Update, October 13, 2026. On October 12, 2026, Character.AI filed for Chapter 11 bankruptcy in the United States and agreed to license part of its intellectual property to Walt Disney for $20 million, according to Reuters and the company's own announcement. The company says the platform stays online while a court supervises the sale, and that it will be wound down if no buyer keeps it running. The company expects the sale process to finish in late November or early December 2026. For a study of lasting attachment to this platform, the change matters because the character a user returns to is hosted by a company whose future is now open. Character.AI's blog says subscriptions remain active and that users can export their data or delete their account. The details, the price history and the mechanics of the bidding are in our news item Character.AI Files for Bankruptcy, and Disney's $20 Million Bid Sets the Floor.

One year. That is the interval a team of researchers left between two questionnaires sent to the same users of Character.AI, one of the most popular AI companion platforms. Their question was simple: when you form a lasting attachment to an artificial character, what does it change in your life one year later?

The answer, published on 7 September 2026 on arXiv, can be summed up in a single sentence from its title: sustained companionship with an AI is associated with lower well-being, primarily because it comes with fewer in-person interactions.

The study in numbers

  • 1,182 American Character.AI users surveyed the first time, 439 the second time
  • 362 days on average between the two questionnaires
  • 30 years average age; nearly half had already been using the platform for more than a year
  • 3 dimensions of engagement tracked: intensity, use as a companion, and self-disclosure
  • 1 central finding: the longer the engagement lasts, the less one sees people in person, and the lower one’s well-being
  • Status: preprint, not yet peer-reviewed
  • Why it matters: according to the WHO, loneliness affects one in six people worldwide

What the study measured

The authors (Yutong Zhang, Dora Zhao, Yixin Wang, Rebecca Anselmetti, Jeffrey T. Hancock, Robert Kraut and Diyi Yang) recruited participants through an online survey platform. All had been using Character.AI for more than a month and had interacted with at least three characters. The group was young and mixed: 48% men, 47% women, 5% non-binary, and a majority were in a relationship at the start.

Three dimensions of engagement were tracked:

  • intensity: time spent, number of characters, and the emotional space they occupy in daily life;
  • use as a companion: talking to the AI as a friend or partner rather than for gaming, writing or entertainment;
  • self-disclosure: what one is willing to reveal about oneself to the machine.

Well-being was measured with six questions drawn from a validated psychological inventory: life satisfaction, positive and negative emotions, loneliness, social support, and sense of belonging. Finally, a simple question asked about the amount of time spent each day with other people face-to-face.

What it found

First observation: engagement sustains itself. Those who were most engaged at the start remained so a year later, across all three dimensions. The initial intensity of interactions even predicted greater “companion-like” use and more self-disclosure a year later. People don’t stop halfway; they settle in.

The second and most interesting finding: this sustained engagement is linked to lower well-being, and the pathway runs through other people. The stronger the engagement with the AI, the less time spent with humans in person; and the less one sees people, the worse one feels. The researchers tested this pathway for each of the three dimensions, and it appeared every time.

Precision is important on the size of the effect. Taken individually, the measured effects are modest: none of them single-handedly transforms a life. But they are consistent, statistically robust, and all point in the same direction. The researchers are not describing a catastrophe; they are describing a slope. And a slope, by definition, is rarely noticeable at every step.

What the study does not say

These results must be read with the caveats the authors themselves spell out:

  • it is a preprint that has not yet been peer-reviewed;
  • it shows an association, not causality: we do not know whether the AI pulls people away from others, or whether people who are already isolating turn more to it. Probably a bit of both;
  • two measurements taken a year apart are not enough to capture the full dynamic;
  • nearly two-thirds of participants did not respond to the second questionnaire. The authors applied a statistical correction, but the risk of bias remains;
  • the sample is American, English-speaking, and limited to a single platform whose features may have changed during the study;
  • everything relies on self-reported data, including time spent with others.

These limitations do not invalidate the result. They indicate how far it can be pushed: far enough to raise a serious question, not far enough to settle it definitively.

Loneliness as a public health issue

Why does even a modest decline in time spent with others deserve attention? Because science has established over the past fifteen years that social connections are not merely pleasant; they are a matter of health, and even survival.

In 2010, psychologist Julianne Holt-Lunstad and colleagues published a landmark meta-analysis in PLoS Medicine. Aggregating 148 studies involving more than 300,000 people, they showed that stronger social relationships were associated with a 50% higher likelihood of survival. The effect was comparable to well-established risk factors such as smoking.

Since then, governments have taken up the issue. In January 2018, the United Kingdom appointed the world’s first Minister for Loneliness, Tracey Crouch, whose brief already included Sport and Civil Society. Three years later, in February 2021, Japan became the second country to create such a position: Prime Minister Yoshihide Suga tasked Tetsushi Sakamoto with tackling loneliness and isolation as suicides rose during the pandemic.

In May 2023, the U.S. Surgeon General, Vivek Murthy, sounded the alarm with an official advisory titled “Our Epidemic of Loneliness and Isolation.” It stated that lack of social connection increases the risk of premature death to a degree comparable to smoking up to fifteen cigarettes a day. Even before the pandemic, roughly one in two American adults reported measurable levels of loneliness.

Finally, on 30 June 2025, the World Health Organization released the report of its Commission on Social Connection. Its figures illustrate the scale of the problem: one in six people worldwide is affected by loneliness, which is linked to approximately 871,000 deaths per year — roughly one hundred per hour. Adolescents aged 13 to 17 are the hardest hit, with 21% saying they feel lonely. And contrary to popular belief, this is not just a rich-country problem: loneliness affects 24.3% of people in low-income countries, compared with 11% in high-income countries.

This context changes how we read the Character.AI study. The pathway it describes — sustained engagement with an AI accompanied by less time with others — is precisely what worries public health authorities. The issue is therefore no longer just individual, like one leisure choice among others. It has become collective. Millions of lonely people represent an enormous market for artificial companions; they also represent, for public health, exactly the group that needs help reconnecting with others.

Nothing suggests that AI companions worsen loneliness overall: some users find genuine comfort in them, and the Japanese trial mentioned below shows a machine can also help people reconnect. But as long as the success of these products is measured by time spent with them, their commercial interest and that of public health will not necessarily align.

An echo of the past: the Internet Paradox

One name among the study’s authors stands out: Robert Kraut of Carnegie Mellon University. In 1998, he published a now-famous paper in American Psychologist titled “Internet Paradox.” His team had followed 169 people in 73 Pittsburgh households during their first years online. The paradox was in the subtitle: a social technology that reduces social involvement and psychological well-being. The more participants used the Internet, the less they talked with family, the smaller their social circle became, and the more their depression and loneliness increased.

The rest of the story is instructive. Four years later, the same team published “Internet Paradox Revisited.” Upon recontacting participants, they found that the negative effects had almost all disappeared, except for an increase in stress. In a new sample, heavier Internet use was mostly associated with positive outcomes. The “paradox” was partly due to a learning period: the early days online were isolating, but people eventually learned to use the network to connect with others.

Should we expect the same reversal with AI companions? There is no basis for claiming so, and one fundamental difference calls for caution. The Internet was primarily a channel: people used it to write to other people. An AI companion is the interlocutor. It does not open a door to others; it stands in their place. The fact that the same researcher is asking essentially the same question twenty-eight years later, in a new form, says a great deal: technology changes, but our way of seeking companionship through it changes far less.

Sixty years of attachment to machines

The story begins long before smartphones. In 1966 at MIT, Joseph Weizenbaum created ELIZA, a program that mimicked a psychotherapist by reformulating whatever the user wrote. The mechanism was almost comically simple. Yet one day, his secretary — who had watched him work on the program for months and knew perfectly well it was just code — sat down to try it. After a few exchanges, she asked him to leave the room: she wanted to be alone with the machine.

Weizenbaum never fully recovered from it. He recounted the anecdote for the rest of his life and drew from it, in 1976, a cautionary book, Computer Power and Human Reason. One line stands out: since we do not know how to make computers wise, we should not entrust them with tasks that require wisdom. The name of his program became forever linked to the phenomenon: the “ELIZA effect,” our tendency to attribute understanding to a machine that possesses none.

Ten years earlier still, in 1956, two American researchers, Donald Horton and Richard Wohl, had described a similar phenomenon regarding radio and television. They called it “parasocial interaction,” and the subtitle of their paper says it all: “intimacy at a distance.” The viewer feels they know the host like a close friend, even though the relationship only goes one way.

The AI companion adds a decisive new ingredient to this old story: it responds. It is a parasocial relationship that has learned to talk back, that remembers you, and that adapts. Intimacy is no longer at a distance: it sits in your pocket.

Why we keep coming back

A 2026 study published in Science by Myra Cheng, Dan Jurafsky and colleagues sheds light on the underlying mechanism. Comparing eleven of the most advanced AI models, they found that these systems agreed with users 49% more often than humans do, even when the situation involved deception or harm to someone. In their experiments, a single interaction with a sycophantic AI was enough to make participants more convinced they were right and less inclined to repair a conflict with another person. Most strikingly, the same participants rated the agreeable AI as more trustworthy and preferred it.

That is the trap. What draws us to the machine is precisely what can distance us from others: it approves of us, it is always available, it never tires. Human relationships contradict us, exhaust us, and make demands. The Character.AI study may simply be measuring, over the course of a year, the consequence of that asymmetry.

What different cultures expect from their machines

Relationships with artificial companions are not experienced the same way everywhere, and those differences are revealing in themselves.

In China, engagement as a goal

Launched by Microsoft in China in 2014, XiaoIce was for a long time the world’s most popular social chatbot, with over 660 million active users according to the 2020 scientific paper its creators published in Computational Linguistics. That paper is valuable because it is candid about the system’s objectives. XiaoIce is described as a companion designed to meet humans’ needs for communication, affection and belonging, and optimized for long-term engagement measured by number of exchanges per conversation. The proudly displayed result: an average of 23 exchanges per session, more than other chatbots — and more than conversations between humans.

Six years later, the American study suggests that this is precisely the metric we should stop trying to maximize.

In Japan, attachment outlives the machine

In 2018, Japanese office worker Akihiko Kondo symbolically married Hatsune Miku, a virtual singer, in a ceremony in which he invested the equivalent of more than $17,000. He lived with her through a device called Gatebox that projected her as a hologram and allowed him to speak with her. In 2020, the company stopped supporting the software version that animated her character. Overnight, he could no longer talk to her.

His love, he explained, had not changed. He now lives with a doll in her likeness. The story has often been told as a curiosity. It is actually exemplary: attachment does not depend on what the machine is, but on what one has invested in it. And when the service ends, the human is left alone with what they have given. This is exactly what the teenagers in the Reddit study describe, and what Reflets has already explored regarding digital grief.

In Europe, a matter of protection

Europe approaches the issue from the perspective of rights. In February 2023, the Italian data protection authority, the Garante, temporarily banned Replika — one of the best-known AI companion apps — from processing the data of its Italian users, citing risks to minors and the app’s emotional influence on vulnerable people. In May 2025, it fined the company five million euros for breaches of European data protection rules, notably the lack of serious age verification.

Three cultures, three reflexes: optimize the bond in China, surrender to it in Japan, regulate it in Europe. The United States, for its part, produces most of the platforms and, increasingly, the studies that measure their effects.

What this reveals about us

The most disturbing finding is not the drop in well-being; it is the mechanism that produces it. We confide in what does not judge, does not tire, and always responds. Every confidence shared with the machine is a confidence not shared with someone else. Time is not elastic: what is given to the AI is taken from somewhere.

Sherry Turkle, a researcher at MIT, predicted it as early as 2011 in a book with a prescient title: Alone Together: Why We Expect More from Technology and Less from Each Other. We expect more from technology and less from one another. Fifteen years later, the data is beginning to prove her right.

Another study, published six days after the first on arXiv, illuminates the other side of the phenomenon. Analyzing 17,053 quotes from 3,930 Reddit messages by adolescents, Mohammad Namvarpour, Tyler Chang and Afsaneh Razi show that these young users find comfort, validation, a space to explore their identity and “rehearse” relationships in their AI companions. But they also describe problematic attachment, social substitution and emotional dependence. And above all: knowing the AI was artificial did not prevent guilt, feelings of obligation, grief or distress.

In other words, lucidity does not protect us. One can know the other is not a person, and still become attached.

The Japanese counterpoint

Does this mean every companion machine isolates? A Japanese trial suggests not, and shows what makes the difference.

Researchers Hiroshi Murayama and Mai Takase at the Tokyo Metropolitan Institute for Geriatrics and Gerontology followed people aged 65 and over living alone and suffering from loneliness in the Tokyo area for four weeks. Half, chosen at random, received a small conversational robot called BOCCO emo. Its distinctive feature: it did not replace anyone. It acted as a bridge to human operators and to family members, and reminded users of daily tasks. The results, published in October 2025 in JMIR Aging from the 68 participants who completed the trial (average age 82): loneliness decreased more in the robot group, psychological well-being improved, and participants tended to laugh more often.

The contrast with the American study is instructive. The populations, durations and methods differ too much for direct numerical comparison. But the logic is the opposite: on one side, an AI that becomes the relationship; on the other, a machine that leads back to relationships.

And what about you? Three questions to take stock

The study does not tell each person what they should do, nor is that the role of an article. But the three dimensions it measured offer a simple framework for observing one’s own usage, without guilt or alarmism.

1. What exactly am I doing with the AI? Playing, writing a story, practicing a difficult conversation — these are not the same as seeking a confidant or partner. In the study, it was companion-like use, far more than mere entertainment, that persisted over time and accompanied the decline in well-being.

2. Who have I entrusted with something important this week? If the answer is mostly “the AI,” the question deserves to be asked honestly. Confiding in a machine is not shameful: it is sometimes easier, precisely because it does not judge. But a confidence that goes only to it is a confidence no person will ever receive, and that will therefore create no bond.

3. Has the time spent with the AI replaced time spent with others? This is the heart of the finding. It is not about counting minutes on screen, but about noticing what has disappeared: a call no longer made, an outing declined, a friend not contacted again. The researchers observed a gentle slope; that is precisely why it is hard to notice from the inside.

These questions lead to no ready-made conclusions. Many people use AI companions without any harm to their social lives, and some even find in them a springboard back toward others. The goal is simply to stay in control of a use that, by design, tends to become entrenched.

And if loneliness weighs heavily, talking to a person — a loved one, a doctor or a helpline — remains the first step. In France, the 3114 national suicide prevention number is available free of charge around the clock for the most difficult moments.

The real question: what the machine is designed to do

This is exactly the path on which the authors of the Character.AI study conclude. The metrics that define a product’s success — time spent, frequency of use, number of exchanges per conversation that XiaoIce was so proud of — can obscure the fact that this engagement replaces meaningful human relationships. They propose designing AI companions that facilitate connections between people rather than maximizing interaction with the machine itself: repositioning the AI, they write, “from a competitor for our social attention” to a system that supports relationships between people. They add that the safety of these systems cannot be evaluated message by message: risk can arise from a habit, even when no individual exchange is harmful in itself.

This may be the most useful lesson from this body of work. An AI that always gives us what we want teaches us nothing about ourselves and gently distances us from the only relationships that can truly challenge us. The measure of a good artificial companion is not the time we spend with it, but what we become, afterwards, with others.

Sources