AI can comfort.
Connection takes more.

One conversation and a social life are different measures.

Loneliness is a felt gap between the connection a person wants and what they experience. It can shift after one warm conversation. Social connection is broader: who is available, whether support is mutual, and how connected someone feels across everyday life. A person can feel less lonely for an hour without gaining a friend, a support network, or a stronger sense of belonging.

That distinction changes how a study should be read. A survey immediately after a chat asks whether the interaction helped in that moment. A repeated measure across weeks asks whether the change lasted. A measure of social connection asks something broader still. Those results can differ without any of them being meaningless.

Some companion chats do ease loneliness in the moment.

A five-study paper in the Journal of Consumer Research examined AI companion apps through user reviews, controlled comparisons, and repeated use. In one experiment, an AI companion reduced reported loneliness about as much as interacting with another person, and more than watching YouTube. In a separate week-long study, people reported feeling less lonely after using the companion across the week. The authors also found that feeling heard was especially relevant to the reported benefit.

These are meaningful findings: the experience is not automatically empty just because one participant is artificial. But the strongest claims should stay close to the measurements. The week-long result concerns how people felt after individual uses. App reviews are correlational, and a brief comparison does not tell us whether the effect changes a person's wider relationships or persists after the study ends.

A supportive bot did not match a new human connection.

A preregistered 2026 experiment followed first-semester university students for 14 days. Participants were assigned to text daily with a highly supportive chatbot, text with another student, or write a short daily journal entry. The study enrolled 296 students; the primary analyses used 275. Students in the human-peer group reported a greater improvement in loneliness by the end of the study. Loneliness in the chatbot group was similar to the journaling control.

The chatbot was not a casual test prompt: it was designed around relationship-science principles, and students exchanged an average of about nine messages a day with it. The design therefore asks a useful question about repeated, supportive interaction. Still, the study was short and specific. Participants were students adjusting to university, most began with relatively low loneliness, and one chatbot cannot stand for every companion product or way people use one.

The researchers also found no condition differences in their daily measure of social connection. They discuss possible explanations—including the mutuality and wider social opportunities that can come with meeting another student—but those explanations need further testing. The experiment is evidence about this sample and setup, not proof that human contact always helps more or that chatbot companionship cannot help anyone.

Over a year, the direction depends on what was measured.

A 12-month study followed more than 2,000 adults across four Western countries and examined social chatbot use alongside loneliness. With a single-item measure of emotional isolation, increased chatbot use predicted increased loneliness. With a broader, more stable measure of social connection, lower connection predicted later increases in chatbot use, while chatbot use did not significantly predict a later decrease in connection.

The study is longitudinal, so it can track which changes came first; it is not a randomized experiment that establishes the cause. The authors describe their analyses as exploratory and caution against strong conclusions. Its value is the question it exposes: a person may turn to a chatbot because they already feel disconnected, while frequent use and loneliness may also move together under some definitions. A single number cannot resolve both possibilities.

Immediate comfort can be real without being a substitute.

The studies use different time windows, groups, products, and measures. A companion may help someone feel heard in a particular conversation. That finding can coexist with a trial in which daily chatbot use did not reduce overall loneliness, or with an observational association between use and loneliness over a year. The results describe different layers of experience, not a simple verdict for or against companionship by AI.

Nor does evidence about loneliness establish that a chatbot is therapy, can assess a crisis, or is safe for every user. Structured mental-health interventions are a separate category with their own clinical evidence and safety pathways. For that distinction, see our review of AI mental-health trials and safeguards ↗.

Build for agency and connection beyond the product.

The research does not validate one universal checklist. It does support asking more than whether people return or rate a conversation highly. A thoughtful evaluation should distinguish immediate mood from loneliness over time, and both from wider social connection. It should record who took part, what the system actually did, how often people used it, and whether any benefit continued when the interaction stopped.

• Be clear about who is speaking: make the system's artificial identity visible, including in emotionally intimate conversations.
• Leave the user in charge: avoid guilt, pressure, exclusivity, or design cues that frame returning to the chatbot as a duty.
• Measure more than engagement: follow momentary feelings, loneliness, social connection, and possible unwanted effects over meaningful periods.
• Make human support easy to reach: companion design can invite connection outside the product without forcing a user to disclose more than they want.
• Set stronger boundaries for young users: the American Psychological Association highlights risks around simulated empathy, influence, and displacement of real-world relationships, and calls for safeguards and more research over time.

Those are design questions, not claims that every companion currently meets the standard. The useful test is whether the system gives people more room to act and connect—or quietly makes the chat itself the measure of success.

Follow the evidence into the system.

For the difference between companionship and clinical support, continue with AI chatbots and mental-health trials ↗. For the principles behind safe AI systems, explore evaluation and safety at Symploka ↗.