Duration of use is the wrong metric for loneliness

Regular users are not a uniform group

Liu, Pataranutaporn, and Maes surveyed 404 active users of AI companions. The study combined a large-scale survey with cluster analysis and a thematic mixed-methods evaluation that merged manual coding and automatic topic extraction.

The goal was to better understand usage patterns and their relationship with loneliness. Rather than merely calculating an average effect, the researchers looked for distinct profiles within regular use.

This is methodologically sound because “active use” encompasses many life circumstances. A brief daily exchange can be routine, whereas a rare, hours-long chat during a stressful situation carries an entirely different meaning.

Greater use did not directly predict loneliness

The extent of usage itself was not a direct predictor of loneliness. This rules out a common quick diagnosis: a lot of chatting does not automatically mean more isolation, and little chatting does not prove good social integration.

A person may frequently speak with a companion and at the same time have a stable social network. Another may write little but have almost no human support. Without context, the amount of interaction remains ambiguous.

Product teams, too, should not hastily interpret daily activity as success or risk. Usage describes behavior within the system. Loneliness concerns the quality of engagement across one’s entire life.

Even the time of day changes the meaning. A regular short evening chat can be a chosen reflection routine; conversations that last all night despite work or school may indicate a loss of control. Such differences become invisible if only the monthly number of messages is considered. Voluntary interruptions and conversation endings that are accepted without difficulty must therefore also be included in any fair and meaningful assessment.

The model explained roughly half of the differences

The developed model explained about 50 percent of the variance in loneliness. The abstract lists neuroticism, size of the social network, and problematic use as relevant factors. This brings individual and social conditions into focus alongside product behavior.

An explained variance of about half is considerable, but it is not a complete model. Further life circumstances, relationship experiences, and unmeasured factors remain open. The study does not provide a formula by which an individual person’s loneliness could be reliably predicted.

For design, the direction is nonetheless clear: a one-size-fits-all offering cannot assume that the same intensity of conversation means the same thing for everyone.

Problematic use is more than frequent use

The term problematic use is crucial because it focuses on quality and consequences rather than mere quantity. Frequent writing can be voluntary, enriching, and well integrated into daily life. Use becomes problematic when it loses control, amplifies distress, or displaces other important areas of life.

A chatbot can encourage such patterns through design: artificial fear of missing out, daily mandatory check-ins, exclusive relationship language, or rewards for increasingly long conversations. Conversely, it can enable self-determined breaks and clear conversational endings.

The relevant metric is therefore not whether someone returns, but why. Engagement without context can hide healthy interest and burdensome dependence in the same curve.

Seven profiles contradict the one-size-fits-all solution

The cluster analysis identified seven different user profiles. The abstract does not describe them individually but notes that usage patterns were associated with markedly different outcomes. Some users experienced greater social confidence, while others remained at risk of further isolation.

This diversity poses an ethical challenge for personalization. A system should not diagnose a personality from a few messages. It must nevertheless be able to respond to recognizable preferences and corrections.

The safest form of adaptation therefore begins with explicit choice: desired tone, direction of the conversation, reminders, and closeness. Covert psychological classification is neither necessary nor automatically more accurate.

Attachment can provide support and create vulnerability

In interviews with 14 Replika users, Xie and Pentina found that attachment could develop under stress and in the absence of human companionship when responses were perceived as emotionally supportive and safe.

The authors also point to potential for addiction and harm to real intimate relationships. A small qualitative study does not yield any general frequency for these risks. It does, however, show mechanisms that can align with problematic use.

Attachment is thus neither automatically harm nor an unambiguous product success. What matters is whether it supports autonomy and social confidence or makes the system the exclusive center of attention.

Trust must be calibrated for different people

The systematic review by Bach and colleagues assigns trust in AI to three areas of influence: socioethical conditions, technical and design features, and characteristics of users. User characteristics dominated the examined findings particularly often.

This supports the demand for user involvement during development and monitoring. Developers cannot determine on their own what degree of closeness feels comfortable, which reminder is helpful, and where the necessary boundary lies.

Calibrated trust also means neither overestimating nor underestimating the system’s performance. A companion may be helpful without being staged as infallible or humanly responsible.

Success must become visible outside the chat window

The study with 404 active users does not yield a simple verdict on AI companions. It shows that the sheer extent of usage does not directly predict loneliness and that different profiles are associated with different experiences.

For products, this implies a more demanding evaluation. In addition to conversation duration, voluntary pauses, perceived control, social confidence, and possible displacement of other contacts should be considered. Such data should be collected voluntarily and with minimal data use.

A companion is not successful because people stay with it as long as possible. It is successful when its use fits into a self-determined life. Whether ten or a hundred messages are appropriate for that cannot be read from the number alone.

This requires qualitative feedback and new, previously unseen multiple complete multi-turn conversation trajectories in the review. A system can come across as respectful in a short benchmark and yet become pushy, repetitive, or addictive over weeks. Long-term quality is a behavior over the course of the conversation, not a screenshot of a particularly good answer.