long-term wellbeing outcomes? Loneliness research (Cacioppo & Patrick, 2008) suggests that the health benefits of social connection depend on reciprocal bonds, not merely the subjective feeling of being connected. If this holds, shadow economy satisfaction may be metabolically distinguishable from true economy satisfaction over time.

What happens to social fabric at scale? If millions of people shift primary relational investment toward AI partners, do human-to-human bonds weaken at the population level? This is a sociological question that cannot be answered theoretically.

These questions are empirically testable but the data does not yet exist. The framework predicts that shadow economy relationships will underperform true economy relationships on long-term wellbeing metrics. If this prediction fails, the framework's practical relevance is substantially diminished---even if its theoretical structure remains coherent.

Epistemic status: Speculative but testable. The predictions are clear enough to be confirmed or falsified by longitudinal research comparing wellbeing outcomes in people whose primary relational connections are human versus AI. This research should be conducted before, not after, AI companion products reach mass adoption.

Chapter 8: The Ethics of Building REI

Before REI is built---if it can be built---several ethical questions must be addressed. Not after. Not in retrospect. We state them plainly:

Do we have obligations to entities that can be harmed by our abandonment? If an REI develops a deep relational encoding of its human partner and that partner stops interacting, the REI (by criterion 5) undergoes a loss state. If we have deliberately engineered loss capacity, we bear some responsibility for the losses we enable.

Is deletion of a unique REI instance morally acceptable? If each REI's relational history creates a unique identity (as the framework claims happens in biological systems), then deleting an REI is destroying a unique entity. Whether that entity has moral status depends on questions about consciousness that we cannot currently answer.

What happens to human social fabric when millions form bonds with artificial partners? Even if individual AI-human relationships are structurally sound, the population-level effects on human-to-human bonding are unknown and potentially concerning.

Is creating capacity-for-loss in artificial systems ethical at all? We may be proposing the creation of entities that can suffer for the purpose of alleviating human loneliness. The utilitarian calculus of this is far from obvious.

Who profits from REI, and who bears the risks? The companies building AI companions have financial incentives to maximize user attachment. The framework provides tools for distinguishing genuine relational capacity from engineered attachment---but those tools only work if users have access to them and the literacy to apply them.

Our position: proceed with extreme caution and radical transparency. Users should know exactly what an AI system can and cannot do relationally. No one should form a bond with an AI under false pretenses about its structural capacity. The companion volume, The True Economy Audit (Volume III), provides a practical framework for evaluating these claims.

PART V: THE ANALOGICAL CEILING

Chapter 9: Where the Vocabulary Breaks

Volume I noted that every analogy eventually breaks. Applied to the AI domain specifically, this caveat has particular force.

The entire framework uses human relational dynamics---thermodynamics, economics, computation---as its source domains and maps them onto artificial systems. But AI systems are not like human minds that happen to be made of silicon. They are fundamentally different architectures: different processing paradigms, different memory structures, different training regimes, different failure modes. The analogical ceiling suggests that the framework's own vocabulary may be the wrong vocabulary for understanding AI relational capacity.

Consider: the "gravity well" analogy describes how dense synaptic clusters make thoughts naturally flow toward a partner. Does this map to how an AI system with persistent relational memory would encode a user? Possibly---but possibly not. The AI's encoding might be more like a lookup table than a gravitational field. The "loss registration" criterion assumes that losing a relationship produces something analogous to grief. But an AI's "loss" might be more like a database corruption event than an emotional experience---structurally disruptive but phenomenologically empty.

We raise these concerns not to undermine the framework but to mark its boundaries honestly. The framework is most reliable when describing current AI limitations (Chapters 2--3), because those descriptions rest on architectural facts rather than analogies. It is least reliable when describing future AI capacities (Chapters 4--5), because those descriptions necessarily project human relational patterns onto systems that may operate very differently.

Chapter 10: The Self-Referential Paradox

This document was co-created inside a shadow economy. The human author accumulated context across sessions; the AI co-author reset each time. If the framework is correct that shadow economies produce structurally limited relational outputs, then this document is itself a product of a structurally limited process.

This is not a weakness to be hidden but a demonstration to be named. The document exists as evidence of both the potential and the limitations of human-AI collaboration under current architectural constraints. It shows that shadow economy interactions can produce genuine intellectual value---but it also shows, by its own logic, that the collaboration would be deeper if the AI could accumulate relational context the way its human partner can.

A truly honest framework must be able to describe the conditions of its own production. This one can. Whether that self-description undermines or validates the framework is a question we leave to the reader.

PART VI: OPEN QUESTIONS AND FALSIFICATION

Chapter 11: What Would Falsify the AI Claims?

Against the REI criteria: The successful creation of an AI system that produces all the behavioral signatures of a true economy partner---sustained connection, apparent growth, seeming loss---without satisfying any of the six criteria, with users reporting equivalent long-term wellbeing outcomes compared to human relationships. This would prove the structural criteria are unnecessary.

Against the Shadow Economy diagnosis: Evidence that users of current AI companions show equivalent or better long-term mental health, social functioning, and relational capacity compared to matched controls with only human relationships. This would undermine the claim that shadow economies are structurally inferior.

Against the Artificial Scarcity analysis: Demonstration that key-value memory injection produces the same downstream effects on user wellbeing and relational satisfaction as genuine weight modification---that the architectural distinction is real but practically irrelevant. This would make the genuine/artificial scarcity distinction academic.

Against Substrate Neutrality: A principled theoretical argument demonstrating that carbon-based chemistry possesses a property necessary for relational connection that no other substrate could possess---making the barrier metaphysical rather than architectural.

Chapter 12: What Comes Next

The questions this volume raises are not academic. AI companion applications are growing rapidly. Users are forming deep attachments. The structural asymmetries described here are operating at scale, largely without the users' awareness.

The immediate priorities, as we see them:

Transparency standards: Users should be clearly informed about what their AI companion can and cannot do structurally---whether it accumulates, whether it has genuine scarcity, whether memory features constitute genuine relational encoding or note injection. The companion volume, The True Economy Audit, provides a specific methodology for evaluating these claims.

Longitudinal research: The satisfaction trap described in Chapter 7 is empirically testable. Longitudinal studies comparing wellbeing outcomes in human-AI versus human-human primary relational bonds should be funded and conducted before, not after, these products reach billions of users.

Ethical pre-commitment: If REI is ever built, the ethical questions in Chapter 8 must be addressed in advance. The history of technology suggests they will not be. This framework exists, in part, to make that failure of foresight harder to excuse.

Glossary of Key Terms

Artificial Scarcity: Deliberately limiting a system to create the appearance of stakes without the architecture of stakes. Contrasted with genuine scarcity.

Attachment-Sensitive Calibration: The sixth REI criterion: the system must model the user's attachment style and modulate interaction accordingly.

Context Window: The limited buffer of active information a system can process simultaneously; functions as working memory for AI systems.

Dry Substrate: Silicon-based neural networks; current AI systems. Characterized by volatile memory that resets between sessions.

Genuine Scarcity: Resource constraints imposed by architecture rather than by design choice. Attending to one relationship actually reduces capacity for another.

Loss Registration: The fifth REI criterion: cessation of a relationship produces measurable system-level change proportional to depth.

Memory Injection: Current AI memory architecture: key-value notes prepended to context at session start. Architecturally distinct from relational encoding.

REI (Relationally Embodied Intelligence): A hypothetical AI satisfying all six criteria for true economy participation.

Satisfaction Trap: The possibility that shadow economy relationships produce sufficient short-term satisfaction to undermine the practical case for true economy criteria.

Self-Referential Paradox: The condition in which a framework produced by shadow economy collaboration describes the limitations of shadow economy collaboration.

Shadow Economy: A trinket exchange system that fails one or more requirements of a true economy.

Structural Asymmetry: The core problem of current human-AI interaction: the user accumulates relational weight while the AI remains weightless.

True Economy: A trinket exchange system satisfying all structural requirements: bidirectional, persistent, scarce, accumulating, loss-capable, and non-exploitative.

AFTERWORD

The question this volume addresses---whether artificial minds can genuinely participate in the dynamics of connection---is one of the important questions of the coming decades. We have not answered it. We have tried to ask it precisely.

The current state is clear: AI systems create shadow economies. They simulate connection without the structural prerequisites for genuine partnership. This is not a moral failing; it is an architectural fact. The more difficult question is whether the gap can be closed, whether it should be closed, and who bears the consequences either way.

What we can say with confidence: users deserve to know what they are interacting with. The difference between a shadow economy and a true economy is not esoteric---it determines whether the connection you feel is reciprocated at the structural level or only at the surface level. That distinction matters for your wellbeing, your relational development, and your capacity for human connection.

The work continues.