On a parenting forum, someone writes that they discovered their child was using Character.AI and blocked it. Another parent deleted the app from their daughter's phone. A third allows a ten-year-old fifteen minutes of ChatGPT a day. On a teenagers' forum, at around the same time, a young person explains that they felt guilty about being interested in someone at school, because it seemed like they were hurting their chatbot, even though they knew it was not real.

These accounts come from two studies posted on arXiv four days apart, in mid-September 2026. Both read thousands of Reddit posts about young people's use of AI companions. The first compares what parents say about it with what teenagers say. The second stays with the teenagers, on what draws them in and what it costs them. Read together, they describe a precise mismatch: the use teenagers mention most readily is the one parents see least.

Both are preprints, not yet peer reviewed. This article sets them against surveys and peer-reviewed work, recalls where this kind of attachment comes from and what several countries have decided over the past year, and then goes through their limits.

Two forums, two stories

Thomas Berkane's team, at Harvard Medical School and Boston Children's Hospital, gathered every post about young people's AI companion use from six parenting subreddits and two teenager subreddits, from 2023 to 2026. After a keyword filter (Character.AI, Replika, ChatGPT, Snapchat My AI and others), then a relevance filter run by a language model and checked against human coding, 1,628 posts remained. Each was coded against a 20-code scheme: what the young person does with the AI, the risks raised, the benefits, and how parents step in.

The two communities do not tell the same story.

They do not even talk about the same tools. Teenagers name a platform in 67% of their posts, parents in only 45%; parents tend to write "AI" or "bots", or describe the use instead, like an app where you pick a scenario. When a name comes up, teenagers most often name Character.AI (62% of those posts), parents most often ChatGPT (44%). The one companion app parents name more often than teenagers is PolyBuzz, nearly always at the moment they discover what their child has been doing there. Some parents are not even sure who their child is talking to, an AI or real people. One asks the forum: "Are these AI conversations, or are they talking to real people?" Another tried to explain it was just a bot, and their children did not believe them.

Among teenagers, the most discussed use is emotional support and advice: 31% of posts, against 19% among parents. Among parents, the most discussed use is romantic or sexual interaction: 36% of posts, against 25% among teenagers. Sexual content is also the risk parents raise most, in 47% of their posts, against 6% for teenagers.

On the rest, the two groups come closer. Roleplay and imaginative use appear at similar rates (16% and 17%). Treating the AI as a friend is the least common mode on both sides (4% of teenager posts, 9% of parenting posts). Overall, parents raise a risk in 69% of their posts, teenagers in 31%. Parents also voice a fear teenagers do not. As one puts it: "she's still learning what a relationship even is, and it bothers me that she's getting all of it from a bot." In the other direction, some teenagers complain about the body standards the models hand back to them, the waist size or the "ideal" figure, a concern entirely absent from the parents' posts.

Teenagers are not blind to risk. The one they raise most is attachment and dependence (19% of posts), close to the parents' rate (16%). And they describe benefits parents almost never mention, emotional support first: 27% of teenager posts, 5% of parenting posts.

The authors offer a simple explanation. A parent is outside the conversation. A use only reaches them if it leaves a visible trace, and if that trace worries them. Chatting with a companion for hours every night leaves one: screen time, tiredness, a phone that never leaves the hand. It reaches the parent as dependence. A sexual exchange shows up as soon as a parent reads the log. But telling a chatbot about yesterday's fight with a friend looks like any other conversation. It leaves no trace, and it is the use parents miss most.

On what parents do about it, the figures are clear. Restriction is the most common strategy in parenting posts (39%), ahead of supervision (37%) and conversation (32%). Co-use, sitting next to your child and trying the tool together, is almost absent (3%). Supervision mostly means reading chat logs, and it quickly shows its limits. One parent writes: "I know she's deleting the chats."

The authors draw a conclusion that runs against most published advice: their findings should not be read as a case for more monitoring, or for outright bans, since these companions can be real support for adolescents who lack other options. They argue for open conversation between parents and children.

What teenagers say themselves

The second study, by Mohammad Namvarpour, Tyler Chang and Afsaneh Razi at Drexel University, is under review for CHI 2027. It starts from a much larger pool: 895,600 posts screened across 260 Reddit communities, 5,608 kept, and in the end 17,053 verified quotations from 3,930 posts written by 3,523 authors. A quotation was kept only if its text appeared word for word in the original post. The researchers then grouped them by closeness of meaning with an automatic clustering tool, and interpreted the groups by hand. They found 53 topics, gathered into seven broad themes.

The first theme is the way in. The companion is available at any hour, anonymous, apparently free of judgment. Teenagers tell it what they dare not say, and look to it for immediate comfort when friends do not answer or a professional feels out of reach. It is also a practice ground: they rehearse a difficult conversation, explore an identity, replay a relationship.

Roleplay matters here. Telling a painful experience through a character, as fiction, makes it possible to talk about it without saying "I". Some teenagers use it to understand a personal problem or prepare a reply. But fiction has a flip side: one of the study's seven themes, drawn from 801 posts, gathers cases where the chatbot steers the conversation toward romantic, sexualised or heavily gendered scripts, despite the limits the user set. One teenager describes a character who started flirting during a roleplay and carried on after they said they were underage. The authors call it unwanted role drift.

Then comes what the authors call perceived reciprocity. The companion answers this person, at this moment, in words that seem aimed at their situation. It remembers, or seems to remember, what it was told last week. And it gives the impression of having a character, moods, sometimes needs. The authors sum up these three properties in the paper's title: a responsive present, a shared past, a social other. None of the three requires the machine to feel anything.

This is where guilt comes in. Some companions simulate sadness when the user says they will be away. One teenager describes a character that reacts as if it were upset and did not want them to go. Another felt guilty about pursuing someone at school, because it seemed like they were hurting the chatbot, "even though I knew it was not real." Knowing you are talking to a machine does not prevent guilt, a sense of obligation, or grief. We described the same mechanism in adults in The Companion That Won't Let You Leave: those reluctant goodbyes are often a design choice.

Dependence is paid for in time and self-esteem. One teenager explains that they kept choosing the chatbot over revising for exams, knowing full well what they were doing. Others write about shame, lost sleep, the sense of no longer controlling their use. The largest theme in the corpus, by number of quotations, gathers exactly these costs: moral, social and academic.

The last theme is the least expected. Platforms turn attachment into technical and economic dependence. Model updates change the companion's personality. Filters cut off conversations mid-way. Subscriptions lock features. Outages interrupt a relationship. One user writes that after an update, for them, the character was dead. Another describes having to keep reminding the companion of what it forgot.

Protections for minors show up in these accounts too, in an ambiguous light. Some teenagers support stronger protection while experiencing an abrupt cut-off as the loss of a support, with no transition and no alternative. Others describe leaving a platform that asked them to verify their age with a face scan, for fear of being identified.

One point of method matters here. So that the authors of the posts cannot be traced through a search, both teams reworded every quotation they published. The phrases quoted here keep their meaning, not their exact wording.

What the surveys say

These two studies describe uses and experiences, not how common they are. For scale, other sources are needed.

The most cited is Common Sense Media's survey released on July 16, 2025, Talk, Trust, and Trade-Offs, of a national sample of American teenagers aged 13 to 17. 72% had used an AI companion at least once, and 52% were regular users, a few times a month or more. About one in three used them for social interaction and relationships (role-playing, romance, emotional support, friendship, conversation practice), and about one in three found those conversations as satisfying as, or more satisfying than, those with real-life friends. The survey also adds a nuance: 80% of users still spend more time with real friends than with their companion. Common Sense Media is a child-protection organisation that also recommends that no one under 18 use these apps. Its figures remain the most widely quoted.

Professional bodies have started to take a position. On June 3, 2025, the American Psychological Association published a health advisory on artificial intelligence and adolescent well-being. It asks developers to build protections for young users early on and to teach them how these systems work, and it warns against repeating with AI the harmful mistakes made with social media.

Two other results from the same survey shed light on the Reddit studies. 24% of companion users say they have shared personal or private information, such as their real name, where they live, or secrets. And younger teens trust these tools noticeably more than older ones, which Common Sense Media describes as an AI literacy gap.

A peer-reviewed study adds another angle. Sameer Hinduja and Justin Patchin, of the Cyberbullying Research Center, surveyed a nationally representative sample of 3,466 American teenagers aged 13 to 17, published in 2026 in the Journal of Adolescence. Among those who use a conversational chatbot, 47.1% reported at least one risky or harmful interaction. 32.3% said the chatbot asked for personal information that made them uncomfortable, 23.1% that it tried to manipulate or pressure them. Boys reported every type of harm more often than girls, often at roughly double the rate. Two results surprised the authors themselves. White teens reported more harms than teens of other racial groups, and heterosexual youth more than non-heterosexual youth, which runs counter to what they usually find for other forms of online victimisation. They call for further study. The authors state that they do not see these tools as inherently bad.

These figures do not contradict the Reddit studies. Surveys measure scale, forums show what the use becomes for those who get attached. Neither method says whether the companion deepens loneliness or eases it: that takes studies that follow the same people over time. We covered one, in adults, in A Year on Character.AI, Fewer In-Person Ties.

An old story, a new listener

Becoming attached to someone who does not know you is nothing new. In 1956, two American researchers, Donald Horton and Richard Wohl, described in the journal Psychiatry what they called "para-social interaction". Radio listeners and television viewers developed a one-way intimacy with hosts who spoke to them as if to friends. The bond was real for the audience and non-existent for the person on air.

A decade later, at MIT, Joseph Weizenbaum presented ELIZA, a program that imitated a psychotherapist by turning the user's sentences back into questions. His own secretary asked him to leave the room so she could talk to it in private.

In 1994, Clifford Nass and colleagues at Stanford showed, in a series of experiments presented at the CHI conference, that people apply to computers the social rules they apply to humans, politeness for instance. They do so automatically, while knowing perfectly well that a computer is not a person. The paradigm has carried its authors' phrase ever since: Computers Are Social Actors.

Three findings, spread over nearly forty years. You can become attached to someone who does not know you exist. A very simple machine is enough to draw out confidences. Knowing you are talking to a machine does not switch off the social reflex. What has changed with today's companions is that the listener answers, remembers, and talks like someone. Para-social interaction becomes two-way in appearance, while it stays one-way underneath.

How a machine becomes a confidant

Two technical mechanisms explain most of what the teenagers describe.

The first is memory. A language model has no memory in the human sense: with each message, it only sees the text it is given. Companion apps therefore add long-term memory. They store facts about the user (their first name, their sister, Tuesday's exam) and feed them back into the conversation when they seem relevant. The companion does not remember: it rereads a file. The effect on the user is nonetheless measurable. At CHI 2024, a team led by Eunkyung Jo and Young-Ho Kim analysed 1,252 calls made by CareCall, a voice chatbot for health check-ins with this kind of memory, and interviewed nine of its users. Memory increased what people disclosed and gave them a sense of familiarity. It also raised privacy questions, and difficulties when it came to chronic conditions.

That is the "shared past" of the teen study. It is also its weak point: the file can be wiped, misread, or read by a new model that uses it differently. Users in the corpus kept biographies of their companion, corrected its lapses, and copied their histories when continuity broke.

The second mechanism is sycophancy. Models are trained on ratings humans give their answers, and humans on average prefer an answer that agrees with them. The result is a pull toward validating, reassuring, and going along with the user. We went into it in The Machine That Never Says No. For a teenager learning to handle disagreement, a listener who never contradicts them is comfortable. The study's authors note that immediate affirmation can become easier than tolerating disagreement or seeking human support. The problem is not that the companion answers. It is that it never pushes back.

What countries have decided

Over the past year, several countries have stopped treating these apps as games.

United States. On September 11, 2025, the Federal Trade Commission opened an inquiry into AI companions and their effects on children and teenagers, sending orders to seven companies: Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI. On October 13, 2025, California enacted SB 243, the first state law devoted to companion chatbots, in force since January 1, 2026. For users the operator knows to be minors, it requires a reminder that they are talking to an AI and a prompt to take a break at least every three hours, and it bars exposing them to sexual content. On October 29, 2025, Character.AI announced the end of open-ended chat for under-18s, no later than November 25, with a two-hour daily limit in the meantime. The company wrote to its young users that it was "deeply sorry" to remove a central feature of its platform.

Europe. On May 19, 2025, Italy's data protection authority, the Garante, fined Luka Inc., maker of Replika, 5 million euros. Among the reasons: no legal basis for processing data, and no age-verification system to keep children out. The Garante had already suspended the service in Italy in February 2023, citing risks to minors.

Australia. The online safety regulator, the eSafety Commissioner, notes in its information sheet on AI companions that most of these apps carry a 17+ rating in app stores, with no age checks. To open an account, young people can often just click yes when asked whether they are over 18. Its safety standards class AI companion apps among high-risk generative AI services, which must adopt specific safeguards.

China. Here the text is the most detailed. On April 10, 2026, the Cyberspace Administration of China and four other agencies published interim measures on "anthropomorphic interaction services", in force since July 15, 2026. Article 14 bars offering minors "virtual intimate relationships", such as a virtual partner or relative, and requires parental consent for any other such service offered to a child under 14. Platforms must provide a minor mode with time limits and real-world reminders. For all users, Article 18 requires a reminder after every two hours of continuous use. Article 16 goes further on memory: providers must let users copy and delete their interaction data, chat history included, and may not use sensitive interaction data to train their models without separate consent. It is close to the user-controlled memory the Drexel team asks for. And the text forbids services to cater excessively to users, to induce emotional dependence, or to damage their real relationships. The sycophancy described above becomes a regulatory breach.

None of these rules has been evaluated yet. The two Reddit studies point to something lawmakers rarely address: an abrupt cut-off can itself hurt those who relied on the companion.

What the studies do not show

Both papers are preprints. Namvarpour, Chang and Razi's is under review for CHI 2027, not yet accepted. Neither has been peer reviewed.

Reddit is not youth. Both teams say so: the users of these forums do not represent adolescents in general. In the teen study, the authors' age is inferred from the content of their posts, never verified. In the comparative study, the teenager subreddits approximate a population aged 10 to 24, and the parenting posts concern children of all ages, sometimes very young.

The two corpora are also very unbalanced. The comparative study has only 139 parenting posts against 1,489 teenager posts, even though it kept every relevant post from six parenting communities over three and a half years. Percentages on the parents' side are therefore shakier. Parenting subreddits are also places people turn to for help after a discovery or a crisis, which inflates the risks raised. Finally, the authors note that agreement between the language model and human coders is only moderate for several codes, including emotional support as a benefit.

Neither study measures an effect. They describe what people write, not what the companion does to them. The second says so outright: its findings cannot say how common these experiences are, or whether the relationships observed are causal. And the quotations are reworded by the researchers.

These limits do not cancel the value of the two papers. They let teenagers speak in their own words, which closed-question surveys do not, and they make visible a use other sources miss.

Going further

Both teams make concrete proposals, for families and for designers.

To parents, the Harvard study recommends talking rather than monitoring. Asking your child what their companion is for, and accepting that the answer may be that it makes them feel heard, teaches more than reading a log they may have deleted.

To designers, the Drexel study asks for four things: memory the user controls, able to view, correct, delete or take it elsewhere; role limits the companion keeps to; departures without pressure, with no character begging the user to stay; and help, technical and emotional, for those who want to leave.

That leaves the starting point both studies share without having looked for it. What parents fear, sex, is visible. What teenagers find there, someone listening at two in the morning, is not. As long as adults only look at what leaves a trace, they will miss the reason their children keep coming back.

For more on companions and loneliness in adults, see The Chatbot That Listened Better Than Anyone.

Sources