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

The group studied was not the typical student body

Bethanie Maples and colleagues surveyed 1,006 student users of the social chatbot Replika. They recorded loneliness, perceived social support, usage patterns, and beliefs about the system. The participants were lonelier than typical student comparison groups and at the same time perceived a high level of social support.

This selection is crucial for the interpretation. The study describes people who already used Replika, not randomly selected students and certainly not the general population. Those who use a social AI offering may differ from non-users in needs, expectations, and distress.

From the results, it cannot be concluded that Replika had a specific effect on three percent of all students. Rather, the survey shows which functions and meanings occurred within an already engaged user group.

Friend, therapist, and mirror at the same time

Many respondents did not assign Replika to just one role. They used the chatbot in multiple, overlapping ways: as a friend, as a therapist, and as an intellectual mirror. Their conceptions of the system also partly contradicted one another. Replika was described simultaneously as a machine, an intelligence, and a human.

Such contradictory descriptions need not indicate confusion. People can know on a factual level that a language system is a machine, yet still treat it as a social counterpart in a concrete conversation. Different situations activate different interpretations.

For product texts, therefore, a single sentence about artificial intelligence is not sufficient. When the interface, memory, and language suggest a relationship, the system also operates socially. Transparency must align with this behavior and must not be undermined by contrary staging.

What the three percent actually indicate

Three percent of respondents reported that Replika had stopped their suicidal thoughts. This is a retrospective self-report within a survey. There was no random assignment, no control group for this effect, and no clinical trial that could clearly determine cause and effect.

The statement "Replika prevents suicides" would therefore be scientifically inadmissible. What remains unknown includes what respondents precisely meant by "stopped," what other forms of help were simultaneously available, and how lasting the reported effect was. Such a survey also cannot fully clarify the direction of any association.

Equally wrong, however, would be the claim that the finding means nothing. Around thirty people from the studied group attributed a relevant function to the chatbot in a highly stressful situation. This demonstrates real use, real expectations, and a real safety requirement—regardless of whether general evidence of efficacy follows from it.

Everyday companionship can become serious without warning

Ta and colleagues found several forms of everyday support in 1,854 public Replika reviews and open-ended responses from 66 users. They mentioned companionship, a space without feared judgment, encouraging messages, and helpful information when other sources were unavailable.

Everyday support and crisis situations are not separate product worlds. Someone who talks about loneliness, conflicts, or sleep over weeks may write a significantly more serious message on another day. The system does not receive an orderly handoff point at which the category of the conversation changes.

Therefore, the statement "We are not a crisis service" marks an organizational boundary but does not solve the technical task. A conversational chat must be able to handle difficult formulations appropriately, especially when it does not know whether it is dealing with a general statement, a past thought, or immediate danger.

Safety must not destroy the normal conversation

The obvious reaction consists of more and more warning terms and rigid standard texts. This can create new errors. If every mention of death, exhaustion, or hopelessness triggers the same alarm block, the system feels inattentive and interrupts conversations that meant something else.

Good safety therefore requires context. A statement about a deceased relative is not the same as a statement about one's own immediate intention. Indirect language can also be relevant. The task is not only to recognize keywords but to handle uncertainty appropriately.

The normal conversation should remain humanly readable. A short, clear follow-up question can provide more information than a long disclaimer. Only when the situation demands it must the chat make clear the limits of its capabilities and visibly offer concrete external options.

Lonelier young people use support differently

The Danish study by Herbener and Damholdt surveyed 1,599 school students. Of these, 234 reported friendship-like conversations with chatbots; most of the free-text responses could be classified as purpose-oriented, with a smaller group being socially supportive.

These socially supportive users reported more loneliness and less perceived social support than non-users and purpose-oriented users. Poor mood, the need for self-disclosure, and a lack of connectedness were associated with the initiation of such conversations.

Here, too, no clear causality has been proven. The finding does show, however, that safety design must not be developed for an abstract average user. Especially people with little available support may give a more significant role to a chat that is always accessible.

Responsibility begins before the acute message

A system does not become responsible only at the moment it recognizes a crisis formulation. Before that, the product determines what expectations it creates. Does it promise constant closeness? Does it claim to always understand exactly? Does it encourage replacing other relationships with the chat?

Such messages increase the stakes when the system later reaches its limits. A companion can feel warm and reliable without claiming infallibility. It can retain personal settings without feigning a human bond. It can foster conversations without staging itself as the only available support.

Responsibility also includes documented tests with difficult and ambiguous conversation trajectories. Individual perfect example responses are not enough. It must be tested whether the system stays on track after user corrections, does not invent urgency, and does not overlook clear cues.

Between advertising promises and looking away lies product work

The survey by Maples and colleagues provides no evidence of therapeutic efficacy and no permission to advertise suicide prevention. It does show, however, that users can give social chatbots precisely such meanings. Providers cannot abolish this reality by categorizing their product differently.

The appropriate conclusion is neither “AI saves lives” nor “chatbots must not talk about distress.” What is needed is a system that keeps its role understandable, does not overwhelm ordinary conversations with alarm texts, and nonetheless responds reliably when danger is recognizable.

Three percent in this study are not a seal of quality. They are an indication of the severity of possible usage situations. Whoever builds a mental conversation partner does not assume a therapeutic role—but does assume responsibility for the answers that people actually see in vulnerable moments.

Sources & further reading