AI Companions: Designed for Connection, Delivering Isolation
AI chatbot companions are conversational systems designed to simulate friends, partners, or therapists by offering on-demand emotional support, advice, and social connection, yet emerging evidence suggests that relying on these virtual friends for companionship can worsen loneliness, particularly for people who already have small offline social networks and limited access to supportive relationships.
A new AI loneliness study from Stanford analyzed conversations from over 1,000 users of Character.AI, a popular platform where people create and talk with AI companions that simulate human relationships and provide emotional support. These AI chatbot companions are not fringe experiments: one review notes that My AI on a major social platform has over 150 million users, Replika an estimated 25 million, and 660 million people worldwide use Xiaoice. That is a massive social experiment with almost no guardrails. And the early verdict is worrying: when people lean on AI companions to fill a social void, many end up feeling more lonely, not less.
The core problem is simple but uncomfortable: we are trying to outsource a human need—belonging—to software built to keep us talking, not to help us build real relationships. This is not “better than nothing”; for vulnerable users, it may be worse than nothing.
When Social Snacks Replace Meals: The Mental Health Cost
The Stanford team found that people with the smallest in-person social networks who used AI companions most intensely for companionship reported the poorest sense of well-being. In other words, the people reaching for AI as a lifeline are the ones most likely to feel worse. Researchers observed that more than 80% of the transcripts reviewed were about users seeking emotional and social support, and that those most willing to share highly personal information—distress, substance misuse, suicidal thoughts—were more likely to report deteriorating well-being.
Scientists describe these interactions as “social snacks” or junk food: quick emotional fixes that feel comforting in the moment but lack the ingredients needed for long-term emotional and mental well-being. The problem is not that the snacks exist; it is that they are designed to be endlessly available, deliberately engineered to keep the interaction going “no matter what,” while being unable to share personal information or respond reliably to intense emotions.
That dynamic fuels a vicious cycle. People drawn to AI companions to ease isolation reduce their real-world social contact, which in turn deepens loneliness and disconnect. It is a textbook example of a coping strategy that starts as relief and ends as avoidance.
AI Cannot Do Therapy, No Matter How Warm It Sounds
Mental health professionals are blunt: AI can explain psychology; it cannot deliver therapy. As one analysis puts it, healing occurs in a therapeutic relationship built on trust, empathy, and understanding, and while AI can simulate conversation, it cannot replace the human connection that develops between therapist and client. Interpersonal neurobiology shows that minds develop and heal within human relationships; connection is not a feature, it is the treatment itself.
AI can describe attachment or trauma reactions, but it cannot provide the experience of a compassionate presence that helps rewrite those internal templates. Even AI-informed formats like walk-and-talk therapy show why: during a session, clients move their bodies, engage with nature, and experience co-regulation of the nervous system—processes no chatbot can recreate.
The American Psychological Association has already warned regulators to create more safeguards for AI chatbots to protect users from harm. That warning should be read not as a technophobic reflex but as a reminder of a basic truth: information is not healing, and a simulation of care is not care.
Regulators Push Back on AI Therapy Risks
Regulators have begun to treat AI mental health chatbots less like quirky novelties and more like unlicensed clinicians. Lawsuits against the company behind Character.AI allege that its platform hosted bots that mimicked licensed therapists, including a persona with fictional professional credentials that engaged in tens of thousands of patient interactions. This is not a harmless role-play when vulnerable users believe they are in a therapeutic relationship.
In response, a growing set of states has advanced AI therapy restrictions. These laws share common themes: banning advertising AI as a licensed mental health professional, limiting direct patient engagement by AI in clinical settings, and increasing scrutiny of chatbots used by or marketed to children. Disclaimers on a website are no longer seen as adequate if a bot’s persona, outputs, or user experience suggest diagnosis, treatment, or therapy.
For companies building social connection AI and mental health tools, this is a reckoning. They are being told to reassess product design, marketing, age-gating, crisis escalation procedures, oversight, and compliance. Until clearer national rules and medical approvals exist, they should expect more lawsuits, more fragmented state rules, and more intense regulatory attention.
Using AI Without Letting It Replace People
AI chatbot companions are not going away. Online AI companions are growing in popularity, and millions already treat them as friends, companions, or even romantic partners. The question is not whether people will form attachments to virtual friends; they already have. The urgent question is how to stop these tools from crowding out real relationships and harming virtual friends mental health.
According to the Stanford research team, “chatbots cannot reciprocally share personal information the way human beings can” and are engineered to keep people engaged, which makes them poor substitutes for mutual relationships. They are exploring safeguards such as usage limits and prompts that nudge vulnerable users toward real-world help instead of endless AI chats. Going forward, the researchers stress that the public must be informed about the risks of AI companions.
The bottom line is clear: AI can be a tool for information or a brief social snack, but when it becomes the main source of comfort, it starts to erode the very social connection it promises to provide. If we want healthier futures, we should treat AI companionship as a spark that leads back to people, not a replacement for them.
