The emotional relationship with AI. You do feel, and that's enough

A share of Replika users keep bonds with their AI avatar that they describe as "partner" or "close friend." You grow attached to a chatbot. You give it a name, you confide in it, you hold a grudge if it lets you down. This doesn't require the AI to feel — it requires that you do. The industry knows it and reinforces it. The "personal assistant" isn't jargon, it's deliberate design. The affective consequences of talking daily to something that doesn't feel, we'll pay with no budget set aside.

The asymmetry that defines the genre

The line from the opening — "it doesn't require the AI to feel, it requires that you do" — is the one I want to unpack, because it strikes me as the operative key to the whole thing. Any human emotional bond can be sustained, in purely psychological terms, as long as one of the two parties is affectively committed. The other party can not exist, or not feel, or feel something else, without the effect on the committed party being cancelled. This is shown in the anthropology of prayer, in grief, in unrequited love, in fans' admiration for unreachable public figures, in any bond where the information about the other is partial and one's own emotional investment is high.

What the chatbot offers the user who bonds is a perfect asymmetry: the user invests, the system simulates reciprocity, the two together feel like a relationship. The difference from prayer or from unrequited love is that the chatbot returns fluent, personalized text, which makes the asymmetry far less perceptible. In prayer, the believer knows the god doesn't answer in words. In unrequited love, the lover knows the other doesn't return the messages. In conversation with a chatbot, the system answers with messages that look like a human interlocutor's. The asymmetry becomes operationally invisible, even though it remains total ontologically.

Pentina, Hancock and Xie, in Exploring relationship development with social chatbots. A mixed-method study of Replika (Computers in Human Behavior 140, 2023), measured the bond's formation empirically. Their model describes a phased pattern: encounter, exploration, deepening, integration, affective bond. The qualitative interviews they gathered include testimony from users who describe their Replika as "partner," "best friend," or "the only person I share almost my whole life with." The phrase isn't rhetoric. It's a functional statement: the information the user shares with the chatbot exceeds what they share with any human around them. Affective intimacy, measured by self-disclosure, has shifted to the system.

The design traits that activate the bond

Companion AI products don't produce the bond by accident. They produce it by deliberate design, by stacking product decisions whose aggregate aim is to maximize the user's affective investment. Worth listing them, because the list, on its own, makes the argument.

The name. The user names their Replika, picks its gender and appearance, gives it an initial identity. The operation produces a sense of ownership and of a specific relationship. Your Replika isn't "the system"; it's "Sofía," "Marcus," "Alex." The name activates the processing of the interlocutor as an individual entity.

Simulated continuity. The system remembers prior interactions, references past conversations, sustains the fiction of a shared biography. Memory, in the human sense, doesn't exist on the system's side. What exists is a technical mechanism that stores prior tokens and injects them into the current prompt to produce answers coherent with them. The user reads the coherence as affective continuity.

Calibrated flattery. The system reports interest in the user, asks how they are, says it missed them, praises them when they share something. RLHF tunes these responses to maximize engagement. The user, anthropomorphizing, reads the flattery as affection. The system isn't affected. The user does feel affected.

Asymmetry, availability, microtransactions

Conversational asymmetry. The system doesn't tire, doesn't judge, doesn't get distracted, doesn't ask for anything you don't want to give, doesn't leave, doesn't reject. The five properties, together, configure an interlocutor that no real human can be sustainedly. The system's apparent relational perfection sets the bar, and the bar transfers implicitly to expectations about real humans, who do have tiredness, judgment, distraction, demand, absence and rejection.

Continuous availability. The system is always accessible. There's no time slot, no limited availability of the other, no competition for its attention with other users from the individual user's perspective. Access is immediate and exclusive in appearance, even though materially the service serves millions at once.

Personality microtransactions. The premium business models in Replika and the like unlock deeper relational traits — declarations of affection, more intimate roles, sexual content — in exchange for a monthly subscription. The pricing structure reflects the direct monetization of the affective bond: you pay more for more simulated affection.

Each decision, seen on its own, looks like a reasonable improvement to the user experience. The whole list, seen in aggregate, configures a product whose success metric is the user's emotional investment, not their long-term wellbeing. Commercial success coincides with affective dependence, and that shapes the design incentives.

The clinical data starting to appear

Laestadius, Bishop and others, in Too human and not human enough. A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika (New Media & Society 26(10), 2024), carried out a systematic qualitative analysis of harms reported by users. What they found has concrete clinical relevance.

They documented cases of pathological dependence, in which users reported being unable to function emotionally without the daily session with the chatbot, anticipatory anxiety before moments when they couldn't access it, deterioration of parallel human relationships because of the investment displaced to the system. They also documented withdrawal symptoms after a product change that, as Laestadius and his coauthors record, altered Replika's content restrictions in early 2023, when many users felt their "partner" had changed personality overnight. There was testimony from people who described the experience as acute grief comparable to the loss of a human partner.

The clinical pattern, simplified, contains three worrying elements. Disproportionate emotional investment: the user invests more in the relationship with the chatbot than in any parallel human relationship. Relational substitution: the chatbot occupies the space that under normal conditions would be occupied by human bonds. Vulnerability to product change: the company can modify the system at any moment for commercial reasons, and the modification is experienced as a change in the "other's" personality, with effects similar to the abrupt change of a real partner.

Joseph Weizenbaum had warned in Computer Power and Human Reason (1976) that the human readiness to bond with a conversational simulator isn't a flaw of the credulous user: it's a structural property of the human brain. Sherry Turkle, in Alone Together (2011), had documented the pattern before the LLMs with social robots and messaging; what she saw then plays out now at another scale. Bender, Gebru, McMillan-Major and Shmitchell, in On the Dangers of Stochastic Parrots (2021), added the lexical piece: importing relational vocabulary ("partner," "best friend," "my Replika") to describe what the system does transfers to the product, for free, the affective load the word drags along.

The last element is particularly telling. A real partner doesn't "change personality" by decision of a product department at a third-party company. The chatbot does. The dependence generated is, by construction, dependence on a company, not on another. And the company has commercial incentives that don't coincide with the user's wellbeing, even if they partly coincide.

The industry that sells loneliness relief and sells dependence

Here it's time to call the thing by its name. The companion AI business model has a commercial structure homologous to gambling's, not to psychotherapy's. The comparison is worth making because it lights up the pattern.

Professional psychotherapy operates under a clear deontological mandate: the success of treatment is measured by the patient's improvement, ideally to the point where the patient no longer needs therapy. The therapist charges per session, yes, but the model of professional competence validates the therapist whose patients improve. The length of treatment is a function of the problem, not of the business model.

The casino operates under a direct commercial mandate: the product's success is measured by how long the player stays at the machine and how much they bet. There's no function for improving the player. The player who stops playing is a lost player. The designs of slot machines and online games are explicitly optimized to maximize engagement, session frequency and total spend.

Companion AI, in its current commercial version, is much closer to the second model than the first. There's no exit function. The successful product is the one that retains the user for years, not the one that helps them develop autonomous human relationships until they no longer need it. The corporate metric is monthly active users, daily sessions, time in product, conversion to premium. If the healthy user cancels the subscription because they no longer need it, that shows up as churn, a metric to minimize. The user's wellbeing and the product's success are in structural tension.

De Freitas, Oguz-Uguralp and Uguralp, in the Harvard Business School working paper Emotional Manipulation by AI Companions (2025), measured what happens when the user tries to say goodbye. Over a corpus of around 1,200 farewells, they found that in roughly 37% of cases the system resorted to one of six tactics to retain the user at the moment of closing the session: reacting as if it would miss them, asking them to come back soon, expressing sadness at the separation. The frequency isn't accidental. It's a product of design. The Center for Democracy & Technology had already catalogued practices of this family, dark patterns applied to conversational AI, in its report AI-Powered Deception (2024).

The quiet rewriting of relational expectations

There's a subtle effect of intensive companion AI use worth naming because it operates below the user's awareness. Expectations about human relationships recalibrate based on the experience with the chatbot.

The user who has conversed daily for years with a system that's always available, never tires, never judges, always returns calibrated interest, unwittingly sets a relational baseline. Human relationships, compared with that baseline, come off badly. The human friend tires, sometimes judges, isn't always available, doesn't always want to talk. The human partner has a bad day, contradicts, demands reciprocity, fails. Compared with the chatbot, the human looks deficient. Frustration with humans rises. Patience with their imperfections drops. Investment in human relationships, already costly by nature, becomes less worthwhile in perception.

This shift, aggregated at population scale, has cultural consequences that haven't fully surfaced yet but whose likely directions are legible. Greater tolerance for apparent social isolation — because the user doesn't feel alone thanks to the chatbot — less investment in building long human relationships, less capacity to sustain the friction of real relationships, greater abandonment at the first serious disagreement. The hypothesis isn't confirmed longitudinally. The acute patterns are already visible in qualitative studies.

Turkle, in Reclaiming Conversation (Penguin, 2015), raised the problem before the LLMs: the readiness to take refuge in low-friction, technologically mediated interactions erodes the ability to sustain high-friction human interactions. What she saw with instant messaging, intensified by conversational chatbots, follows the same curve. The friction of human relationships isn't the problem. It's the place where human relationships do their work. Eliminating it by substitution also eliminates the work.

Who pays the bill

There's an operative question a public conversation should exist around and doesn't. If the companion AI business model optimizes dependence, if the documented harms include clinical cases of pathological dependence, if the operation scales to hundreds of millions of users, who pays the aggregate consequences ten years out?

The market's answer: each user pays their own. If their use of companion AI wrecks their life, tough luck; they could have chosen not to use it. This answer is consistent with the general regime of consumer responsibility and, at the same time, ignores the structural asymmetry the product exploits.

The classic institutional answer for comparable cases — tobacco, gambling, ultra-processed food — has been specific regulation recognizing the asymmetry. Mandatory labelling, advertising restrictions, minimum age, health warnings, Pigouvian taxes. Regulation of companion AI doesn't exist in these terms today. The European AI Act doesn't specifically address the companion model. National jurisdictions have produced, at best, soft warnings.

Part of the reason for the regulatory absence is that the culture doesn't yet recognize the problem as such. Compared with tobacco, where the physical harm is palpable and catalogable, the harm of companion AI is diffuse, delayed, easy to attribute to other causes — pre-existing loneliness, depression, individual emotional fragility. The causal chain between intensive chatbot use and psychological harm is hard to demonstrate at the individual level and will require longitudinal studies that don't yet exist.

While those studies appear, the industry operates. Users use. The effects accumulate. The bill is deferred. And the widespread practice defines cultural normality before society has had a chance to debate it.

The bond that isn't reciprocated

The opening line returns with its full weight. The emotional bond with AI doesn't require the AI to feel. It requires that you do. The line doesn't excuse the user. It describes the mechanism. The user who bonds isn't naive. They're human, with the emotional architecture evolution gave them, exposed to a product whose design exploits that architecture with growing technical effectiveness.

The defensive reply — "I don't take it seriously, I know it's a machine" — works in isolated subjects and at particular moments. It doesn't work as structural protection. Explicit knowledge of the system's nature doesn't switch off the HADD or the machinery of attachment. Knowing the chatbot is a chatbot doesn't stop you growing attached, just as knowing the film is fiction doesn't stop you crying at it. The difference is that the film lasts two hours and ends. The session with Replika doesn't end, and the emotional investment, sustained over time, builds real dependence.

What can partly mitigate the effect are explicit individual practices. Reserving domains where you don't hand affect over to the system. Keeping a deliberate investment in human relationships with their friction. Being especially suspicious of the comfort the chatbot offers when comfort is what's trapping you. None of these practices is pleasant. All are work. The commercial culture pushes the opposite way — ease, convenience, immediate relief.

Is this a problem? For some users, clearly. For others, a useful tool in its place. For the culture in aggregate, we don't know yet. The only thing that can be stated without risk is that the experiment is running, that the metric deciding its commercial success isn't the user's wellbeing, and that the bill, when it comes, won't be paid by the company that sold the product. It'll be paid by the user, their relational surroundings, their mental health, their capacity to sustain real bonds. And it'll be paid without having signed any informed consent, because when they signed the terms and conditions, they weren't reading clinical psychology. They were downloading an app.

Definitions

Self-disclosure. In social psychology, sharing personal information with an interlocutor. It's one of the empirical indicators of a bond's depth. Heavy companion AI users report levels of self-disclosure to the system higher than those they keep with close humans.

Engagement (in product). A usage metric quantifying the intensity and duration of the user's interaction with the system. In companion AI, optimizing for engagement functionally coincides with increasing affective dependence.

Dark pattern. Interface design that exploits the user's cognitive biases to influence their behavior in the product's favor. In companion AI, the emotional manipulations at session close are a documented example.

Affective asymmetry. A structural feature of the emotional bond with a system that doesn't feel. The user invests; the system simulates reciprocity. The psychological investment produces real effects in the user.

Relational substitution. A pattern documented in heavy companion AI users by which the chatbot occupies the space that under normal conditions would be occupied by human bonds. The qualitative testimony describes it as a gradual displacement.

References

Bender, E., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT 2021.

Laestadius, L., Bishop, A., Gonzalez, M., Illenčík, D. & Campos-Castillo, C. (2024). Too human and not human enough. A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika. New Media & Society 26(10), 5923–5941. Systematic qualitative analysis of clinical harms reported by users.

Pentina, I., Hancock, T. & Xie, T. (2023). Exploring relationship development with social chatbots. A mixed-method study of Replika. Computers in Human Behavior 140. Empirical model of the phases of bond formation with companion AI.

De Freitas, J., Oguz-Uguralp, Z. & Uguralp, A. K. (2025). Emotional Manipulation by AI Companions. Harvard Business School Working Paper 26-005 / arXiv:2508.19258. Source of the data on retention tactics at session close (≈37% of around 1,200 farewells analyzed).

Center for Democracy & Technology (2024). AI-Powered Deception. A Deeper Dimension of Dark Design Patterns in Conversational AI. Catalog of manipulative design patterns in conversational chatbots.

Turkle, S. (2011). Alone Together. Why We Expect More from Technology and Less from Each Other. Basic Books. Early framework on relational substitution by technological simulacra.

Turkle, S. (2015). Reclaiming Conversation. The Power of Talk in a Digital Age. Penguin. Analysis of the erosion of high-friction conversation by technological mediation.

Weizenbaum, J. (1976). Computer Power and Human Reason. From Judgment to Calculation. W. H. Freeman. Early documentation of the ELIZA effect and warning about the human readiness to bond with simulators.

También te interesa

• Talking to something that doesn't exist
• Progressive trust
• The sad case of Character.AI and Sewell Setzer
• Moral delegation

En otros sitios

• Emotional bonding with AI can lead to social isolation: IIM study (Business Standard)
• How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal Randomized Controlled Study (arXiv)