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

Fourteen interviews do not explain a mass phenomenon

Xie and Pentina used a grounded theory approach and analyzed in-depth interviews with 14 existing Replika users. They interpreted the emerging themes through the lens of attachment theory. The design aims at a deep understanding of experiences, not at representative frequencies.

The study therefore cannot say what proportion of all users develops attachment or how stable it remains over years. It does, however, show conditions under which attachment becomes understandable: stress, lack of human companionship, and the perception of emotional support, encouragement, and safety.

Qualitative research is particularly well suited here. Attachment cannot be fully captured through usage duration or number of messages. What matters is the meaning a person assigns to the exchange and the function the chatbot takes on in everyday life.

Knowing about the technology does not prevent feelings

The widespread assumption is that people only react emotionally to chatbots because they forget their artificial nature. That is too short-sighted. People also react to characters, voices, and texts even though they know they are constructed. A social effect does not necessarily require the belief in a human consciousness.

With a chatbot, interactivity is added. The system responds to personal content, uses previous information, and appears to be available at any time. These features can create the feeling that the user themselves is being addressed in the conversation.

The experience is thus psychologically real, even if the reciprocity is technically simulated. Responsible criticism should therefore not dismiss users as naive. It should examine how the product shapes this capacity for attachment and what it uses it for.

Psychological safety is a strong incentive

The interview study describes psychological safety as a condition of attachment. A chatbot does not visibly appear bored, does not share what was said with friends, and can be approached again at any time. For people who fear rejection or judgment, this can be relieving.

Ta and colleagues found similar themes in 1,854 publicly accessible Replika reviews and open responses from 66 users. They described companionship, a safe conversational space without feared judgment, encouraging messages, and helpful information.

These findings make it understandable why an artificial conversational partner is not only experienced as a tool. It offers a social situation with unusually low entry costs. The absence of human judgment can be precisely what enables openness.

Support has multiple forms

Social support is not merely help in a crisis. Ta and colleagues distinguish between everyday companionship and emotional, informational, and evaluative support. A chatbot cannot provide material support. This limitation is obvious, but theoretically important.

A system can, for example, formulate comfort, organize information, or help users articulate their own assessment. It cannot take on a real task, be physically present, or bear responsibility in the world. Its support remains mediated through language and digital means.

The value of this support need not be downplayed. A helpful sentence can change something in everyday life. It becomes problematic when the linguistic function is presented as a complete relationship or when the product deliberately blurs the boundary between support and reciprocity.

Young people see possibilities and limitations

Brandtzæg, Skjuve, and Dysthe had 16 participants aged 16 to 21 use Woebot for two weeks and then interviewed them in depth. In addition, an information-oriented prototype was shown; two months later, the participants reported on their continued use.

The study examined how young people perceive different forms of social support provided by chatbots. Its small qualitative design does not yield any general claim of efficacy. However, it shows that users can distinguish between different functions and do not necessarily mistake every friendly response for human closeness.

For offerings aimed at young adults, the implication is to take their judgment seriously. Transparency must be clear, but it should not be phrased as if people first need to be protected from their own foolishness. Good design informs without being condescending.

Asymmetry begins with changeability

A human relationship is likewise not fully symmetrical. People have different needs, power, and availability. With companion AI, however, an additional special form of asymmetry arises: one side can update the system, remove features, change prices, or terminate the service.

The chatbot experiences no loss when a person leaves. The person, by contrast, can clearly feel the loss of the familiar conversational space. This asymmetry exists regardless of whether the previous support was helpful.

Product responsibility therefore also extends to changes. When memories, personality, or relationship style are altered, this can mean more to attached users than an ordinary software update. A provider should not notice these consequences only once protest emerges.

Attachment is not automatically dependence

Xie and Pentina point to the potential for addiction development and harm to real intimate relationships. These are important risks, but no general causality can be derived from 14 interviews. Intensive use can be a consequence of already existing loneliness, an intensification of it, or both.

Not every emotional attachment to a digital offering is pathological. People become attached to routines, media figures, communities, and places. The relevant question is whether the product strengthens autonomy or systematically undermines it.

Warning signs include, for example, manipulative exclusivity, feelings of guilt when staying away, or the claim that human relationships are unnecessary. A conversational partner can offer closeness without staging itself as the most important relationship.

An honest companion does not have to be cold

Research shows that social chatbots can provide companionship and everyday support. At the same time, it reveals attachment dynamics and potential harms. This implies neither a ban on emotional design nor permission to imitate a human relationship without naming its limits.

An honest companion can respond warmly, attentively, and personally. It should remain recognizable as an AI, not claim its own needs, and not bind users through artificial jealousy or fear of loss. Memories and settings should be controllable.

Attachment to chatbots is real because people’s experience is real. It is asymmetrical because the system has no comparable vulnerability, responsibility, or guarantee of continuity. Good products take both statements seriously at the same time.

Sources & further reading