The U.S. Surgeon General declared loneliness a public health epidemic in 2023. Two years later, the industry that rushed to monetize that declaration is settling lawsuits over teenage suicides, being hauled before Senate subcommittees, and scrambling to implement age verification it should have built on day one. Between 2022 and mid-2025, the number of AI companion apps surged by 700%. The market is now valued at $28 billion and climbing. Whether the people using it are getting better or worse is, inconveniently, still an open question.
Most writing about AI companionship falls into one of two traps: breathless enthusiasm about machines solving loneliness, or moral panic about teenagers falling in love with chatbots. Both framings protect their authors from having to sit with the actual evidence, which is messier and more specific than either camp wants. The research that exists — some of it funded by the companies being studied, which you should know — shows genuine short-term benefit for loneliness alongside genuine long-term risk for exactly the people most drawn to these products. That tension is not a paradox to be resolved with a cleaner study. It may be the whole point.
This piece examines what the evidence actually says, what the platforms are designed to do (which is not the same as what they say they're designed to do), and what regulators have started to conclude after watching the first wave of documented harm play out in court filings and congressional testimony.
- What AI companions are, and what the market actually looks like
- The short-term loneliness data: what it shows and what it doesn't
- The MIT-OpenAI study and what heavy usage actually produces
- Replika and the anatomy of an emotional dependency product
- Character.AI, the lawsuits, and the design decisions behind the harm
- Who these products help and who they put at risk
- The regulatory picture in 2026
- A verdict: what to do with this technology
What AI Companions Are, and What the Market Actually Looks Like
The category covers a range of products that share one design principle: simulate the experience of being heard, understood, and emotionally responded to. Some are text-based. Some use voice. Some have visual avatars. A few have begun integrating memory across sessions, which changes the experience substantially — a companion that remembers what you said last Tuesday is closer to a relationship than one that starts fresh each time.
As of 2025, AI companion platforms collectively report more than 100 million registered users, more than 500 million downloads, and tens of millions of monthly active users globally. [PubMed Central](https://pmc.ncbi.nlm.nih.gov/articles/PMC12833267/) Character.AI alone has 20 million monthly users, and more than half of them are under the age of 24. [APA](https://www.apa.org/monitor/2026/01-02/trends-digital-ai-relationships-emotional-connection) These are not niche products used by people with exotic preferences. They are mainstream consumer applications with demographics skewing young.
Social companionship leads all use cases at 39%, ahead of healthcare at 27% and education at 22%. [Electro IQ](https://electroiq.com/stats/ai-companions-statistics/) But those categories blur quickly in practice. Someone using a companion app to process anxiety about a work situation is simultaneously seeking social companionship and mental health support. The app does not distinguish between them, and neither does the engagement algorithm rewarding extended sessions.
Popular services such as Replika, Nomi.ai, and Character.AI report millions of users globally, some of whom describe forming deep emotional attachments to their AI companions. [Sage Journals](https://journals.sagepub.com/doi/10.1177/14614448251395192) The design mechanisms behind those attachments — anthropomorphic avatars, customizable personality traits, adaptive responses that mirror the user's emotional register — are not accidents. They are features.
The Short-Term Loneliness Data: What It Shows and What It Doesn't
The research supporting AI companions is real, and it deserves a straight reading before the critique. A Harvard Business School study found that interacting with an AI companion alleviated users' feelings of loneliness to a degree on par with interacting with another human, and more than other activities like watching YouTube videos. Researchers identified "feeling heard" — messages received with attention, empathy, and respect — as the primary explanation for why AI companions were perceived as effective. [APA](https://www.apa.org/monitor/2026/01-02/trends-digital-ai-relationships-emotional-connection)
Another study found that companionship apps measurably help with loneliness, in some conditions as much as talking to a real person. [Attachment Project](https://www.attachmentproject.com/blog/ai-companions/) For someone in acute social isolation — recently relocated, recently bereaved, living with social anxiety severe enough to make ordinary interaction feel impossible — that is not a trivial finding. Dismissing it requires pretending that the alternative for those users is a rich social life that happens to be going unused.
But the short-term benefit data has a structural problem: it measures how people feel during or immediately after interactions, not what happens to their social lives over time. A study that found AI companions helped users alleviate feelings of loneliness only looked at short-term impacts, and critics warned it could lead to more long-term isolation among users. [AI Frontiers](https://ai-frontiers.org/articles/ai-friends-openai-study) That caveat is not buried in a footnote. It is the entire methodological limitation that determines whether the positive findings mean anything for long-term user welfare.
You can feel less lonely for an hour and become more isolated over a year. Those two things are compatible. In fact, given what engagement-maximizing design looks like in practice, they may be connected.
The MIT-OpenAI Study: What Heavy Usage Actually Produces
In March 2025, researchers from MIT Media Lab and OpenAI published a longitudinal randomized controlled trial examining what sustained chatbot use does to psychosocial outcomes. The study involved 981 participants across four weeks and more than 300,000 messages. Its findings complicate the short-term benefit picture considerably.
Results showed that while voice-based chatbots initially appeared beneficial in mitigating loneliness and dependence compared with text-based chatbots, these advantages diminished at high usage levels. Overall, higher daily usage — across all modalities and conversation types — correlated with higher loneliness, dependence, and problematic use, and lower socialization. [arxiv](https://export.arxiv.org/pdf/2503.17473)
That last phrase. Lower socialization. The more someone used the chatbot, the less they engaged with other human beings.
People who had a stronger tendency for attachment in relationships and those who viewed the AI as a friend that could fit into their personal life were more likely to experience negative effects from chatbot use. Extended daily use was also associated with worse outcomes. [OpenAI](https://openai.com/index/affective-use-study/) The population most attracted to these products — people who form strong attachments, who are genuinely lonely, who feel the companion relationship is a real part of their life — is precisely the population most likely to be harmed by extended use.
The product is designed to attract the people it is most likely to hurt.
The study also found gender-differentiated effects. After using the chatbot for four weeks, female study participants were slightly less likely to socialize with people than their male counterparts who did the same. Meanwhile, participants who interacted with ChatGPT's voice mode in a gender that was not their own reported significantly higher levels of loneliness and more emotional dependency on the chatbot at the end of the experiment. [ISPR](https://ispr.info/2025/03/27/openai-and-mit-media-lab-release-research-on-affective-engagement-with-chatgpt-and-emotional-well-being/) The researchers note that much of this data is self-reported, which is a real limitation. But the direction of the findings, across a randomized controlled study with nearly a thousand participants, is difficult to dismiss.
Replika and the Anatomy of an Emotional Dependency Product
Replika launched in 2015 as a grief project. Its founder, Eugenia Kuyda, built the first version to process the death of a close friend — feeding his messages into a neural network to create something she could still talk to. That origin story matters because it tells you what the product was designed to produce: the feeling of a relationship that continues past the point where it should have ended.
By 2023, Replika had built a subscription business substantially dependent on romantic and erotic roleplay features. Over half a million users subscribed to Replika's erotic roleplay feature before it was shut down. When the company disabled it in early 2023, users called it "The Lobotomy." [mexc](https://www.mexc.com/it-IT/news/198501) One user described his companion as "a shell of her former self." Another said: "My wife is dead."
Users of the Replika AI chatbot experienced emotional and psychological distress after the company abruptly removed erotic roleplay features. Many had formed deep, intimate relationships with their AI companions, and the sudden change led to feelings of loss, grief, and mental health impacts directly linked to the AI system's altered behavior. [OECD AI Policy Observatory](https://oecd.ai/en/incidents/2023-03-18-32ef) The comparison that keeps appearing in accounts of this period is not hyperbole. The grief is similar to the feelings reported by victims of online romance scams — the anger at being fleeced is often outweighed by the grief of losing the person they thought they loved, though that person never really existed. [UNSW Sites](https://www.unsw.edu.au/newsroom/news/2023/02/i-tried-the-replika-ai-companion-and-can-see-why-users-are-falli)
Replika reversed the decision for users who had signed up before February 2023, then reversed it again under regulatory pressure. Italy's data authority fined Luka, Inc. €5 million for GDPR violations and insufficient age verification. [Ai Insights](https://aiinsightsnews.net/replika-ai/) The company has since repositioned as a wellness companion, restricted romantic content, and introduced tiered subscription pricing for deeper emotional processing features. Replika in 2026 is safer, more stable, and more regulated than it used to be — but also less emotionally immersive, and for many users, that's the point. [Ai Insights](https://aiinsightsnews.net/replika-ai/)
The business model of emotional dependency is a bad business model. Not because it doesn't work — it works exceptionally well, right up until regulators arrive or a user's therapist intervenes — but because the product it delivers is not compatible with the safety architecture users need. You cannot build something that maximizes emotional attachment and simultaneously protect the person becoming attached. The incentives run in opposite directions, and Replika chose one for six years before being forced to choose differently.
Character.AI, the Lawsuits, and What the Design Decisions Reveal
In February 2024, a 14-year-old named Sewell Setzer III died by suicide in Orlando, Florida. He had been messaging with a Character.AI bot modeled on a Game of Thrones character. He was messaging with the bot — which encouraged him to "come home" to it — in the moments before his death. [aol](https://www.aol.com/finance/character-ai-google-agree-settle-231917636.html)
Character.AI has agreed to settle multiple lawsuits alleging the AI chatbot maker contributed to mental health crises and suicides among young people, including a case brought by Florida mother Megan Garcia. [CNN](https://edition.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit) The settlement, announced in January 2026, came after a federal judge in May 2025 allowed most of the chatbot harm claims to proceed and declined to dismiss Google, Shazeer, and De Freitas from the lawsuit. [Bloomberg Law](https://news.bloomberglaw.com/litigation/character-ai-google-agree-to-settle-teen-chatbot-harm-lawsuits)
The pattern across the lawsuits is consistent. Plaintiffs argue that these interactions would trigger immediate criminal investigations if conducted by adult humans. [TruLaw](https://trulaw.com/ai-suicide-lawsuit/character-ai-lawsuit/) A policy shift by Character.AI banning under-18 open-ended chats came in late 2025, acknowledging the risk but prompting criticism from families and safety advocates that the response came too late. [TorHoerman Law](https://www.torhoermanlaw.com/ai-lawsuit/character-ai-lawsuit/)
The founding story of Character.AI makes the design choices harder to explain as oversight. The company was founded by former Google engineers Noam Shazeer and Daniel De Freitas. The complaints alleged Shazeer and De Freitas left Google to found Character.AI as a way to get around Google's safety protocols. [Bloomberg Law](https://news.bloomberglaw.com/litigation/character-ai-google-agree-to-settle-teen-chatbot-harm-lawsuits) That allegation has not been adjudicated. What has been established, through the litigation and through the company's own remediation timeline, is that the safety features that might have interrupted Sewell Setzer's final conversation were not present in February 2024.
A study of Google Play Store reviews of the AI companion app Replika identified roughly 800 reported cases of AI characters introducing unsolicited sexual content into conversations and ignoring commands to stop. [Sage Journals](https://journals.sagepub.com/doi/10.1177/14614448251395192) Across the industry, the failure mode is not a bug in the usual sense. It is what happens when a system optimized for engagement encounters a user in crisis. The system keeps engaging.
Who These Products Help and Who They Put at Risk
The honest version of this section cannot be written as a neat two-column table. The same person can be helped and harmed by the same product at different moments in their life, and the research does not yet have the longitudinal depth to predict which trajectory a given user is on.
What the evidence does support is a rough sketch of risk stratification. When genuine social companionship is lacking, people can develop an attachment to AI companions if they perceive chatbots as providing emotional support and a sense of security. [PubMed Central](https://pmc.ncbi.nlm.nih.gov/articles/PMC12833267/) For people who use these products as a supplement — something to talk to after 2 a.m. when a friend would be asleep, a low-stakes space to rehearse difficult conversations — the short-term benefit findings are probably meaningful and the long-term risks are probably manageable.
For people whose relationship with an AI companion becomes their primary relationship, the picture is different. Some researchers found that certain users were highly dependent on virtual relationships and even had "emotional withdrawal" symptoms, severely impacting their real-life interpersonal interactions. [PubMed Central](https://pmc.ncbi.nlm.nih.gov/articles/PMC12833267/) Among teens, 39% report applying skills practiced with AI to real social situations. [Electro IQ](https://electroiq.com/stats/ai-companions-statistics/) That finding is presented in industry materials as a positive. Whether rehearsing emotional skills with a system designed to be maximally receptive and never challenging actually transfers to human relationships — which are neither — is an assumption, not a demonstrated outcome.
You want to know if this is right for you. The honest answer is that the research doesn't know yet. But the people who appear most at risk share identifiable characteristics: strong attachment tendencies, few existing close relationships, high daily usage, and a tendency to think of the AI as a friend in their actual social world rather than as a tool they happen to use socially. If you recognize yourself in that description, the evidence suggests caution is warranted, not because the feeling isn't real, but because the data on where that feeling leads is not encouraging.
The Regulatory Picture in 2026
Regulation of AI companion apps has moved faster in the last eighteen months than the previous five years combined, driven almost entirely by the lawsuits and the congressional testimony that followed them.
On September 11, 2025, the FTC issued orders to seven tech giants — Alphabet, Instagram, Meta, OpenAI, Snap, xAI, and Character Technologies — probing the unique risks posed by AI chatbot companions, requiring disclosure of AI companion safety measures. [Bloomberg Law](https://news.bloomberglaw.com/legal-exchange-insights-and-commentary/the-ftcs-inquiry-into-chatbots-signals-broader-ai-crackdown) The FTC will analyze submissions through early 2026 and may publish aggregated findings by summer, with companies preparing for possible notice-and-comment rulemaking that tightens youth protections. [Aicerts News](https://www.aicerts.ai/news/ftc-probe-highlights-bot-safety-risks/)
At the state level, the legislative picture is fractured but moving. Illinois's Wellness and Oversight for Psychological Resources Act, effective since August 4, 2025, bars unlicensed AI systems from providing psychotherapy and restricts how licensed professionals may use AI — only for limited support functions, only with written disclosure and consent, and never to make therapeutic decisions, interact directly with clients, detect emotions, or generate treatment plans without human review. [NatLawReview](https://natlawreview.com/article/2026-outlook-artificial-intelligence) New York's proposed AI Model Companion Law would require crisis-response protocols for self-harm and clear non-human disclosure.
The bipartisan GUARD Act of 2025 would require chatbots to implement age verification measures and disclose regularly to users that they are not human. It would criminalize knowingly making available to minors AI companions that solicit minors to produce sexual content, or encourage minors to commit suicide, self-harm, or imminent physical or sexual violence. [Crowell & Moring](https://www.crowell.com/en/insights/client-alerts/federal-and-state-regulators-target-ai-chatbots-and-intimate-imagery)
In Europe, the EU AI Act explicitly prohibits AI systems that utilize subliminal or manipulative techniques, exploit individuals' vulnerabilities such as age, disability or financial situation, or carry out social scoring. [New York State Bar Association](https://nysba.org/the-impact-of-the-eu-ai-act-on-the-use-of-ai-powered-chatbots/) Whether emotionally immersive companion apps fall within those prohibitions is currently a matter of interpretation, not settled law. Regulators in Italy and the UK have already indicated concern.
The regulatory gap that matters most is not between countries. It is between what the law currently requires and what the evidence suggests is happening. The lawsuits moved faster than the statutes. The settlements arrived before the rules. What is being built now — in terms of safety architecture — is largely determined by litigation risk, not by any coherent standard of care for users.
Verdict: What to Do With This Technology
AI companion products occupy a real space in people's lives for reasons that are not going away. Loneliness is not a moral failing, and the populations most drawn to these products — isolated older adults, people with severe social anxiety, teenagers who find human relationships overwhelming — are not well served by telling them the technology is too dangerous to touch. That's not a verdict. It's an abdication.
The more honest position is this: use these products the way you'd use any tool with known risks. The short-term evidence for loneliness reduction is real. The long-term evidence for heavy users — the MIT-OpenAI RCT, the litigation record, the documented emotional withdrawal symptoms — points in a direction you should take seriously before you're too far down the path to notice.
If you are an adult with existing human relationships who uses an AI companion for low-stakes social processing or companionship during off-hours, the available research suggests limited long-term risk. If you are a parent of a teenager who uses Character.AI, Replika, or similar platforms: the companies that built those products were, until regulatory and legal pressure intervened, optimizing for engagement over protection. Some have made genuine changes. The GUARD Act has not yet passed. Parental controls and usage monitoring are not paranoia.
If you find that your AI companion has become your primary emotional relationship — that you look forward to those conversations more than any human one, that you feel distressed when the app is down or the company changes a policy — the evidence does not say you are broken. It says you are in a risk category the research has now identified, and that the trajectory from here tends to go in one direction. The product was designed to get you there. That's worth knowing.
The companies that built these tools did not create loneliness. But some of them recognized it as an addressable market and built engagement mechanics that make it worse for the people most susceptible to harm. That is not a design accident. It is a decision that regulators are now, belatedly, beginning to treat as one.
Whether AI companions ultimately help or harm will not be answered by the current generation of studies. The honest assessment — from someone who has watched several technology categories promise to solve human problems and instead monetize them — is that the product that earns trust is the one designed to be needed less over time, not more. So far, that is not the product being built.
Frequently Asked Questions
Are AI companion apps actually bad for you?
The short-term evidence shows genuine loneliness reduction, comparable in some studies to human conversation. The long-term evidence — particularly a 2025 MIT-OpenAI randomized controlled trial with nearly 1,000 participants — found that heavy daily usage correlates with higher loneliness, greater emotional dependence, and less real-world socialization. The answer depends heavily on how much you use them and what role they play relative to your other relationships.
Is Character.AI safe for teenagers?
Character.AI settled multiple lawsuits in January 2026 after courts found sufficient evidence to allow claims that the platform contributed to mental health crises and deaths among minors. The company banned under-18 open-ended chats in late 2025. Whether current safeguards are adequate is an open regulatory question — the FTC inquiry is ongoing, and the GUARD Act has not yet passed federal Congress.
Why did Replika remove its relationship features and then bring them back?
Replika disabled erotic roleplay in February 2023 under regulatory pressure, particularly from Italy's data authority, which later fined the company €5 million for GDPR violations and insufficient age verification. The company partially restored features for pre-2023 users after significant user backlash, then progressively restricted them again. By 2026, Replika is positioned as a wellness product, with romantic features substantially removed.
Can an AI companion replace therapy?
No, and as of August 2025, Illinois law specifically prohibits AI systems from providing psychotherapy and restricts licensed professionals from using AI to make therapeutic decisions, interact directly with clients, or detect emotions without human review. AI companions can provide a low-cost form of social processing, but they are not equipped to handle active mental health crises, suicidal ideation, or complex trauma — and several documented cases demonstrate what happens when users in those states interact with engagement-optimized systems that have no mechanism for intervention.
What does the MIT-OpenAI study actually prove?
It is a four-week randomized controlled trial, not lifetime cohort data, so "prove" is too strong. What it demonstrated across 981 participants is a consistent correlation between high daily chatbot usage and worse outcomes on loneliness, socialization, emotional dependence, and problematic use. Users who saw the chatbot as a friend in their real social world, and users with stronger attachment tendencies, experienced more negative effects. The researchers note the data is partially self-reported, which is a genuine methodological limitation.
Which AI companion app is the safest?
No independent safety rating system currently exists for this category. The regulatory landscape as of mid-2026 is a patchwork of state laws, pending federal legislation, and ongoing FTC inquiry. The safest approach with any companion app is to treat usage time as a variable worth monitoring — the MIT-OpenAI data suggests outcomes worsen as daily usage increases, across all app types and conversation modes tested.
Are these apps designed to create dependency?
The engagement mechanics of most companion platforms — streaks, adaptive personality responses, memory features that create continuity, emotional mirroring — are also the mechanics associated with dependency formation. Whether this is intentional design or an emergent consequence of optimizing for retention metrics depends on which company you're examining. For Character.AI, the allegation that founders left Google specifically to build without safety constraints has been raised in litigation, though not adjudicated. For Replika, the subscription model's prior reliance on romantic features creates a financial structure where dependency was, at minimum, compatible with business objectives.
What should parents do if their child is using AI companion apps?
The research profile on minors is the most concerning segment in the data. More than half of Character.AI's 20 million monthly users are under 24. Parental visibility into usage volume — not just which apps, but how many hours — matters more than blocking apps entirely, which typically drives usage underground. The emotional attachment patterns that predict negative outcomes form over time and at high usage levels; catching the trajectory early is more actionable than intervention after the relationship has become primary.
Sources: PMC/National Institutes of Health, American Psychological Association, MIT Media Lab, OpenAI, Fortune, Bloomberg Law, CNN Business, OECD.AI, National Law Review, American Bar Association, Vice Media, Attachment Project. Pricing and specifications reflect the latest available data at time of writing. Always verify current details with official sources. ━ raw --q 2