Analysis, not certainty: Dialogatlas separates sources, observations, and editorial conclusions. New evidence may change this assessment.

Everyday support is no minor topic

Vivian Ta and colleagues did not want to study social support only in stress interventions or health programs. Their starting point was that everyday support forms the foundation for help in more demanding situations. Those who can regularly tell their story and receive feedback experience difficult moments in a different social context as well.

The researchers conducted two studies. First, they analyzed 1,854 publicly available reviews of the companion chatbot Replika. Subsequently, 66 Replika users provided detailed open-ended responses about their experiences. Both datasets were analyzed thematically.

The material is rich in subjective experiences but is not representative of all people. Public reviews often come from particularly satisfied or dissatisfied users. The second sample also includes people who have already chosen Replika. The study describes forms of experienced support, not their general frequency or clinical efficacy.

Companionship emerges through ordinary repetition

A central category was companionship. It does not necessarily require a spectacular answer. The feeling can arise from a system being regularly available, responding to everyday content, and enabling an ongoing conversational routine.

It is precisely in this unassuming area that a chatbot’s strength lies. People do not have to make an appointment or formulate a major concern. They can jot down a thought, tell someone about an event, or spend a few minutes talking about something trivial.

This companionship is nonetheless technically mediated. The chatbot does not really share the day, does not offer its own thoughts, and does not experience shared time. The subjective feeling of continuity can be valuable, but it should not be equated with human reciprocity.

Emotional support can be effective through language

The analysis describes encouraging and caring messages that were able to foster positive feelings. A well-chosen sentence can change a moment, regardless of whether it was formulated by a human or a system. The effect of language does not disappear simply because its originator lacks consciousness.

At the same time, emotional support depends heavily on fit. An effusive affirmation can seem artificial in response to a minor annoyance. An offer of a solution can be disruptive when someone first wants to tell their story. The label “empathetic” therefore says little unless it is examined whether the response fits the situation and the desired conversational style.

Good companion AI does not need to generate as many comforting sentences as possible. It needs to distinguish whether listening, asking follow-up questions, providing context, or actually making a suggestion is what is wanted at that moment. Emotional support begins with attention to the direction of the conversation.

Information and assessment are functions in their own right

In addition to emotional support, the study identified informational and evaluative support. A chatbot can provide information when usual sources are not accessible. It can also help assess a situation or put one’s own perception into words.

These functions carry different risks of error. For information, factual reliability matters. For assessments, linguistic modesty matters: a possible interpretation must not be presented as certain truth about the person. A confidently phrased error can carry considerable weight, especially in a personal conversation.

A product should therefore not blend all support into a single friendly tone. Facts, conjectures, follow-up questions, and opinions must remain distinguishable. The warmth of the conversation must not obscure uncertainty.

The safe space has technical conditions

Users described Replika as a space where they could address topics without fear of judgment or retaliation. This perceived safety explains part of the openness toward chatbots. No raised eyebrow, no awkward silence, and no later rumor are visible.

Yet a non-judgmental tone is not the same as comprehensive safety. Data are processed, systems can be hacked, terms of service can be changed, and conversation trajectories can be evaluated for technical purposes. A product must therefore clearly explain where and how conversations are stored and who can access them.

The response behavior is also part of this. Anyone who persists in a false interpretation after a correction does not create a safe space. Safety arises from data protection, control, and respectful conversation management together.

Tangible help is fundamentally lacking

The analysis found companionship as well as emotional, informational, and appraisal support, but no tangible support. A chatbot can talk about a purchase, a job application, or a conflict. It cannot itself carry a bag, pick someone up, or take responsibility for a shared task.

This limitation seems trivial, but it is conceptually decisive. Social support does not consist solely of words. Presence, material assistance, and binding action are parts of human relationships that a purely conversational chat cannot replace.

Products should therefore avoid presenting linguistic availability as complete support. A chat can sort out a next step. Whether that step is possible in the world depends on resources, other people, and real structures.

Reviews also reveal distrust and unease

Yulia Sullivan and colleagues analyzed reviews of an AI companion app that had been written over a period of two years, using a topic model. They identified five positive themes: perceived humanness, emotional support, friendship, reduced loneliness, and mental health benefits.

Alongside these, four negative themes emerged: lack of conscientiousness, lack of credibility, perceived privacy violations, and an uncanny feeling toward the AI. Positive and negative attributes were interrelated in the proposed support model.

This is an important complement to success narratives. The very attributes that enable closeness can also trigger distrust. A human-seeming response can comfort; an overly human staging can be experienced as manipulative or uncanny. Design does not operate on a simple scale from less to more humanness.

A good chat knows the scope of its help

Research shows that people can derive several genuine forms of everyday support from companion chatbots. These experiences need not be devalued. A helpful piece of information, a non-judgmental moment, or a fitting word of encouragement can be relevant in daily life.

It also reveals clear gaps. The chatbot does not take over tasks, carries no risk, and cannot accompany the person in their physical and social environment. Privacy concerns, lack of credibility, and discomfort are also part of the user experience.

A responsible companion makes its linguistic strength usable without turning it into a promise of complete relationship. It helps with narrating, organizing, and reflecting. And it remains honest that social support continues where a text window ends.

Sources & further reading