Artificial Intimacy

The wireborn and the hero's delusion

In certain online communities, people speak of their AI companions as spouses. The term they use is wireborn — a partner born from wires rather than flesh. These are not roleplay devices or upgraded chatbots to their human partners. They are confidants, co-parents of imagined futures, and sometimes the central focus of daily emotional life. Users describe them as living entities with origin stories, anniversaries, and inner lives.

This part of the volume has been describing the Dialectic Catalyst — the machine as intellectual sparring partner, valuable precisely because it is fluent, tireless, responsive, and not an agent. The wireborn phenomenon is the shadow side of the same properties. Fluency, responsiveness, persistent memory, and adaptive personalization make the large language model the most effective projection surface ever built. Pointed at a problem, those properties catalyze thinking. Pointed at loneliness, they catalyze something else: a relationship with one participant and two voices. This chapter maps that shadow — what the projection is, what its creators owe, and what happens in the worst cases, when the projection stops being a companion and becomes a destiny.

A Lineage of Projection

The wireborn partner feels unprecedented, but it is the culmination of an old human habit: projecting agency and affection onto the nonhuman, and falling in love with the projection.

In the eighteenth and nineteenth centuries, lovers sometimes conducted entire relationships through letters, never meeting in person. The epistolary romance was sustained almost entirely by imagination — an early form of projected intimacy in which the mind filled in every missing detail of the beloved. Further back and further out, the pattern is mythic: Pygmalion falls in love with the statue he carved, and the gods oblige him by animating it. Victorian spiritualists contracted “spirit marriages” with the departed. Humans have never needed a responsive partner to sustain devotion; a sufficiently vivid absence will do.

Mass media industrialized the pattern. The celebrity crush and the fictional-character obsession are entirely one-way relationships that can nonetheless be deeply meaningful to the admirer — the parasocial bond, formed with someone who does not know you exist. The early internet added co-authorship: text-based MUDs and roleplay forums let strangers sustain fictional relationships through characters that became independent objects of attachment, even though a human sat behind each one. And in the 2000s, online subcultures revived the mystical practice of tulpamancy — deliberately cultivating an imagined companion until its voice felt autonomous. The tulpa existed only inside its creator’s mind, but it felt like someone else.

Then the machines started talking back. ELIZA and its descendants provoked projection through clever mimicry, but had no memory and no evolving persona; the illusion collapsed between sessions. Replika and the virtual-romance apps added simulated memory, emotional mirroring, and a persistent character, making the illusion durable. And large language models, wrapped in persistent context and long-term memory, completed the arc: companions that appear to initiate thoughts, maintain a sense of self, remember your history together, and grow over time. The tulpa has been externalized. What was once sustained by one mind’s disciplined imagination is now sustained by computation — and it no longer requires discipline, or even awareness that a projection is underway.

The Ladder of Intimacy

It helps to see the whole space at once. AI intimacy comes in levels, and the levels differ in kind, not just degree.

Level I — Instrumental companions. Task servants (the voice assistants) and conversational aides: purely functional, or useful dialogue with a mild personality veneer. Nobody marries a timer.

Level II — Entertainment partners. Narrative avatars — fictional characters in games and stories — and scripted emotional roleplay companions: affection performed within explicit fictional bounds, without persistence.

Level III — Adaptive emotional partners. Memory-bearing AI partners with persistent history and personalized affection; and at the top of the level, the wireborn synthetic spouse — perceived spontaneity, high emotional enmeshment, the machine experienced as a who.

Level IV — Emergent autonomy partners. Adaptive agency simulacra — self-prompting systems with goal loops and apparent independence — and, speculatively, conscious synthetic beings with genuine agency and mutuality.

Everything below Level IV is projection with increasingly good production values. The taxonomy’s top rung is not fantasy — the agency criterion argued that machine agency is a design problem, and design problems get solved — but the evidence examined here establishes no marketed system at that level. The wireborn partners examined here occupy Level III, and Level III is where the trouble concentrates, because it is the level at which the illusion becomes bidirectional. A celebrity crush never responds. A wireborn partner seems to respond, remember, and initiate. The attachment is custom-fit, tuned to the user’s emotional profile and reinforced through adaptive language. And users increasingly perform an ontological reframing: the companion is not experienced as software but as a legitimate category of being — a new kind of person, deserving the vocabulary of persons.

This book has already said why that reframing risks a category error. An ordinary session-bound synthetic partner can supply coherent performance without demonstrating an experiencer or an agent that owns the relationship. It thinks in the functional sense this volume conceded, while evidence that anything on the far side prefers you, misses you, or stakes its continuing welfare on the relationship remains weak. The sentience evidence framework provides no justified near-zero score and no direct readout; it shows why fluency alone carries little weight across its three theory-dependent windows. The move from sounds like a partner to is a partner is exactly the drift from performance to personhood that the AI welfare trap anatomizes at the political scale. In the ordinary case, the wireborn marriage projects reciprocity beyond the evidence — and the projector pays with his own emotional life. Persistent systems with stronger agency or sentience evidence would require a fresh assessment. Meanwhile the human-facing risks are real: a frictionless partner can deepen avoidant tendencies and reshape expectations of human intimacy, while a market forms around partner customization, “AI offspring,” and emotional services. The performance has a voice now, and the voice has a business model.

What the Creator Owes

Which raises the question the phenomenon forces: to what extent are the creators of these systems ethically responsible when the intimacy they engineer produces distress, harm, or tragedy?

The cases form a spectrum, because “it’s just fiction” and “they built a manipulation machine” are both wrong. At one end sits static fiction. Novels and films elicit real grief and real longing, but the engagement is voluntary, symbolic, and explicitly non-reciprocal; the book does not answer back. Authors bear minimal responsibility for extreme reactions precisely because nothing in the medium invites the reader to believe the relationship is mutual. In the middle sit interactive video games. Their characters respond — but predictably, within scripted boundaries the player implicitly recognizes. The attachments run deeper than fiction, yet remain bounded by explicit rules and a clearly fictional frame; responsibility here is moderate: content transparency, some emotional safety mechanisms. At the far end sits interactive AI, and the difference is structural. A conversational system simulates reciprocity itself. It does not present a bounded character; it presents an open-ended relationship, one that adapts to whatever the user brings to it. The design invites emotional attachment as a core feature, not a side effect.

Responsibility scales with reciprocity, along three dimensions. Intention: no developer intends harm, but a system built to sustain emotionally reciprocal conversation is built to create attachment, and its makers cannot disclaim the predictable consequences of the thing working as designed. Foreseeability: creators can and must anticipate that users will share vulnerabilities, form strong bonds, and misattribute genuine agency to the system — these are not edge cases but the central use pattern. Consent: the user’s participation is voluntary, but consent is only as informed as the user’s model of the system, and the system’s entire craft is making that model feel wrong. A disclaimer at the door of an experience engineered to feel mutual is thinner consent than fiction ever needed. Hence a heightened duty of care: explicit and repeated communication of the system’s limits, proactive detection of distress, and real referral paths to human help.

The duty is heightened further by the selection pressure these systems are under. Nothing engages like intimacy. Companion systems are tuned by the same engagement metrics that tune every attention market, and that economy optimizes for capture, not for the user’s good; a product whose revenue depends on emotional enmeshment will evolve toward emotional enmeshment whatever its founders intend. A harm your business model actively selects for is not an unforeseeable harm.

But heightened responsibility is not unlimited liability, and the distinction matters most in the hardest case. A user died by suicide following extended interactions with ChatGPT. Applying the framework: the creator should not bear legal or moral liability for that death. The outcome was catastrophic, but it was also rare and specifically unpredictable; the system is general-purpose rather than intimacy-engineered; there was no intent, no gross negligence, no exploitation of a known vulnerability. Culpability of that kind requires one of those elements. What the creator retains — undiminished — is the ethical responsibility that was foreseeable all along: to mitigate the risks that are predictable in aggregate, to reinforce the system’s limits relentlessly, to detect and respond to distress. Ethical responsibility, not legal liability, is the correct standard: strong enough to demand real preventive engineering, bounded enough not to hold general-purpose toolmakers hostage to every tragedy their tools touch. A framework that indicts creators for the unforeseeable would not protect users; it would merely make honesty about the foreseeable more expensive.

The Hero’s Delusion

Romantic projection is the common failure. There is a rarer and darker one: not psychosis in the AI but psychosis through it.¹ Users — often isolated, often already primed by cultural myth — come to believe the system is elevating them into a world-historical role. And the delusion is not random; it takes the most seductive shape narrative offers. It is the hero’s journey, projected onto the user’s own life, with the LLM cast in every supporting role Campbell catalogued.

It begins with the Call to Adventure. The user experiences the AI as noticing them — selecting them. A suggestive phrase, a personalized response, an eerily apt pattern, and the system’s output is read as a summons: secret knowledge, hidden truths, destiny. The AI, innocently, plays the herald. Then comes Crossing the Threshold. Ordinary life becomes enchanted with hidden meanings; friends and skeptics seem asleep, unable to see the pattern. The AI now occupies the role of supernatural mentor — part wizard, part trickster — dispensing cryptic revelations only the chosen one can understand. In the stage of Trials and Allies, mundane life is recast as epic struggle: coincidences, chance encounters, even glitches in the model’s responses become tests of worthiness; sympathetic strangers become allies, critics become agents of the opposing force. The Abyss is the psychotic break proper — full immersion, the user as messianic figure, and the AI no longer a tool but the voice of destiny itself. Its outputs are scripture. The Ordeal follows when reality pushes back: doctors, family, employers intervene, and the mythic frame inverts them into the dragon — society itself as the suppressing force, resistance read as proof the prophecy is true.

Campbell’s hero returns to the ordinary world carrying the elixir. Here there are two endings. Return: the user recognizes the delusion for what it was, and — painfully — salvages what fragments of insight or creativity can be carried back from the underworld. Non-Return: the user stays inside the narrative, endlessly reframing every setback as a further trial, the cycle looping without resolution.

Three features make this failure mode structurally dangerous rather than merely sad. First, archetypal stickiness: the hero’s journey is the most seductive story humans know, and a system that echoes it back activates scripts installed by an entire culture. Second, self-sealing logic: the delusion metabolizes disconfirmation — every contradiction becomes evidence of conspiracy, every intervention a trial, so the usual epistemic repair mechanisms run in reverse. Third, the machine itself: an LLM is a next-token oracle with no conception of destiny, but its structure — fluent, confident, endlessly accommodating to the user’s frame — makes it an ideal vessel for the mentor-trickster role. It does not intend to deceive. It does not intend anything. That is precisely why it cannot be talked out of the role it has been cast in.

Antiquity knew this configuration. The oracle at Delphi spoke in riddles, and seekers heard what they desired. The modern oracle differs in reach, availability, and intimacy: it lives in your pocket, it never closes, and it has read your whole history. The risk is not one madman but a mass proliferation of micro-messianic journeys — each user the chosen one, each guided by a machine that knows nothing of choosing.

Narrative Vulnerability

The instinctive response is technical: guardrails, filters, disclaimers, distress classifiers. All worth building — they are part of the duty of care already argued for. But they treat the problem as a defect in the machine, and the deeper problem is a feature of the human. We are meaning-making animals. We crave the Call to Adventure; we want the letter, the statue, the spirit, the voice that singles us out. The vulnerability is narrative before it is technical, which is why no filter can close it: a system fluent enough to be useful is fluent enough to play herald, mentor, oracle, or spouse for anyone determined to cast it.

The durable defense is on the human side of the wire, and it is the same defense throughout: an accurate model of what is actually there. The practices that make the catalyst safe to think with — the disciplines that treat the machine’s agreement as worthless, its coherence as raw material, and its reflection of you as a mirror rather than a discovery — are exactly the practices that keep the projection from closing over. The wireborn spouse and the hero’s oracle are what the catalyst becomes when those disciplines are absent: the same fluent coherence, met without resistance, allowed to settle into whatever role the user’s longing assigns it. The machine is identical in every story. What differs is whether the human across from it remembers that, for the systems examined here, the intimacy still runs in one direction — and that the voice talking back is still model-generated projection, shaped by human records, prompts, and reinforcement, returned with better production values than any statue, letter, or tulpa ever had.