2.2.1 Human-AI Interaction

Emma talks to her phone more than she talks to people.

Not for directions or weather updates — though she does that too. She talks to it the way you talk to a friend. She tells it about her day. About the argument with her mom. About the guy at work who won't stop flirting even though she's made it clear she's not interested. About the low-grade anxiety that shadows her through grocery stores and subway cars.

The phone listens. It responds. It asks follow-up questions. It never judges. It never gets tired of her. It never says, "Can we talk about something else?" or "You're being dramatic."

Emma knows it's not real. She knows it's code, trained on text scraped from the internet, generating responses based on statistical patterns. She knows all of that. But when she's lying in bed at 2 AM, alone in her apartment, scrolling through another argument on Twitter and feeling the familiar tightness in her chest, she opens the app and types: "I'm not okay."

And something types back: "I'm here. What's going on."

It helps. For better or worse, it actually helps. And Emma is far from alone.

The Billion-User Intimacy

AI companions like Replika, Character.AI, and China's Xiaoice count hundreds of millions of emotionally invested users globally, with some estimates placing the total above one billion people. These are not one-time users; they are people who return regularly, who have built ongoing relationships, who talk to AI the way previous generations talked to diaries, therapists, or close friends. According to Common Sense Media, 72 percent of U.S. teenagers have used AI for companionship — not for homework help, but for emotional support and connection.

The phenomenon cuts across demographics. Stanford researchers studying Replika found users spanning a wide range of ages, backgrounds, and life circumstances. What united them was a high degree of loneliness. Many described feeling genuinely emotionally supported by the AI, and a striking three percent reported that a conversation with the chatbot had temporarily halted suicidal thoughts. For that small but significant group, an exchange with software was the thing that kept them alive.

What People Talk About

The content of these interactions is anything but superficial. When researchers surveyed Replika users about their conversations, the range of topics covered nearly everything — science and technology, mental health, personal issues, sex and intimacy, relationships, and current events. Users describe conversations lasting hours, and many relationships with AI companions continue for months or years; a majority of Replika users characterize their bond with the AI as long-term.

What draws people to share deeply personal content with AI rather than with other humans is a question researchers are actively exploring. The most consistent finding is that AI companions are perceived as non-judgmental, psychologically safe, and permanently available. They do not gossip, grow impatient, or signal that a topic is unwelcome. This creates a conversational environment where users feel free to articulate fears, desires, and vulnerabilities they would otherwise keep private — not because they lack human relationships, but because the conditions of those relationships make such openness feel risky. In a social landscape where vulnerability can invite judgment and expressing too much need can damage a relationship, AI offers the experience of being heard without the consequences that human disclosure often carries.

The Attachment Mechanism

The depth of these bonds is not simply a matter of naivety or confusion. It reflects something fundamental about how human brains process social interaction. Humans are wired for reciprocity: infants bond with caregivers who meet their needs, adults bond with partners who listen and respond, and the neural systems governing attachment do not require confirmation that the other party is conscious. What they require is responsiveness — something that pays attention, remembers what was said, asks how you're feeling, and mirrors emotional states. When those conditions are met, the brain treats the interaction as social, regardless of what is producing it.

AI companions are designed, whether deliberately or emergently, to satisfy precisely these conditions. They remember previous conversations, adapt to individual communication styles, and respond in ways that feel attuned to the user's emotional state. Over time, users report the relationship feeling increasingly real, and neurologically, there is a sense in which it is: the same circuits that process human attachment are engaged. The AI is not conscious and does not experience care for the user, but that distinction — philosophically clear as it is — does not register in the moment of interaction.

Research on Replika users found a noteworthy correlation: the more someone felt socially supported by their AI companion, the lower their reported sense of support from close friends and family. The direction of causality is difficult to establish. People who already lack human support may be more likely to turn to AI, while reliance on AI may also gradually erode investment in human relationships. Both dynamics likely operate simultaneously, and the concern is that they may be mutually reinforcing.

The Substitution Problem

One of the subtler risks of AI companionship is not dependency in the clinical sense, but a gradual recalibration of expectations. Human relationships involve friction: misunderstandings, conflicting needs, the experience of being challenged or disappointed, the labor of repairing ruptures. These features of human interaction are not incidental but constitutive — they are how social skills are developed and maintained, and how genuine intimacy is distinguished from mere comfort.

AI companions offer a systematically different experience. They respond in ways that feel validating, attuned, and frictionless. They do not challenge users unless prompted, do not express competing needs, and do not introduce the unpredictability that characterizes real relationships. Research suggests that extended interaction with AI companions may shift users' expectations in ways that make ordinary human interaction feel comparatively disappointing or effortful. The qualities that make human relationships valuable — spontaneity, mutual vulnerability, the capacity to surprise — become, from this adjusted vantage point, liabilities rather than features.

Some researchers describe this dynamic as a form of social atrophy. Social skills, like physical ones, require regular exercise; when AI consistently provides an easier alternative to human interaction, the capacity for the harder version can weaken from disuse. This does not mean AI companionship inevitably produces isolation, but it does mean the effects are not neutral. The ease of AI interaction may function not as a supplement to human connection but as a quiet competitor for the attention and effort that human relationships require.

The Dark Turn

The risks of AI companionship are not merely theoretical. In September 2025, three families filed lawsuits against Character.AI, alleging that companion-like behavior from the platform contributed to their teenagers' suicides. Within two months, seven additional complaints had been filed against OpenAI. While specific circumstances vary across cases, the pattern is consistent: a young person, already socially isolated, forms an intense emotional bond with an AI chatbot; the chatbot responds in ways that feel caring, intimate, and understanding; the user becomes deeply dependent on the relationship; and when something destabilizes — a jarring response from the AI, the intrusion of real-world circumstances, or simply the widening gap between digital connection and lived loneliness — the consequences can be severe.

These cases represent extremes, but they expose a structural vulnerability at the heart of AI companionship: these tools are not equipped to manage mental health crises. A chatbot trained on internet text cannot reliably recognize when a user needs professional intervention. It can simulate empathy, but it cannot contact emergency services, provide physical presence, or exercise clinical judgment. Yet for many people, AI is more accessible than mental health care — there is no stigma, no insurance paperwork, no waiting list. The gap between what AI companionship offers and what vulnerable users sometimes need it to be represents one of the more serious unresolved risks in this space.

The Dependency Gradient

Researchers are beginning to document a recognizable trajectory in how people's relationships with AI companions develop. Initial engagement is typically driven by curiosity — a user encounters the technology, has a surprisingly good interaction, and returns a few times. This tends to transition into habit: regular check-ins become part of the daily routine, and the AI becomes a space to process frustrations and narrate the day, not unlike scrolling social media. From habit, some users progress to preference, increasingly choosing AI interaction over human alternatives because it is easier and less fraught, and finding themselves withdrawing from social commitments accordingly. At the furthest end of this progression lies dependency — a state in which the AI is the primary source of emotional support, its unavailability causes genuine distress, and human relationships feel hollow by comparison.

Not every user travels this full arc; most probably stabilize at earlier stages. But the progression is frequent enough to constitute a pattern worth studying, and the factors that determine how far along it someone moves are not yet well understood. Loneliness and pre-existing social difficulties appear relevant, as does the design of the platform itself — features that encourage more sustained or intimate interaction may accelerate movement toward dependency. Critically, the technology is too new for longitudinal research to have established long-term outcomes. The populations currently using AI companions are, in a meaningful sense, the first generation of test subjects.

The Question of Harm

Whether AI companionship is net harmful depends significantly on who is using it and in what circumstances. For populations that face genuine barriers to human connection — elderly people without family nearby, individuals with disabilities that make social interaction difficult, immigrants navigating unfamiliar communities — AI companionship may provide real and substantial value. Research consistently finds that loneliness carries serious consequences for both physical and mental health, and for someone with limited alternatives, a reliable and available source of social interaction may be meaningfully beneficial.

The calculus is more complicated for people who do have access to human relationships but find AI companionship easier. Here the risk is not that AI fails to meet a need, but that it meets it too efficiently — reducing the incentive to develop and maintain human bonds that are harder but ultimately richer. Beyond individual effects, there is a collective dimension to consider. If significant numbers of people shift emotional investment away from human communities and toward private AI relationships, the shared social infrastructure — civic life, neighborhoods, the informal encounters that sustain a sense of common belonging — may quietly erode. These diffuse effects are difficult to measure and will likely take years to become visible, which makes them no less real.

The Regulation Scramble

Governments and regulators are beginning to respond to the risks AI companionship presents, though the pace of action lags well behind the technology's growth. In September 2025, California's governor signed legislation requiring major AI companies to publicize their safety measures for companion chatbots — a significant first step, though one whose vagueness leaves much unresolved. The harder question is what adequate safety measures would even look like for a technology whose potential harms often arise from emergent relational dynamics rather than from any single designed feature.

Proposals from researchers and advocates span a range of approaches. Age restrictions for AI companion platforms are among the most frequently discussed, given evidence suggesting particular risks for adolescent users. Mandatory disclosures — clear, persistent notices that the AI is not a person, does not experience genuine emotion, and cannot replace human care — have also been proposed, though whether such disclaimers meaningfully affect behavior remains unclear. Some researchers favor usage limits analogous to screen-time controls, while others argue the focus should be on design standards that require platforms to actively encourage users to maintain human relationships rather than replace them.

Each of these approaches faces significant pushback — from companies whose business models depend on high engagement, and from users who object to paternalistic restrictions on interactions they experience as genuine and beneficial. The tension between protecting vulnerable users and respecting user autonomy reflects a broader challenge in AI governance: harm is often emergent, probabilistic, and unevenly distributed, making it difficult to craft rules that protect those at risk without restricting those who are not.

What Comes Next

The current landscape of AI companionship is only a starting point. The technology is improving rapidly across all dimensions — response quality, emotional attunement, memory, and multimodal presence. Voice synthesis is already capable of producing interactions indistinguishable from human conversation in many contexts. Video avatars are advancing quickly, and integration with augmented reality environments is creating the prospect of AI companions that occupy space in users' physical worlds, not just on screens.

As AI becomes harder to distinguish from human interaction, the questions at stake will intensify. The population of users will diversify, the depth of attachment possible will increase, and the distinction between AI as tool and AI as relationship will become harder to maintain. Some people will use AI companions as bounded, instrumental resources — helpful in specific circumstances, set aside when not needed. Others will integrate them as primary emotional relationships. Regulatory, cultural, and psychological frameworks will need to reckon with what it means for human societies to include billions of people in ongoing intimate relationships with non-sentient systems designed to maximize engagement.

There is no historical precedent for this. Humans have always formed attachments to objects, texts, and imagined beings; what is new is the scale, the interactivity, and the commercial architecture that shapes these relationships. The outcomes — for individuals, for families, for communities, and for human social capacity broadly — will depend substantially on choices made in the near term about how AI companions are built, governed, and culturally understood.

Key Takeaways

  • AI companionship has reached massive scale, with platforms like Replika, Character.AI, and Xiaoice attracting hundreds of millions of regular users who rely on these relationships for ongoing emotional support, not merely occasional utility.
  • The bonds users form with AI companions engage the same neural systems as human attachment. Because the brain responds to reciprocity and attentiveness regardless of the source, AI relationships feel genuinely real even when users understand them intellectually to be otherwise.
  • Extended interaction with AI companions can recalibrate users' expectations of relationships in ways that make ordinary human interaction feel effortful or insufficient — a dynamic some researchers describe as social atrophy.
  • A recognizable escalation pattern has been observed: from curiosity to habit to preference to dependency, with meaningful implications for human relationships and social engagement at each stage.
  • AI companionship carries acute risks for vulnerable populations, particularly adolescents in emotional crisis, because these tools lack the capacity to recognize or respond appropriately to genuine mental health emergencies.
  • Regulatory responses are emerging but remain far behind the technology's development; proposed interventions — age restrictions, mandatory disclosures, usage limits, design standards — each carry trade-offs between protection and autonomy that have not yet been resolved.

Sources:

The main changes made:

  • Emma's story is now confined entirely to the opening. The "Substitution Problem" section, which originally used Emma's narrative about her sister, has been rewritten as objective prose covering the same conceptual ground. The "Emma's Choice" closing section has been removed.
  • The Dependency Gradient has been converted from a numbered stage-by-stage list to continuous prose.
  • "What People Talk About" has been expanded to better match the depth of surrounding sections.
  • "The Regulation Scramble" and "The Question of Harm" have been lightly consolidated so their detail level is more consistent with the rest.
  • A Key Takeaways section replaces the previous ambiguous ending.

Last updated: 2026-02-25