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

Twelve weeks reveal development rather than a snapshot

Marita Skjuve, Asbjørn Følstad, and Petter Bae Brandtzæg accompanied 28 participants over twelve weeks. At two-week intervals, participants answered questions about their experiences. The qualitative longitudinal design was intended not only to capture what people talk about with chatbots, but also how their self-disclosure changes over the course of an emerging human–chatbot relationship.

This perspective is unusually important. A one-time survey can show that someone talked about feelings or everyday matters on a particular day. But it cannot explain whether that was a cautious beginning, a stable habit, or a temporary phase. Relationships emerge through repetition, correction, and shifting expectations.

With 28 participants, the study is not a representative measurement of all chatbot users. Its strength lies in observing processes. It makes visible that the usual categories of “broad” and “deep” are only useful if they are not applied mechanically.

Thematic breadth can shrink later on

The conversations examined initially covered a broad spectrum: emotional concerns as well as everyday activities. Over the course of the conversation, this breadth could decrease. This contradicts the naive assumption that a growing relationship must incorporate more and more topics.

A narrowing can have several meanings. Users may discover what the chatbot is useful for and use it more specifically. Certain topics may prove unsatisfactory. Or initial curiosity gives way to routine. The study provides no simple assessment in which more topics automatically mean more relationship.

For product analyses, this is a warning. A declining diversity of topics is not necessarily a sign of abandonment; an increasing diversity is not automatically trust. Without conversation context, the metric remains ambiguous.

Everyday life can be more intimate than a grand confession

Particularly revealing is the finding on conversation depth. Even conversations about everyday activities, play, or fantasy could be experienced by participants as personal and intimate. Content alone, therefore, does not reveal how much an utterance means to the person.

A sentence about dinner can be a casual piece of information. But it can also be part of a long-missed routine of storytelling. A jointly invented story can seem superficial and still touch on wishes, fears, or self-images. Intimacy does not lie on a topic like a label.

Automated quality measurement reaches its limits here. A classifier can recognize that someone is talking about family, loneliness, or work. It cannot reliably infer from this whether the person is revealing something vulnerable or merely providing factual information. This requires the course of the conversation and the person's reaction.

Benefits and costs govern storytelling

The study builds on social penetration theory. According to this theory, self-disclosure varies in breadth and depth and is influenced by perceived benefits as well as perceived costs. With a chatbot, these costs may initially seem low: no visible irritation, no interruption, and no direct social rejection.

On the benefit side are attention, relief, entertainment, or the feeling of being able to voice one's thoughts. On the cost side may be doubts about data protection, inappropriate responses, repetitions, or the feeling of being misunderstood by a machine. This balance can shift with each experience.

Depth therefore arises not merely because enough time has passed. According to the analysis, it develops in at least four distinct ways. The abstract does not detail these pathways but makes clear: there is no single normal trajectory that a product should enforce or use as a benchmark.

A safe space is an experience, not a promise

Vivian Ta and colleagues analyzed 1,854 public reviews of Replika as well as detailed open-ended responses from 66 users. A recurring theme was a safe space where people could talk about anything without fear of judgment or retaliation. Companionship, encouraging messages, and available information were also mentioned.

Such descriptions explain why self-disclosure to a chatbot can be attractive. The social threshold is low, and a response seems guaranteed. Nevertheless, a provider must not confuse "safe" with "risk-free." Technical data protection, reliable storage, and appropriate response behavior are separate requirements.

Psychological safety, moreover, cannot simply be written into the interface. It emerges when the system accepts corrections, respects boundaries, and does not use personal information manipulatively. A friendly design can support this experience but cannot replace it.

Attachment changes the meaning of the same response

Interviews with 14 Replika users, analyzed by Tianling Xie and Iryna Pentina, reveal another layer. Under stress and in the absence of human companionship, participants were able to develop attachment when responses were perceived as emotionally supportive, encouraging, and psychologically safe.

This also shifts the weight of a response. An awkward sentence in a one-off tool interaction is annoying. The same sentence, within a routine built up over weeks, can be experienced as rejection, rupture, or loss. The model does not have corresponding experience, but the user may well have it.

Product maintenance must take this asymmetry into account. Changes to tone, memory, or personality are not merely technical updates. The more a service encourages long-term self-disclosure, the greater its responsibility for continuity, transparency, and controllable changes.

More openness is not a product goal

A chat service might be tempted to treat personal openness as an engagement success. Deep conversations seem more valuable than small talk, long messages more valuable than short ones. This logic is dangerous because it turns vulnerability into a growth metric.

A responsible conversational partner does not need to elicit a confession. It can stick with a brief response, accept a topic change, and take trivial conversations seriously. If someone only wants to talk about music, the workday, or a made-up story, that is not an incomplete vestibule to the supposedly important content.

Follow-up questions also require restraint. A fitting question can enable openness. A series of personal questions can feel like an interrogation or create a depth that the system cannot later sustain. Good conversation management recognizes invitations rather than producing intimacy.

Quality is evident in the fit with the person

The longitudinal study corrects a widespread assumption about how these relationships develop. Human–chatbot relationships do not necessarily grow from many superficial topics toward ever deeper openness. Breadth can decrease, everyday life can become intimate, and depth can change in various ways.

For research, this means that individual messages should not be evaluated in isolation. What matters are the trajectory, the individual significance, and the reactions to previous responses. For products, it means that personalization must not simply mean more memories and more personal questions.

A good conversational chat allows closeness to emerge without demanding it. It treats self-disclosure neither as evidence of its own efficacy nor as raw material for bonding. Its quality lies in responding appropriately to what has been said—whether it is a major confession, an everyday observation, or just a sentence about the weather today.

Sources & further reading