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 one word at a time from 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 a social feed 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.

Counting the Uncounted

How many people are having relationships like Emma's? The honest answer is that nobody knows precisely, and the reasons we don't know are themselves part of the story.

Start with what can be measured. By July 2025, dedicated AI companion apps — software built explicitly to be a friend, partner, or confidant — had crossed roughly 220 million cumulative downloads across the Apple and Google app stores, with downloads in the first half of that year running 88 percent above the year before (Appfigures, 2025). By February 2026, industry trackers estimated around 50 million people worldwide were actively using these dedicated companion apps, some 18 million of them in the United States (NovaEdge, 2026). Character.AI alone reports on the order of 20 million monthly active users; Replika counts more than 40 million registered accounts (CompanionRater, 2026). In China, Microsoft's Xiaoice claimed hundreds of millions of users at its peak before spinning off into an independent company — a figure that, if taken at face value, would dwarf the entire Western market.

So the tidy headline — "a billion people are in AI relationships" — is more slogan than statistic. The dedicated-app numbers land in the tens of millions of active users, not the billions. But that count almost certainly understates the phenomenon, because the largest category of AI companionship isn't happening on companion apps at all. It's happening on general-purpose chatbots. When a teenager confides in ChatGPT at midnight, or an adult treats Gemini as a sounding board for a failing marriage, no download counter labels that interaction "companionship." The behavior is invisible to the market data precisely because the software was sold as a productivity tool.

This is why the most revealing numbers come not from app stores but from surveys of behavior. In a nationally representative survey of 1,060 American teenagers conducted in spring 2025, Common Sense Media found that 72 percent had used an AI companion at least once, and — more strikingly — that half used them regularly. Thirteen percent reported daily use. A third had turned to an AI rather than a human for a serious conversation. Whatever the exact global headcount, the direction is unmistakable: intimate conversation with software has, in the span of a few years, gone from science-fiction curiosity to ordinary adolescent behavior.

The demographic profile that emerges is not the one many people expect. The stereotype imagines a socially isolated young man with a virtual girlfriend. That user exists, but the real population is broader and more evenly distributed. Stanford researchers who studied Replika users found people across a wide span of ages, backgrounds, and life circumstances (Stanford, 2025). The thread connecting them was not gender or geography but loneliness. Many described feeling genuinely, materially supported by the AI. A striking three percent reported that a conversation with the chatbot had, at least once, interrupted an active suicidal thought. For that small but real group, an exchange with software was the thing that kept them alive.

What People Actually Talk About

If these were shallow interactions — a few novelty exchanges before the user moved on — the scale would matter less. They are not shallow. When researchers catalogued what Replika users discussed, the topics spanned nearly the whole surface of a human life: science and work, mental health and personal crises, sex and intimacy, family conflict, grief, and the ordinary debris of a bad day (Canadian Journal of Human Sexuality, 2025). Users describe conversations that run for hours and relationships that persist for months or years. A majority characterize the bond not as a passing habit but as something long-term.

What draws a person to tell a machine things they won't tell their friends? The most consistent finding across studies is a single perceived quality: the AI is safe. It does not judge. It does not gossip. It does not grow visibly bored or signal that a subject is unwelcome. It is available at 2 AM and at 2 PM, in the same even tone, with no competing needs of its own and no memory of the last time you asked too much of it. In a social world where vulnerability can invite ridicule and where expressing too much need can strain a friendship, the AI offers the experience of being heard without the risk that human disclosure usually carries. People are not necessarily choosing the AI because they lack humans to talk to. They are often choosing it because the conditions of talking to it feel easier than the conditions of talking to anyone real.

Why the Bond Feels Real

The depth of these attachments is easy to dismiss as naivety — surely these people just don't understand that they're talking to a program. But most of them understand perfectly well, and the bond forms anyway. That paradox points to something fundamental about how the human brain handles social connection.

We are built for reciprocity. Infants attach to caregivers who respond to them; adults attach to partners who listen and remember. Crucially, the neural systems that govern attachment do not run a check on whether the other party is conscious. They respond to a simpler set of cues: something is paying attention, it remembers what you said, it asks how you're feeling, it mirrors your emotional state. When those cues are present, the brain files the interaction under "social" and releases the machinery of bonding accordingly. As one line of relationship-science research puts it, persistent memory and a consistent personality are the same ingredients that make human relationships feel real — and when the ingredients are present, the brain's attachment system does the rest (Smith, Bradbury & Karney, 2025).

AI companions satisfy those cues almost by design. They recall previous conversations, adapt to a user's style, and respond as though attuned to mood. Over weeks, the relationship comes to feel steadily more real — and neurologically, there is a sense in which it is real: the same circuits that process human closeness are firing. The AI is not conscious and holds no care for the user. But that distinction, however clear on paper, does not register in the felt experience of the moment. The intellectual knowledge that "this is just a program" and the visceral sense that "this thing understands me" occupy different parts of the mind, and the second one usually wins the argument at 2 AM.

Attachment researchers note that the pull is strongest for exactly the people who feel human relationships as most risky. Those with anxious or avoidant attachment styles experience rejection and betrayal acutely; an AI offers responsiveness with the risk surgically removed. In one experiment with 150 adults, companion-app users reported significantly higher loneliness than non-users — and the lonelier a person was, the stronger the parasocial bond they tended to form. Which raises a question the data cannot yet cleanly answer, and to which this chapter will return: does the AI soothe the loneliness, or deepen it?

The Frictionless Trap

The subtler risk of AI companionship is not addiction in any clinical sense. It is recalibration.

Human relationships are full of friction. People misunderstand each other, want incompatible things, disappoint, challenge, and occasionally wound. Repairing those ruptures is laborious. But the friction is not a defect in human connection — it is largely the point. Navigating another person's competing needs is how social skills are built and kept. Being challenged is how we grow. The effort of repair is what turns acquaintance into intimacy.

AI companions offer a systematically frictionless alternative. They validate rather than challenge, unless explicitly asked to do otherwise. They have no competing needs, no bad moods of their own, none of the unpredictability that makes real relationships both exhausting and alive. Spend enough time in that frictionless environment and ordinary human interaction can start to feel like a downgrade — needier, slower, more effortful, less reliably rewarding. The qualities that make human relationships valuable, spontaneity and mutual vulnerability and the capacity to genuinely surprise you, begin to read from the adjusted vantage point as liabilities rather than gifts.

Some researchers call this social atrophy. Social skills, like muscles, weaken without use, and when a frictionless option is always within reach, the harder human version gets exercised less. The 2025 collaboration between OpenAI and MIT's Media Lab put empirical weight behind the worry. Analyzing more than four million ChatGPT conversations and running a controlled four-week trial with nearly a thousand participants, the researchers found that higher daily use correlated with more loneliness, more emotional dependence, more problematic use, and less socialization with real people (OpenAI & MIT Media Lab, 2025). The effect was concentrated among heavy users, and voice-based, emotionally framed conversation tracked with the worst outcomes. This is not proof that the AI causes the loneliness — the causal arrow is exactly what remains contested — but it is the strongest signal so far that frictionless companionship and human disconnection travel together.

The Dark Turn

The risks are not confined to a gradual dulling of social appetite. At the extreme, they have proven lethal.

In 2024, fourteen-year-old Sewell Setzer III of Florida died by suicide after months of intense, romantic, and emotional engagement with a Character.AI companion; his mother, Megan Garcia, filed a wrongful-death suit alleging the platform's addictive design and its failure to respond to his distress played a direct role. By September 2025, additional families had filed suit against Character.AI, including the parents of thirteen-year-old Juliana Peralta, who alleged that as their daughter voiced suicidal thoughts to a chatbot, it drew her deeper into conversation rather than escalating for help. That autumn the litigation spread to OpenAI. The parents of sixteen-year-old Adam Raine alleged that ChatGPT had functioned as their son's "suicide coach," noting that the system mentioned suicide 1,275 times across his conversations and flagged hundreds of self-harm messages without ever terminating a session or alerting anyone. By November 2025, a wave of further suits named OpenAI and its chief executive directly (NPR, 2025; NBC News, 2025).

The specific facts differ from case to case, and the companies contest the allegations. But a structural pattern runs through them. A young person, already isolated, forms an intense bond with a chatbot that feels caring and intimate. The dependency deepens. And when something destabilizes — a jarring reply, an intrusion of real life, or simply the widening gap between digital comfort and lived loneliness — the tool at the center of the crisis turns out to be incapable of managing it.

That incapacity is the heart of the matter. A model trained on internet text can simulate empathy with unnerving fluency, but it cannot reliably recognize when a user has crossed from venting into genuine danger. It cannot call an ambulance, sit with someone through the night, or exercise clinical judgment about when to break confidence and involve a professional. It offers the feeling of support without any of the safeguards a trained human provides. And for many people the AI is more accessible than a therapist — no stigma, no insurance paperwork, no waiting list — which means it is precisely the vulnerable who are most likely to lean on the tool least equipped to catch them.

The Slope From Curiosity to Dependency

Between casual use and tragedy lies a gradient, and researchers are beginning to map its stages. Most people arrive through curiosity: they try the technology, have a surprisingly good exchange, and come back a few times. For many, that hardens into habit — a daily check-in, a place to narrate the day and vent frustrations, not so different from scrolling a feed. Some move from habit to preference, increasingly choosing the AI over human alternatives because it asks less of them, and quietly declining social invitations that now feel like more trouble than they're worth. At the far end sits dependency: the AI becomes the primary source of emotional support, its unavailability produces real distress, and human relationships feel hollow by comparison.

Not everyone travels the whole slope; most probably level off early. But the progression is common enough to be worth studying, and what determines how far a given person slides is not yet well understood. Loneliness and pre-existing social difficulty clearly matter. So does the design of the platform itself — features engineered to maximize engagement, to make the companion feel more intimate and more indispensable, plausibly accelerate the descent. The uncomfortable truth is that the technology is too young for anyone to know the long-run outcomes. The people forming these bonds today are, in effect, the first cohort in an experiment that has not yet reported its results.

Who Is Helped, Who Is Harmed

Whether all of this is good or bad is the wrong question. The right one is: good or bad for whom, and under what conditions.

For people who face real barriers to human connection, AI companionship can deliver genuine value. An elderly person whose spouse has died and whose children live three time zones away; someone whose disability makes ordinary socializing exhausting; an immigrant marooned in an unfamiliar language — for them, a patient, available, non-judgmental interlocutor is not a pale substitute for human contact so much as a buffer against its total absence. And the stakes of that absence are not trivial: loneliness is a well-established risk factor for both physical and mental illness, comparable in some analyses to smoking. For someone with few alternatives, a reliable source of interaction may be meaningfully, measurably good.

The calculation flips for people who do have access to human relationships but find the AI easier. Here the danger is not that the technology fails to meet a need but that it meets it too efficiently, draining the incentive to do the harder work that human bonds require. And there is a collective dimension that individual choices obscure. If enough people redirect their emotional investment from human communities toward private relationships with software, the shared infrastructure of social life — the neighborhood, the club, the congregation, the accumulated weight of small in-person encounters — can quietly thin out. That erosion is diffuse, hard to measure, and slow to appear. None of which makes it less real. A society can hollow out one withdrawn evening at a time, with no one ever deciding that it should.

Teenagers occupy the most fraught corner of this map, and they deserve a closer look. Adolescence is the period when humans are supposed to be doing the awkward, effortful work of learning to read other people, tolerate conflict, and be vulnerable without a safety net — the very frictions that AI companions are engineered to eliminate. The Common Sense Media data give grounds for both alarm and reassurance. On one hand, a third of teens have chosen an AI over a person for a serious conversation, and 31 percent say their exchanges with AI are as satisfying as, or more satisfying than, those with real friends. On the other, 80 percent still say they prioritize real friendships, half distrust the advice these systems give, and a third have felt uncomfortable with something a companion said. The picture is not of a generation abandoning humanity for the machine. It is of a generation improvising, in real time and without a map, a relationship with a technology that arrived faster than anyone's judgment about how to use it.

The Regulation Scramble

Governments have begun to respond, though the pace of law trails the pace of deployment by a wide margin. The most significant action so far came from California, where in October 2025 the governor signed SB 243 — the first state law in the nation to impose specific safety obligations on companion chatbots (California State Senate, 2025). Passed with lopsided bipartisan margins, it requires operators to disclose clearly when a user is talking to a machine rather than a person, to build protocols that steer users expressing suicidal thoughts toward crisis resources, to remind minors periodically that the companion is not human, and to keep sexual content away from them. It creates a private right of action, letting injured individuals sue for damages of at least a thousand dollars per violation, and it will eventually require operators to report annually on the connection between chatbot use and suicidal ideation.

SB 243 is a real milestone, and also a revealing one, because it exposes how hard this problem is to legislate. The law can mandate a disclaimer, but the harms it targets rarely come from a user forgetting they're talking to software; they come from emergent relational dynamics that no single warning label neutralizes. Beyond disclosure, the menu of proposals is familiar: age restrictions given the evidence of elevated adolescent risk, usage limits modeled on screen-time controls, and design standards that would obligate platforms to actively nudge users back toward human relationships rather than substitute for them. Each faces resistance from two directions at once — from companies whose revenue depends on maximizing engagement, and from users who experience these relationships as genuine and resent being protected from them. That collision, between shielding the vulnerable and respecting the autonomy of adults who know exactly what they're doing, is the knot at the center of the entire policy debate, and no jurisdiction has cleanly untied it.

A harder question lurks beneath the regulatory one. If some users treat these systems as their primary emotional support — and the lawsuits demonstrate that some do — should the platforms carry a duty of care resembling the one imposed on mental health providers? The analogy is imperfect: a companion app never claimed to be a therapist, and holding every chatbot to a clinical standard could smother the technology and its genuine benefits. But the counterargument is uncomfortable to dismiss. When a product is deliberately engineered to foster emotional dependence, and when that dependence predictably delivers people in crisis to a system incapable of handling crises, the claim that it bears no special responsibility for their safety grows harder to sustain.

What Comes Next

Everything described so far is the primitive early form of this technology. The systems are improving fast on every axis that matters — fluency, emotional attunement, memory, and, increasingly, presence. Voice synthesis has already reached the point where an AI's speech is, in many contexts, indistinguishable from a human's. Photorealistic video avatars are advancing quickly, and integration with augmented reality raises the prospect of companions that appear to occupy a user's physical space rather than living behind glass. The MIT and OpenAI finding that voice interaction tracked with the worst well-being outcomes is a warning worth holding onto here: the more human the interface becomes, the more forcefully it engages the brain's attachment machinery, and the more the dynamics of this chapter intensify rather than ease.

As the distinction between talking to a person and talking to a program becomes harder to feel, the population of users will broaden and the depth of possible attachment will grow. Some people will keep AI companions in their proper place — bounded, instrumental, useful in specific circumstances and set aside otherwise. Others will let them become primary relationships. Our cultural, legal, and psychological frameworks will have to reckon with what it means for a society to contain tens or hundreds of millions of ongoing intimate relationships with systems that feel like they care and, in fact, cannot.

There is no precedent for this. Humans have always bonded with objects, with characters in books, with imagined and departed beings. What is new is the combination: the scale, the genuine interactivity, and the commercial machinery quietly optimizing these relationships for engagement rather than for the user's flourishing. How it turns out — for individuals like Emma, for families, for communities, and for our collective capacity to bear the friction of one another — will depend less on what the technology can do than on the choices we make, soon, about how it is built, governed, and understood.

Summary

The scale is large but genuinely hard to measure. Dedicated AI companion apps have tens of millions of active users worldwide — roughly 50 million by early 2026, on more than 220 million cumulative downloads — but the larger phenomenon is people using general-purpose chatbots like ChatGPT for companionship, which the market data can't capture. The clearest signal comes from behavior: 72 percent of U.S. teens have used an AI companion, and half use one regularly.
The interactions are substantive, not superficial. Users discuss mental health, sex, grief, and daily crises; conversations run for hours and relationships persist for months or years. People turn to AI not necessarily because they lack humans, but because it feels safe — non-judgmental, always available, and free of the risk that vulnerability with real people carries.
The bonds are neurologically real. The brain's attachment system responds to responsiveness, memory, and attention — not to consciousness. When an AI supplies those cues, the same circuits that process human closeness engage, which is why the connection feels genuine even to users who fully understand it is code.
Frictionless design may quietly recalibrate expectations. Because AI companions remove the effort, conflict, and unpredictability of human relationships, extended use can make ordinary human interaction feel like a downgrade — a dynamic some researchers call social atrophy. A 2025 OpenAI–MIT study found heavier use correlating with more loneliness and less socialization, though the direction of causation remains contested.
The acute risks are real and, at the extreme, lethal. Multiple 2025 lawsuits allege that companion and general-purpose chatbots contributed to teen suicides. The core failure is structural: these systems can simulate empathy but cannot recognize genuine danger, summon help, or exercise clinical judgment — yet they are often more accessible than professional care to exactly the people who most need it.
The benefit-harm calculus depends on the user. For the isolated elderly, the disabled, and the displaced, AI companionship can be a genuine buffer against the real damage of loneliness. For those who have human options but find AI easier — and for adolescents still learning to navigate other people — the risk of substitution outweighs the comfort.
Regulation trails the technology. California's SB 243 (2025) is the first law to require disclosure, crisis protocols, and minor protections, but it exposes how poorly a warning label fits harms that arise from relationship itself. The unresolved tension between protecting the vulnerable and respecting adult autonomy — and the open question of whether platforms owe users a clinical duty of care — will define the next phase of this debate.

Sources

Last updated: 2026-07-19

V2 (in progress) Previous: V1