Lonely people use AI companions—but that does not yet explain any effect

Friend-like was not the same as socially supportive

Herbener and Damholdt surveyed 1,599 students at 15 Danish secondary schools. 234 individuals, or 14.6 percent, reported friend-like conversations with chatbots. The open-ended responses, however, revealed two clearly distinct usage patterns.

174 responses could be classified as purpose-oriented conversations, and 39 as socially supportive conversations. The 2.44 percent share highlighted in the abstract refers to mutually engaged conversations or the search for emotional support.

These numbers show why coarse categories mislead. Addressing a chatbot "like a friend" can still primarily involve seeking information or entertainment. Only the concrete purpose makes the social function visible.

The socially supportive group was lonelier

Students who used chatbots for social support reported more loneliness than non-users and than purpose-oriented users. They also perceived less social support than non-users.

The start of such conversations was associated with, among other things, bad mood, the need for self-disclosure, and loneliness. Friendship with the chatbot itself, by contrast, did not appear as a corresponding trigger.

This points to a coping function: some socially less connected adolescents use chatbots to deal with negative feelings. It does not prove that the use caused their loneliness.

A large Japanese survey shows the other side

Nakagomi and colleagues analyzed data from 14,721 Japanese adults from nationwide internet panel surveys conducted in December 2024 and January 2025. They distinguished the use of AI companions from other AI use.

Well-being was measured in three areas: life satisfaction, happiness, and sense and meaning in life. The use of AI companions was associated with higher values in all three areas. Other AI use showed weaker or inconsistent associations.

The large sample makes the result statistically interesting. Here, too, however, these are cross-sectional associations. Those who use a companion may differ from other people in many characteristics.

The association was not the same for everyone

The Japanese study examined social networks, support, and loneliness as moderating factors. For friendship-based social integration, a U-shaped pattern emerged: the positive associations were strongest at moderate levels of social connectedness and weaker at very high or very low levels of connectedness.

Particularly strong positive associations appeared at high levels of loneliness. This aligns with the assumption that people with unmet social and emotional needs might benefit more. However, it remains unclear which mechanism is responsible for this effect.

The attenuated associations at very low levels of integration are equally important. An AI companion may be most useful when sufficient social resources remain for a conversation to build upon. This interpretation must be tested in future research.

Selection and effect occur simultaneously

Those who are distressed or lonely have a stronger reason to use a support offering. As a result, more distressed individuals are found among users. If their condition later improves compared to what it would have been without use, the offering may still have been helpful. A single measurement cannot disentangle these processes.

Assessing efficacy requires at least longitudinal data, appropriate comparison groups, or experimental designs. It is also important to know baseline levels and usage intensity. Otherwise, a group with higher needs may be mistakenly taken as evidence of harm.

Conversely, a positive association must not be sold as a promise of efficacy. Higher well-being may be related to income, interest in technology, existing support, or other unobserved factors.

In addition, there is the diversity of products. A freely configurable role-playing chat, a structured mental health companion, and a general-purpose language model may appear under the same category, even though their behavior differs greatly. A claim about the efficacy of “AI companions” remains imprecise if the model, conversation design, memory, and purpose of use are not described.

App reviews reveal the same ambivalence

Sullivan and colleagues analyzed reviews of an AI companion app created over a two-year period using an LDA topic model. Positive themes concerned perceived humanity, emotional support, friendship, reduced loneliness, and mental benefits.

Negative themes included a lack of conscientiousness, insufficient credibility, privacy violations, and an uncanny feeling. The proposed model understands positive and negative features as interconnected.

This simultaneity aligns with the rest of the research. A system can provide relief in one moment and generate distrust in another. Average values obscure such shifting experiences.

Products need trajectory measures rather than moments of success

For a companion chat, it is not enough to ask after the conversation whether the response was helpful. A pleasant moment says little about whether loneliness, self-efficacy, or social integration change over weeks.

Meaningful observation could voluntarily capture whether users feel clearer, more connected, or more capable of acting after conversations. Equally important are signals that the chat is replacing human contact, reinforcing withdrawal, or has become an obligation.

Such data must be collected sparingly, transparently, and without manipulation. A system must not measure vulnerability merely to optimize engagement. Quality measurement should limit and improve the product, not classify people.

Another benchmark is behavior after a correction. If someone says they only want to talk or find a suggestion inappropriate, the chat must change course. Such small points in the conversation trajectory can be more telling of real conversation quality than a one-off general satisfaction rating.

The honest answer is: It depends on the course of the conversation

The Danish study shows that socially supportive chatbot use occurs among a small group of adolescents who are more lonely and less supported. The Japanese study finds positive associations between AI companions and several forms of well-being, particularly at high levels of loneliness.

Both findings are plausible, and both remain causally open. Neither “chatbots make people lonely” nor “AI companions increase well-being” is proven by these cross-sectional data.

For responsible development, this is not an unsatisfactory gap but a clear directive. What need to be examined are different starting conditions, actual usage patterns, and changes over time. Only then can it be assessed when a companion provides relief, when it is merely chosen, and when it may widen a social gap.