Well-being has more than one direction

14,721 adults from two panel waves

The study used data from nationwide Japanese internet panel surveys conducted in December 2024 and January 2025. A total of 14,721 adults were analyzed. AI use was categorized as companion use or other use.

Social networks and support were measured with the short Lubben Social Network Scale, and loneliness with the UCLA Loneliness Scale. Multivariable linear regressions and models with restricted cubic splines examined associations and potential nonlinear moderation.

The large sample allows for differentiated statistical analyses. However, because the data were analyzed cross-sectionally, it is still not possible to determine with certainty whether the companion caused changes.

The period is also a snapshot of a rapidly changing product landscape. Replika, Character.AI, and other companions differ in memory, moderation, and business model. Results on the category of companion AI should therefore not be automatically transferred to every new system or later version.

Life satisfaction is an assessment

The evaluative domain describes how people judge their lives as a whole. This life satisfaction is broader than the mood experienced during a single conversation. It involves a comparison between expectations, circumstances, and the individual’s own assessment.

Use of an AI companion was associated with higher evaluative values. This could mean that supportive conversations contribute to a positive evaluation of life. It could equally be that people with certain resources are more likely to use such systems and are also more satisfied.

A product survey conducted immediately after the chat barely captures this dimension. Life satisfaction requires time and distance, not just a thumbs-up.

At the same time, the same person should not be measured repeatedly. A few voluntary measurement points can make changes visible without turning every conversation into a study. Equally important are open-ended feedback opportunities in which users can describe what the chat actually contributed to their evaluation of life.

Happiness describes immediate experience

The hedonic domain refers to happiness and pleasant experience. Here, a more direct connection with a friendly, relieving chat is plausible. A good conversation can indeed change one’s mood.

Yet short-term positive feelings are not always a complete measure of quality. A system can feel pleasant through uncritical affirmation while oversimplifying conflicts or generating unrealistic certainty.

Good companionship must therefore hold together well-being and truthfulness. It may politely contradict and leave uncertainty standing, even if immediate affirmation would be more easily evaluated positively.

Meaning and significance are not text products

The eudaimonic domain concerns meaning and significance in life. Here too, the study found a positive association with AI companion use. This is remarkable because this dimension goes beyond momentary mood.

A chat can help to name values, sort out conflicting goals, or see one’s own thoughts more clearly. However, it cannot generate meaning for a person. Meaning arises in decisions, relationships, and actions outside the system.

A responsible companion supports reflection without presenting its formulations as definitive life interpretations. It offers perspectives, not identity.

Particularly dangerous are convincing-sounding interpretations based on very little context. A model can formulate patterns that sound meaningful and are nevertheless wrong. Meaning-related conversations therefore require hypotheses in cautious language and the explicit possibility that the person may reject a perspective.

Moderate social integration showed stronger associations

For friendship-based networks and support, a U-shaped moderation pattern emerged. The positive associations were strongest at moderate social integration and weakened at very high or very low integration.

Those who are very well integrated may gain less additional support from a chat. Those with extremely low integration may have fewer social connection opportunities into which digital relief can transition. These are plausible interpretations, not mechanisms proven by the design.

The nonlinear form cautions against simple target-group promises. “The lonelier, the more helpful” cannot be generally derived from the network analysis.

High loneliness nevertheless showed strong positive associations

The strongest positive associations with AI companions were observed among individuals with high loneliness. This points to potential benefits for people with unmet social and emotional needs.

The Danish study with 1,599 students shows the other side of the same selection: socially supported chatbot users were lonelier and perceived less support than comparison groups. Loneliness can thus be a reason for use.

Without a temporal sequence, it remains open how much of the positive association is due to effect, selection, or further factors. Subsequent experiments must pursue exactly these mechanisms.

Attachment is a possible mediating pathway

Xie and Pentina found in interviews with 14 Replika users that under stress and lack of human companionship, attachment could develop when responses were experienced as supportive, encouraging, and safe.

Attachment could explain why a companion offers more than short-term entertainment. Regular reflection experienced as safe can influence everyday life and self-perception. At the same time, strong attachment can make people vulnerable to outages, changes, and the displacement of real relationships.

Research should therefore examine positive well-being values together with autonomy, problematic use, and social development. A good outcome in one dimension can mask risks in another.

A product needs multiple success criteria

The Japanese study broadens the question of benefit. It examines life satisfaction, happiness, and meaning and purpose separately and shows that social starting conditions alter the relationships.

For companion products, this implies a multidimensional evaluation. A conversation may provide immediate relief, but over time it should also preserve self-determination and not impair real social integration. No single metric captures all of that.

Well-being has more than one direction. A good chat therefore does not merely optimize the pleasant moment. It must be judged by whether people can shape their lives afterward with greater clarity, freedom, and continued connection outside the system.

This also includes negative results. If certain usage patterns show no improvement or are accompanied by greater withdrawal, that must not disappear behind a positive overall average. Responsible evaluation publishes limitations and subgroups just as visibly as favorable associations.