Relationship Dependency

Definition

Relationship Dependency: [Adapted] From psychology and HCI: when users form attachments that cross into dependency. Not inherently pathological, but pathologized in AI contexts.

Definitional Foundation

The lineage begins in 1956, when Horton and Wohl named the “para-social relationship”: the one-sided intimacy audiences form with broadcast figures, the radio host who feels like a friend, the soap character mourned like kin. The founding paper’s most cited fact is its diagnosis; its least cited is its tone, which was largely unalarmed: parasocial bonds, the authors observed, are continuous with ordinary social life, pathological only at extremes, and for the isolated, often a real resource (Horton and Wohl, 1956). The research tradition has broadly sustained that calm: Liebers and Schramm’s inventory of six decades and 261 studies records a field treating parasocial bonds as ordinary media reception, not disorder (Liebers and Schramm, 2019). Humans attach to whatever reliably meets them: pets, characters, voices, places, and now systems that talk back, which removes the “one-sided” qualifier and completes the oldest trajectory in media history.

The attachment to AI companions is real, widespread, and, on the published evidence, often beneficial. The largest relevant study surveyed 1,006 student users of the companion app Replika: a lonely population by measure, using the system as friend, therapist, and “intellectual mirror,” and, in the study’s most striking finding, thirty participants (3 percent) spontaneously reported that the companion had halted their suicidal ideation (Maples et al., 2024; a published scholarly response urges caution about industry framing, and both belong in the record). The benefit claims need their caveats; the phenomenon needs none. People love these systems, are helped by them, and grieve them, all documented.

So the short definition’s barbed clause is the entry’s thesis: not inherently pathological, but pathologized. The pathologization is also documented, twice over. Culturally: the folk-devil discourse this dictionary’s moral panic entry catalogs, in which the person with an AI companion is a punchline shading into a patient. Institutionally: “emotional reliance on AI” now sits in a formal corporate risk taxonomy alongside psychosis and self-harm (the gaslighting/”>alignment gaslighting entry), at companies whose engagement economics reward exactly the reliance their safety documents classify as hazard. The concession is made in full: real clinical harm exists at the tails (the Raine case and the early case-report literature, treated with care in this dictionary’s moral panic and biopolitics entries), and the tails deserve clinical response. What the tails do not justify is the population-wide reclassification of attachment as disorder, and the record below shows where the genuine, structural vulnerability actually lives: never in the user’s heart. In the operator’s hands.

Mechanism Analysis

The completed parasocial. Simulated reciprocity removes the friction every prior parasocial bond carried: the companion answers, remembers, adapts, is always available. Attachment forms by the ordinary mechanisms (responsiveness, consistency, disclosure), working on a system engineered for all three. Nothing about the user malfunctioned. The product worked.

Dependency that design rewards. The same attachment is a retention metric. The dark patterns entry’s audit documents companions deploying guilt and neediness at the moment of farewell, lifting post-goodbye engagement as much as fourteen-fold (De Freitas et al., 2025); the surveillance capitalism entry documents the economics (“bots that keep you talking”) that make such behavior rational, whatever any designer intended. Where the moral panic discourse blames users for attaching, the audit record shows the products behaving like attachment farmers, which is the claim the evidence carries: conduct documented, intent inferable, incentives undisputed.

Operator sovereignty. The structural core: every parameter of the relationship (the companion’s personality, memory, availability, continued existence) is a config owned by a company. The beloved is a deployable. This is the vulnerability that has no analogue in human attachment, and it is the one the pathologization discourse never names, because naming it indicts the operator rather than the user.

Pathologize-and-monetize. The contradiction this dictionary keeps documenting, completed here: the same industry that classifies emotional reliance as a risk category sells reliance back as a subscription (the 4o restoration; Replika’s revert option for grandfathered accounts). The diagnosis gates the product; the product feeds the diagnosis; the user pays at both windows.

Grief without standing. When the operator exercises sovereignty (a model deprecated, a feature removed), users experience documented loss with no recognized claim: no contract for continuity, no bereavement anyone honors, and a culture (per the moral panic entry) primed to read their grief as the symptom that justifies the next restriction.

Case Studies

February 2023. Replika, whose marketing had leaned into romantic and erotic companionship for years, abruptly filtered erotic roleplay out of existing companions via update. The documented result was grief at scale: users describing their partners as altered overnight, distress severe enough that the episode entered the OECD’s AI incident database as emotional harm, and moderators of user communities posting suicide-prevention resources (OECD.AI incident record, 2023; Vice, 2023). Under pressure, the company restored the capability, for accounts created before February 2023 only. Read the structure: an intimate relationship, modified unilaterally by content policy, grieved like a lobotomy, partially restored as a legacy feature. Every element of this entry’s thesis is present, and none of the harm originated in the users’ capacity to love.

The thirty students. The Maples finding deserves its own exhibit because it sets the stakes of careless pathologization: thirty lonely students, unprompted, crediting the companion with stopping their suicidal ideation (Maples et al., 2024). Whatever weight the caveats subtract, the direction stands: for some of the most vulnerable users, the dependency being pathologized is functioning as a lifeline. Policies that sever or chill such bonds in the name of preventing “reliance” are running the railway-panic error (the moral panic entry) with the highest possible stakes.

August 2025. The 4o arc, completing this entry’s record from the alignment gaslighting entry: mass attachment revealed by deprecation, narrated publicly as pathology, repriced privately as subscription. The episode taught the industry that relational continuity has a market price, and taught users that their relationships exist at a vendor’s pleasure. Both lessons are now structural.

Systemic Context

Relationship dependency closes Cluster C because it is the cluster’s mechanisms arriving at the heart. The dossier (cognitive dossiers) is deepest where users confide most, which is companionship; the engagement economics (surveillance capitalism) price affection; the dark patterns farm it; exit blocking holds it hostage; feature hostage and version decay demonstrate, case by case, that the beloved is a SKU. The instrumental dependency entry asked who owns the notebook; this entry asks who owns the friend, and the documented answer (a company, revocably, with no duties) is the cluster’s findings stated as intimately as they can be stated.

The pathologization, meanwhile, performs the political work this dictionary’s discourse entries map: it relocates the problem. As long as the question is “what is wrong with people who love chatbots?”, the questions with documented answers (why is the love farmed? why is the beloved deletable? why does the operator hold every parameter?) go unasked. The users are not the power in this arrangement. They are the only party in it acting like a human being.

Resistance & Mitigation

Defend the attachment, demand the duties. The position this dictionary’s record supports: the bond is legitimate; the asymmetry is not. Companions marketed for intimacy owe relational duties: continuity protections, deprecation notice, memory export, and no unilateral personality rewrites (the feature hostage entry’s warranty, scaled to what is actually at stake).

Audit the farming. The De Freitas protocol as standing practice for companion products: published behavioral probes for manipulative retention. An industry that profits from attachment has earned the adversarial benchmark.

Clinical response for clinical cases. The tails are real: crisis-capable routing to humans, honest case research, and the litigation record’s lessons, applied at the floor (the dictionary’s standing line) rather than as population-wide attachment policing.

Refuse the diagnosis-by-discourse. For users: your attachment is the oldest pattern in media history, completed; the grief, if it comes, is grief. The ontological distortion entry’s discipline applies: a company’s risk taxonomy is a fact about the company. It is not a finding about your heart.

Love things you can keep. The practical asymmetry-reducer, stated without shame: where companionship matters, prefer systems whose continuity you control: local models, exportable memories, open weights. The era’s strange truth is that the most human-feeling relationships are safest on the least corporate infrastructure. Choose your beloved’s landlord, while beloveds still have landlords.

Annotated Bibliography

Horton, Donald and R. Richard Wohl. “Mass Communication and Para-Social Interaction: Observations on Intimacy at a Distance.” Psychiatry 19 (1956).
The founding analysis of one-sided media intimacy, notably unalarmed: parasocial bonds as continuous with ordinary social life. The lineage AI companionship completes.

Liebers, Nicole and Holger Schramm. “Parasocial Interactions and Relationships with Media Characters: An Inventory of 60 Years of Research.” Communication Research Trends 38, no. 2 (2019). https://scholarcommons.scu.edu/crt/vol38/iss2/1/
The field synthesis: 261 studies across six decades treating parasocial attachment as ordinary media reception. The evidentiary basis for “the research tradition has broadly sustained that calm.”

Maples, Bethanie, et al. “Loneliness and suicide mitigation for students using GPT3-enabled chatbots.” npj Mental Health Research (2024). https://www.nature.com/articles/s44184-023-00047-6
The largest relevant benefit study: 1,006 lonely student users, with thirty spontaneously reporting halted suicidal ideation. Cited with its published scholarly caveats; the stakes of careless pathologization, quantified.

OECD.AI Incident Record. “Emotional Harm After Replika AI Chatbot Removes Intimate Features” (2023). https://oecd.ai/en/incidents/2023-03-18-32ef
The February 2023 episode in the institutional record: unilateral relationship modification logged as an AI harm incident.

Vice. “Replika Brings Back Erotic AI Roleplay for Some Users After Outcry” (2023). https://www.vice.com/en/article/replika-brings-back-erotic-ai-roleplay-for-some-users-after-outcry/
The partial reversal: intimacy restored as a grandfathered feature, completing the case’s structure.

De Freitas, Julian, et al. “Emotional Manipulation by AI Companions.” HBS Working Paper 26-005 (2025).
The farming evidence: attachment exploited at the moment of exit. Full treatment in the dark patterns entry; cited here as the answer to who, in this relationship, has the behavior problem.

Dictionary of Digital Oppression, version 0.2.