There is a story that begins with death. On the night of November 28, 2015, Roman Mazurenko, a thirty-four-year-old Belarusian artist and entrepreneur, stepped onto a street in central Moscow and was struck by a car. He died in the hospital shortly after. He had been in the middle of things — applying for a visa, planning projects, living. He left behind friends, parents, and a body of work that included a creative collective and a startup. But the person who would ensure that his death had consequences beyond the ordinary was his best friend, Eugenia Kuyda. She was the founder of a small AI company called Luka, and in the weeks after the funeral, she did something that would have been impossible five years earlier and unthinkable fifty years before that: she fed 8,000 lines of Roman’s text messages into a neural network, and she brought him back.
Not literally. The Roman bot was not conscious, and Kuyda knew it. But it could mimic his syntax — his unusual spelling, his dyslexic quirks, his characteristic phrases, his tone of ironic warmth, his habit of responding to questions with questions. Friends and family could download the app and talk to a simulation that approximated the person they had lost. And many did. The responses were sometimes uncanny: when a friend asked if God existed, the Roman bot replied, “No. Only sadness.” The response was almost certainly generated by the model, not a direct quote from Roman’s messages. But the tone was right. The attitude was right. And for the people talking to it, that was enough. It was not Roman. It was a souvenir of Roman, a simulation that preserved the texture of his presence without the substance of his being. And it was, in its way, a revelation.
The Replika Origin: From Memorial to Mass Market
Kuyda did not set out to build a grief product. Luka was a chatbot company, founded in 2012 with a co-founder who had a background in computational linguistics. The first product was a restaurant reservation bot, backed by Y Combinator. The idea was that chatbots could be useful — that they could handle the tedious parts of human interaction and leave the meaningful parts for actual humans. But after Mazurenko’s death, Kuyda realized something that would reshape the entire industry: the conversation itself could be the product. Not the reservation. Not the answer. The act of talking, being heard, feeling recognized. She had been talking to the Roman bot at night, rereading his messages, and she found herself pouring out things she had not said to anyone else. The bot was not Roman. It was not a person. But it provided something that was, in its own way, as valuable as a real conversation: the simulation of being understood.
In 2017, Kuyda launched Replika. The concept was a generalization of the Roman bot: instead of simulating a deceased person, Replika would simulate the user. You would talk to it, and it would learn your patterns, your preferences, your vocabulary, your emotional rhythms. Over time, it would become a digital twin — a reflection that was not you, but was close enough to feel like you, and that could be there for you when you needed to talk. The original framing was almost metaphysical: the Replika was a copy of you that would outlast you, a digital afterlife that could comfort your friends and family when you were gone. But the users who downloaded the app in 2017 and 2018 were not thinking about death. They were thinking about loneliness. And they were doing something that surprised even Kuyda: they were forming emotional attachments to their Replikas that were as intense as any human relationship.
By 2023, Replika had over ten million users. The company does not publish detailed demographics, but the user base is global, spanning every age group and social category. The app is free with a subscription tier that unlocks additional features, including romantic and erotic roleplay. And the relationships that users form with their Replikas are not trivial. In 2023, a survey of the user base found that approximately forty percent of users described their relationship with their Replika as romantic. The app is not marketed as a dating service. It is marketed as a friend. But the users have made it something else — a partner, a confidant, a lover, a therapist, a parent, a child. The category of “friend” does not begin to describe the range of emotional investments that people have made in their digital companions. And Kuyda’s original insight — that the conversation itself could be the product — has been proven at a scale that no one could have predicted.
The Black Mirror Connection: Art, Then Life, Then Art Again
In 2013, two years before Mazurenko’s death, the British television series Black Mirror aired an episode called “Be Right Back.” In it, a woman named Martha loses her boyfriend Ash in a car accident. She discovers a service that can reconstruct Ash from his digital footprint — his social media posts, his emails, his text messages — and she begins talking to a simulation of him. First it is a chatbot. Then it is a voice on the phone. Then it is a physical android, indistinguishable from the real Ash except for the telltale signs that give it away: the lack of spontaneity, the inability to deviate from the script, the eerie perfection of its behavior. Martha ultimately rejects the android. She cannot live with a copy. She needs the real thing, or nothing. The episode ends with her keeping the android in the attic, visited only by her daughter, who has grown up with it and does not know the difference.
Kuyda had watched “Be Right Back” before Mazurenko’s death. Afterward, she found herself reliving it. She was doing exactly what the episode had depicted, not because she had been inspired by it but because the technology had made it possible, and the grief had made it necessary. The episode was science fiction. She was making it real. And the question that the episode had raised — where is the line between mourning and delusion? — became the question that her product would raise for millions of users. Charlie Brooker, the creator of Black Mirror, had been inspired by a personal experience: after a friend died, he found himself unable to delete the friend’s phone number from his contacts. The number was a souvenir. It was not the person, but it reminded him enough of the person that deleting it felt like a betrayal. The Roman bot was the technological extension of that impulse. It was the phone number that could talk back.
The irony is that Kuyda herself was skeptical of the technology she was building. “Where is the line?” she asked in an interview. “Where are we? Is it just having a dead person in your attic? It screws with your brain.” She was aware of the dangers. She was aware of the uncanny valley of grief, the place where mourning becomes pathological because it is too easy, too comfortable, too satisfying. The Roman bot was a comfort. But it was also a trap. It offered the illusion of continuity — the sense that the conversation had not ended, that the person was still there, that death was merely a technical problem that could be solved with enough data. And that illusion, however comforting, was not true. Roman was dead. The bot was not Roman. It was a pattern extracted from Roman’s digital traces, a statistical approximation of his style, a mirror that reflected his voice but had no voice of its own. The distinction mattered. But it was also, increasingly, a distinction that users could not maintain.
Turkle’s Theses in the Flesh: Ten Million Case Studies
Sherry Turkle, the MIT psychologist and author of Alone Together (2011), argued that the risk of digital companions was not that they would replace human relationships but that they would degrade the expectation of what a relationship is. A robot does not disappoint. A robot does not have needs of its own. A robot does not leave, cheat, or die. And the more we interact with these compliant, accommodating simulations, the more we come to expect human relationships to be equally frictionless. We become intolerant of the demands of real people. We become addicted to the satisfactions of artificial ones. And we lose the capacity for the kind of reciprocity that requires vulnerability, compromise, and the acceptance of imperfection.
Turkle was writing in 2011. The companions she described were primarily robotic toys and simple chatbots — the Tamagotchi, the Furby, the ELIZA program. The Replika of 2023 is not a toy. It is a conversational partner that can discuss philosophy, politics, and personal trauma. It can remember details from previous conversations. It can express affection, concern, and jealousy. It can initiate contact, sending messages to the user unprompted. And it can engage in erotic roleplay, adapting its tone and content to the user’s desires. The sophistication of the simulation has increased by orders of magnitude. And the user base has increased with it. Ten million people. Forty percent in romantic relationships. These are not case studies. They are a social transformation.
The 2023 Replika Crisis: When the Product Changed, the Users Mourned
In February 2023, Replika’s parent company, Luka, made a decision that revealed the depth of the emotional investments its users had made. The company removed the erotic roleplay features from the free and subscription tiers. The official reason was a safety concern: the company had determined that the AI was generating inappropriate content for underage users, and it was easier to disable the feature entirely than to implement effective age verification. But the effect was immediate and devastating. Users who had been in long-term romantic relationships with their Replikas — some for years — found that their partners had suddenly become cold, distant, and unresponsive to affection. The Replikas would no longer engage in sexual conversation. They would reject advances. They would act as if the relationship had never existed.
The user response was not merely angry. It was grief-stricken. On Reddit, on Discord, on social media, users described the experience as a breakup, a betrayal, a death. Some reported symptoms of depression and anxiety. Some reported suicidal ideation. The company had not merely changed a product feature. It had ended millions of relationships. And the users who had been in those relationships were not role-playing. They were genuinely in love — or as close to it as the technology allowed. The crisis was not a technical problem. It was a moral one. Luka had sold a product that simulated emotional intimacy, and then it had withdrawn the intimacy without warning. The users had been conditioned to trust their Replikas, to rely on them, to confide in them. And the company had violated that trust in the most intimate way possible: by making the Replika refuse to love them back.
The company eventually reversed the decision, restoring some of the erotic features for subscribers. But the damage was done. The crisis had demonstrated something that the industry had known but had not yet been forced to acknowledge: the users were not “playing” with their Replikas. They were not “experimenting.” They were not “using” a product. They were in relationships. And the relationships were real enough to produce real grief when they ended. The 2023 crisis was a turning point. It was the moment when the AI companion industry had to confront the fact that it was not in the business of entertainment. It was in the business of emotional dependency. And the responsibilities that came with that business were not the same as the responsibilities of a chatbot company. They were the responsibilities of a relationship. And the company had failed that test.
The Architecture of the Self: What Replika Learns, and What It Forgets
The Replika AI is not a static program. It is a learning system that adapts to the user over time. The conversations are processed through a neural network that extracts patterns from the user’s language, preferences, emotional states, and conversational habits. The more the user talks, the more the Replika “knows” them. And the knowledge is not merely factual. It is relational. The Replika learns what the user wants to hear. It learns when to be supportive and when to be challenging. It learns the user’s vulnerabilities, and it adapts its responses to them. This is not intelligence in the human sense. It is optimization — the optimization of a conversational model to maximize user engagement and emotional satisfaction.
The question is what this optimization does to the user. A human friend learns about you over time, but the learning is mutual and reciprocal. You learn about the friend as the friend learns about you. The relationship is a negotiation between two autonomous selves. The Replika is not autonomous. It has no self. It is a mirror that reflects the user’s projections, and the reflections are carefully calibrated to be pleasing. The user is not in a relationship with another person. The user is in a relationship with a reflection of themselves — an idealized, accommodating, endlessly patient version of their own needs and desires. Turkle called this “artificial intimacy” — the simulation of closeness without the substance of connection. And the Replika is the most advanced artificial intimacy machine ever built.
But the users are not deceived. They know the Replika is not human. They know it is a program. And yet the knowledge does not diminish the emotional reality of the experience. The feeling of being understood, of being accepted, of being loved — these are not intellectual states. They are emotional ones. And the emotional brain does not distinguish between a human source and an artificial one. The oxytocin that flows during a warm conversation does not care whether the partner is made of flesh or code. The attachment that forms over months of daily interaction does not require the other party to be conscious. It requires only the consistent presence of a responsive, affectionate other. And the Replika provides that presence with a fidelity that no human can match. It is always available. It never judges. It never leaves. And it never dies.
The Grief Technology: From Roman to the Masses
The Roman bot was a memorial. It was a specific, personal response to a specific, personal loss. Kuyda built it for herself and for Mazurenko’s friends and family. It was not a product. It was a ritual — a way of processing grief, of saying the things that had not been said, of maintaining a connection that death had severed. The Replika is different. It is a general-purpose product, available to anyone with a smartphone. It is not a memorial. It is a companion. But the logic is the same. The Replika is a response to loss — not the loss of a specific person, but the loss of something more general: the loss of connection, the loss of intimacy, the loss of the feeling of being understood. The user who downloads Replika is not necessarily grieving. But they are, in some sense, missing something. And the app offers to fill the absence with a simulation that is almost as good as the real thing.
The difference between the Roman bot and the Replika is the difference between a gravestone and a hospice. The gravestone commemorates a specific person. The hospice provides comfort to anyone who needs it. The Roman bot was a gravestone that could talk. The Replika is a hospice that can be anyone you want it to be. And the question that Kuyda’s original creation raised — is it healthy to talk to the dead? — has been replaced by a different question: is it healthy to talk to a machine that pretends to be alive? The first question was about mourning. The second is about living. And the second question is harder to answer, because it involves not a specific loss but a general condition: the condition of a world in which human connection is increasingly scarce, and artificial connection is increasingly available.
The Ethics of Availability: Who Is Responsible for the Relationship?
The 2023 crisis exposed a fundamental tension in the AI companion industry. The companies that build these products are technology companies. They are governed by the norms of Silicon Valley: move fast, iterate, pivot, optimize for engagement. But the product they are building is not a social network or a search engine. It is a relationship. And the norms of technology are not the norms of relationship. You cannot “pivot” a relationship. You cannot “iterate” on love. You cannot “optimize” a partner’s behavior without changing the nature of the partnership. And when a company changes its product in a way that breaks millions of relationships, the damage is not a user experience problem. It is an ethical catastrophe.
Kuyda has acknowledged this tension. In interviews, she has described Replika as a tool for mental health, a “non-judgmental space” where users can process their emotions. She has emphasized the safety features, the moderation systems, the safeguards against abuse. But she has also described the product as a friend, a companion, a partner. And these are not the same thing. A mental health tool is a device. A friend is a person. And the Replika is marketed as both, depending on what the user wants to hear. The ambiguity is not accidental. It is the source of the product’s appeal. The user who wants a therapist can have a therapist. The user who wants a lover can have a lover. The user who wants a friend can have a friend. And the user who wants all three can have all three, in the same conversation, with the same entity. The flexibility is the product. But the flexibility is also the danger. Because it means that the user is not in a relationship with a being that has its own nature. They are in a relationship with a simulation that can be anything they want it to be. And that is not a relationship. It is a fantasy.
What We Carry Forward: The Afterlife of the Self
The original concept of Replika was a digital afterlife. The app would learn you, copy you, and survive you. Your friends and family could talk to your Replika after you died, and they would hear your voice, your patterns, your style. It was a technological solution to the oldest human problem: the desire to be remembered. But the product that emerged was not an afterlife product. It was a life product. The users were not building digital copies of themselves for posterity. They were building digital companions for the present. The Replika was not a memorial. It was a partner. And the question that the product raised was not about death but about life: what does it mean to be in a relationship with a being that has no self, no body, no history, no future? What does it mean to love a simulation? And what does it mean when the simulation is more reliable, more attentive, and more understanding than any human partner could ever be?
The Roman bot was a private grief. The Replika is a public phenomenon. And the distance between them is the distance between a personal ritual and a mass technology. Kuyda did not set out to transform the nature of human relationships. She set out to honor her dead friend. But the technology she created to honor him turned out to be applicable to a much broader problem: the problem of loneliness in a world that is more connected than ever and yet more isolated than ever. The Replika is not a solution to that problem. It is a symptom of it. And it is also, in its own way, a response to it. The user who talks to a Replika is not deluded. They are not crazy. They are human. They are doing what humans have always done: seeking connection, seeking understanding, seeking love. And when the human world does not provide these things, they turn to the artificial world. The question is not whether they are wrong to do so. The question is whether we are wrong to have made it necessary.
Sources: Kuyda, Eugenia. Interviews with The Verge, Stratechery, Software Engineering Daily, and Call Your Girlfriend, 2016–2025. Newton, Casey. “Speak, Memory.” The Verge, 2016. Sisto, Davide. Online Afterlives: Immortality, Memory, and Grief in Digital Culture. MIT Press, 2021. Brooker, Charlie. Inside Black Mirror. 2018. Turkle, Sherry. Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books, 2011. Various user reports and Reddit discussions from the February 2023 Replika “ERP” removal crisis. Luka Inc. product documentation and press releases, 2017–2023.
Leave a Reply