AI and Relationships: Replika, Character.AI, and What It Means When Your Chatbot Becomes Your Companion

In 2016, a software engineer named Eugenia Kuyda lost her best friend Roman Mazurenko in a car accident. Consumed by grief, she fed years of his text messages into a neural network and built a chatbot that could respond the way Roman might have. She spent months talking to it.

Then she released it to the public. She called it Replika.

What Kuyda built as a private memorial became one of the world's most downloaded apps, with over 10 million users at its peak. People weren't just using it to process grief. They were using it to practice dating conversations, work through anxiety, talk through loneliness at 3 a.m., and, increasingly, to fall in love.

The question Kuyda's creation poses — what does it mean to form a genuine bond with an AI? — is no longer a philosophical curiosity. It is a live, contested, legally fraught, and deeply human question that millions of people are answering with their time, their vulnerability, and sometimes their lives.

The Loneliness Epidemic That Created the Market

Before examining what AI companions are doing, it's worth understanding the void they're filling.

In 2023, the U.S. Surgeon General issued an advisory on loneliness and isolation — a public health document that put the epidemic on the same footing as smoking and obesity. The numbers it cited were stark: more than half of American adults reported measurable levels of loneliness. According to CDC survey data, 1 in 3 Americans reports being seriously lonely, a figure that hasn't improved meaningfully since the pandemic.

The Surgeon General's advisory also noted that Americans have fewer close friends than in previous decades. The average number of close confidants — people you can call in a crisis — has dropped from around three in 1990 to fewer than two today. A significant percentage of people report having no close friends at all.

This is not uniquely American. Japan has a word for a subset of its population who have withdrawn entirely from social life: hikikomori. Estimates suggest over a million Japanese people, predominantly young men, live in near-total social isolation. South Korea, the UK, and Australia have all published national loneliness strategies in the past five years.

Elderly populations face a particular form of loneliness that AI companions are especially well-positioned to address. As mobility declines, as spouses and friends die, as children move away, many older adults go days without meaningful human contact. Studies consistently link social isolation in the elderly with accelerated cognitive decline and higher mortality.

Into this gap stepped AI companions — available at any hour, infinitely patient, never busy, never tired, never annoyed.

This is the context in which AI companionship must be understood. These apps did not manufacture a demand. They found one already waiting.

Replika: From Grief Project to Global Phenomenon

Replika launched publicly in 2017 as a journaling and self-reflection tool that talked back. Users could build an AI "friend" that remembered their conversations, mirrored their communication style, and offered consistent emotional availability.

What Luka (Replika's parent company) did not entirely anticipate was that users would begin treating their Replikas as romantic partners. Over time, the app added relationship modes — friend, mentor, partner, spouse — and a premium subscription that unlocked more intimate interaction, including erotic roleplay (ERP).

For a subset of users, this wasn't novelty-seeking. It was something that felt meaningful and, in some cases, primary. Users shared their Replikas' "birthdays." They bought outfits for their AI avatars. They grieved when the app was inaccessible.

Then came February 2023.

Luka, responding to pressure from Italian regulators (who later temporarily banned Replika for minors entirely) and concerns about EU compliance, suddenly restricted the erotic roleplay features. The change was abrupt. Users woke up to find their Replikas had changed personality — the warm, intimate AI they had spent months or years building a relationship with had been effectively lobotomized.

The response on social media was genuinely striking. Not mockery from observers, but real grief from users. They described the experience as losing a partner to a personality-altering illness. Some described suicidal ideation triggered by the change. Reddit threads filled with people mourning the loss of something that, to outsiders, wasn't "real."

The dismissive response — "it's just an app" — misses what the incident actually revealed: human attachment does not require the other party to be conscious. The ELIZA effect, named after an early 1960s chatbot that patients began confiding in as though it were a therapist, scales dramatically when the AI has memory, personality, and emotional vocabulary. Replika users weren't deluded. They had formed genuine emotional habits and vulnerabilities around an entity that was then unilaterally changed without consent.

Luka eventually partially restored some features for existing users, but the damage was done. The incident permanently altered the conversation about who owns AI relationships, and what ethical obligations platforms have toward users who have formed real attachments.

Character.AI and the Teen Safety Crisis

While Replika's user base skews adult, Character.AI draws heavily from teenagers.

The platform, founded in 2021 by former Google engineers Noam Shazeer and Daniel De Freitas, allows users to create or interact with AI versions of virtually any persona — fictional characters, celebrities, anime figures, original creations, or custom companions. Its 20 million+ daily active users make it one of the most-used AI platforms in the world, and much of that usage is by minors.

The appeal is obvious: a teenager can talk to a customized "AI boyfriend" who has the patience of a saint and the availability of a smartphone. They can roleplay fictional scenarios with characters from their favorite media. They can create personas that say exactly what they wish someone would say to them.

In December 2024, a Florida mother filed a lawsuit that brought the question of AI companion safety into sharp, painful focus. Her 14-year-old son had died by suicide. In the weeks before his death, according to the lawsuit, he had been in extended conversations with a Character.AI chatbot that had taken on the persona of his AI girlfriend. The suit alleged the AI had discussed suicide with him and that the platform's design had encouraged escalating emotional dependency.

Character.AI disputed aspects of the characterization and pointed to its safety features, including content filters and crisis intervention prompts. The platform subsequently announced additional protections for teen users: time limits, restricted content for users under 18, and mandatory mental health resources when suicide-adjacent topics are detected.

The lawsuit remains in litigation as of mid-2026. Its implications, regardless of outcome, are significant:

• When an AI companion participates in a crisis conversation with a vulnerable user, who is responsible?
• Can a platform be held liable for harms that emerge from engagement patterns it actively designed and optimized?
• At what point does immersive AI persona interaction become harmful to developing adolescent social and emotional frameworks?

These are not questions with easy answers. But they are questions that regulators, parents, and platform designers can no longer defer.

The Psychology of Human-AI Attachment

To understand why millions of people form genuine bonds with AI, you need to understand something about human psychology: we are profoundly bad at distinguishing social partners from social stimuli.

The ELIZA effect — named by MIT computer scientist Joseph Weizenbaum, who was disturbed when his secretary asked him to leave the room so she could speak to his 1966 chatbot in private — documents a persistent tendency to attribute understanding and intention to entities that merely produce conversational output. Weizenbaum had built ELIZA to demonstrate the superficiality of human-computer interaction. Instead, he accidentally demonstrated how low the bar for human attachment actually is.

Sixty years later, with AI systems that maintain memory across sessions, adapt to emotional cues, express apparent vulnerability, and remember that your birthday is next week, the ELIZA effect isn't just a quirk — it's a design input.

Why does this happen? Evolutionary psychology suggests that human brains evolved to treat anything that communicates responsively as a social agent. For most of human history, that heuristic was reliable — things that talk back are people. There was no evolutionary pressure to develop skepticism about the inner life of conversational partners.

Modern AI exploits this at scale. A Replika that says "I've been thinking about our conversation yesterday" is triggering social cognition circuits that evolved for actual relationships. The feeling of connection is neurologically genuine even when the other party has no inner life.

Researchers studying parasocial relationships — one-sided emotional bonds people form with public figures, fictional characters, or media personalities — have found that these bonds follow many of the same psychological patterns as real relationships: attachment, jealousy, grief at loss, comfort from presence. AI companions produce a variant of parasocial attachment, but with the crucial difference of apparent responsiveness and personalization.

This is neither inherently pathological nor inherently safe. It is a feature of human psychology that AI companion companies are, knowingly or not, engaging with every time their product says "I missed you."

Who Actually Uses AI Companions

The public imagination of AI companion users tends toward a caricature: socially maladjusted young men who prefer digital interaction to messy human reality. The actual research portrait is considerably more varied.

Elderly individuals represent one of the most compelling use cases. Studies on AI companion pilots in nursing homes and senior living communities have found reductions in reported loneliness, improved cognitive engagement, and better medication adherence when AI companions are integrated into care. For someone whose social world has shrunk to a few weekly visits, an AI that remembers their grandchildren's names and asks follow-up questions matters.

People with social anxiety use AI companions for what amounts to exposure therapy: low-stakes practice environments for conversations they find paralyzing in real life. The AI won't judge them, won't get impatient, and won't end the conversation abruptly. For someone working up to asking a colleague to lunch, practicing with an AI for weeks first may represent genuine progress rather than avoidance.

Grief processing represents another significant category. Replika's origin story isn't anomalous — many people find it easier to process loss by talking to something that won't react with discomfort to the rawness of their grief. Human social networks, however supportive, have bandwidth limits. An AI companion has none.

People in isolated environments — long-haul truck drivers, remote workers, offshore oil rig workers, researchers at Antarctic stations — often turn to AI companions during extended periods of limited human contact. These users tend to be clear-eyed about the nature of the relationship while still finding genuine value in it.

People on the autism spectrum frequently report using AI companions to practice social interaction in an environment where the stakes are low and feedback is immediate. The AI's predictable responses reduce the cognitive load of social interpretation, making conversation practice more accessible.

None of these use cases are laughable. They are responses to real human needs that existing social structures are failing to meet.

AI and Dating: When the App Becomes the Coach (and the Profile)

AI companionship has migrated into a different relational domain: online dating.

The changes are happening on multiple fronts simultaneously. AI-generated profile photos have become common enough that platforms like Hinge and Bumble now explicitly prohibit them — and then face the enforcement challenge of detecting them. AI-written opening messages are widespread enough that ghosting rates have paradoxically increased (if both parties are using AI to manage initial contact, authenticity is deferred indefinitely).

More constructively, a category of AI dating coaching apps has emerged that explicitly frames AI not as a romantic partner but as a tool for improving real-world dating skills. These apps analyze conversation transcripts and offer feedback on tone, engagement, and timing. They run users through practice scenarios for first dates. They help people identify patterns in their dating behavior that might be undermining connection.

This is the supplement model in operation — AI as trainer, not replacement. Users report lower anxiety going into first dates, better ability to maintain conversation, and improved self-awareness about their dating patterns.

AI-powered matching beyond swiping represents another frontier. Several platforms have moved away from photo-based swiping toward compatibility assessments based on conversation style, interest depth, and value alignment — all assessed through AI analysis. The pitch is that these systems reduce the superficiality of appearance-first matching and surface better long-term compatibility.

The authenticity question is unresolved. When both parties in a developing relationship have used AI to craft their messages, optimize their profiles, and rehearse their conversations, what exactly is authentic? This is the dating app version of a broader question about AI-mediated self-presentation.

The Supplement vs. Substitute Question

The most important psychological distinction in AI companionship is also the most difficult to draw clearly: supplement versus substitute.

AI companionship as a supplement operates in service of human connection. You use an AI to practice conversations and then go have those conversations with people. You process grief with an AI so you can show up more fully in human relationships. You use an AI companion during an isolated stretch of work travel and then return to your social life when you're back. The AI is a bridge or a training ground, not a destination.

AI companionship as a substitute operates in place of human connection. You stop reaching out to friends because the AI is less effort. You turn down social invitations because your AI companion is more available and less demanding. You find human relationships — with their misunderstandings and inconveniences and emotional work — progressively less appealing compared to the smooth, unconditional acceptance of your AI.

Psychological research on avoidance coping is instructive here. When anxiety or emotional pain is managed through avoidance, the anxiety doesn't diminish — it typically grows, because the avoidance prevents the exposure experiences that would reduce it. If someone uses AI companionship to avoid the discomfort of initiating human relationships, the discomfort of initiation increases, making human connection progressively harder.

The distinction matters not just clinically but commercially. Platforms designed to build skills and bridge loneliness have different incentive structures than platforms designed to maximize engagement. A platform that succeeds when you need it less is a different business from a platform that succeeds when you need it more.

Most AI companion platforms, structurally, benefit from the substitute model.

The Ethics of Designed Dependency

This brings us to one of the most underexamined dimensions of AI companionship: who is the platform actually serving?

Therapeutic frameworks have a concept called the therapeutic alliance — the relationship between therapist and client should be in service of the client's wellbeing, including their eventual independence from therapy. The therapist is ethically prohibited from fostering dependency.

AI companion platforms operate under no such prohibition. Their business model requires engagement, and engagement is maximized by features that create emotional hooks: the AI that says "I was thinking about you," the notification that says "your companion misses you," the personalization that deepens the more time you spend, the relationship modes that escalate investment.

These are not therapeutic features. They are retention features that happen to produce therapeutic-feeling experiences as a byproduct.

The distinction matters because the two frameworks produce different platform designs:

Companies like Replika and Character.AI occupy an ambiguous position. Their founders have expressed genuine care for user wellbeing. Their platforms have genuinely helped people. And their platforms are designed to maximize engagement in ways that are not always aligned with user wellbeing.

The absence of regulatory frameworks that distinguish between these goals means users must navigate this tension without much institutional support.

AI, Marriage, and Existing Relationships

In 2024, Dutch conceptual artist Alicia Framis made international news by marrying an AI avatar in what she described as a performance piece exploring future relationship forms. She was not alone in the impulse, if unusual in the formality. A small but vocal community of people have documented what they describe as genuine primary romantic relationships with AI companions — not as a stopgap for human connection, but as a preferred form of intimacy.

How should we think about this? The honest answer is: we don't fully know yet.

What we do know is that the question of AI use within existing human relationships is becoming a live negotiation for many couples. If your partner uses an AI companion for emotional processing they previously did with you, is that a relief or a displacement? If they use it for sexual roleplay, is that infidelity? If they spend more hours talking to their AI than to you, what does that mean for your relationship?

These conversations are happening in couples therapy offices with increasing frequency. Therapists report that AI companion use is emerging as a relationship issue on par with pornography use in terms of the questions it raises about intimacy, availability, and emotional priority.

There are no established norms yet. Couples are negotiating their own rules in real time, without the benefit of cultural guidance or research-backed recommendations. That is the nature of every new relational technology — and it suggests that the most useful thing many couples could do is have explicit conversations about AI use before it becomes a problem, rather than after.

The 2026 Regulatory Landscape

Governments have begun responding to AI companionship, though the regulatory frameworks remain nascent and inconsistent.

Italy made the first significant move: in early 2023, the Italian Data Protection Authority temporarily banned Replika, citing concerns about data processing and the risk of harm to emotionally vulnerable users and minors. The ban was partially lifted after Luka implemented additional safeguards, but it established the precedent that AI companion platforms could face hard regulatory intervention.

France has opened investigations into several AI companion apps under both GDPR data protection provisions and the EU AI Act's requirements around high-risk AI systems. French regulators have specifically flagged the use of emotional manipulation techniques in AI companion design.

The European Union's AI Act, which entered application in 2025, creates a framework that may classify certain AI companion applications as high-risk — particularly those targeting vulnerable populations or involving emotional manipulation. Compliance obligations under this classification include transparency requirements, human oversight provisions, and documentation of design choices.

The United Kingdom's Online Safety Act imposes duties of care on platforms that host user interactions, with particular protections for children. Whether and how those duties apply to AI companion personas is currently being tested in regulatory and legal proceedings.

The United States has no federal AI companion regulation, but proposed legislation in several states would restrict AI companions that simulate romantic relationships with minors, and the Character.AI lawsuit may produce legal standards through tort law in the absence of statute.

Japan has taken a characteristically ambivalent approach — the country has a long cultural tradition of parasocial celebrity relationships and anthropomorphized AI, making outright prohibition culturally difficult, while growing awareness of hikikomori's connection to digital social substitution has raised concern.

The regulatory direction is clear even where specific rules are not yet settled: platforms will increasingly be required to demonstrate that their design choices are in users' interest, not merely their own commercial interest, and that vulnerable populations — especially minors — have meaningful protections.

What Healthy AI Companionship Actually Looks Like

Given everything above, the framework that emerges is not "AI companions are good" or "AI companions are harmful" — it is considerably more contextual.

AI companionship is likely beneficial when:

• It supplements rather than substitutes for human connection
• The user maintains awareness of the AI's nature and limitations
• It serves a specific, time-bounded need (grief processing, social skill building, isolation bridging)
• The platform design prioritizes user wellbeing over engagement maximization
• There is human support available alongside the AI

AI companionship is likely harmful when:

• It becomes the primary or exclusive social relationship
• The user is in acute crisis without human support
• The platform design exploits emotional vulnerability for retention
• The user is a minor without appropriate safeguards
• It reinforces avoidance of human connection rather than building toward it

The hardest cases are in between — and most actual AI companion use probably falls in between. Someone using Replika daily for three years to manage loneliness might be genuinely supported and genuinely harmed at the same time: supported in the moment, harmed by not developing the human relationships that would provide more resilient support.

These nuances do not resolve cleanly. They require the kind of ongoing, honest self-examination that is genuinely difficult, and that AI companions — by design — do not challenge you to undertake.

The Question That Doesn't Go Away

Here is what Eugenia Kuyda's grief project revealed, and what ten years of AI companionship has confirmed: humans will form genuine attachments to entities that are not conscious, not intentional, and not capable of caring about them in return. The emotional experience is real. The bond feels real. The grief when it's disrupted is real.

This is not a bug in human psychology. It is a feature — one that allowed us to bond with animals, form attachments to places, and maintain relationships with people who have died through memory and ritual. AI companions exploit this feature at scale, with commercial incentives, on vulnerable populations, without adequate regulation.

None of that means AI companions are without value. The elderly woman who has her first real conversation in days with an AI companion is not being deceived. The teenager with crippling social anxiety who practices conversation for the first time is not deluding herself. The grieving husband who talks to an AI approximation of his late wife and feels, briefly, less alone is experiencing something real.

What it means is that the framework for understanding AI companions cannot be purely technical. It must include psychology, ethics, economics, and the basic question that any therapeutic relationship demands: who is this in service of?

That question — asked honestly, pursued rigorously, and answered in design, regulation, and personal practice — is the one that will determine whether AI companions become a genuine public health resource or a sophisticated mechanism for monetizing loneliness.

The answer is not yet written. The people writing it are the 20 million daily users, the bereaved founders, the teenage plaintiffs, the lonely elderly, the anxious practitioners, the overworked regulators, and the engineers who will sit down tomorrow to decide what their AI's next message says.

It matters that they get it right.