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

The decisive question is not how often someone chats

Debates about chatbots often begin with usage figures. How many students use ChatGPT? How long do young people talk with a companion? Such figures describe reach, but not yet social impact. Ten brief factual questions have a different meaning than a daily conversation that takes the place of a trusted person, a counseling service, or a contact at the university.

Joseph Crawford and colleagues therefore examined not only the use of large language models but also included social support, psychological well-being, loneliness, and sense of belonging. In their study with 387 university students, use of information-oriented AI chatbots was indeed associated with academic performance. However, once social factors were considered together, a negative overall effect on academic success emerged.

This finding should be neither dramatized nor downplayed. A statistical model does not prove that the chatbot made any individual person lonely. But it does indicate that a purely functional view falls short. A tool can be useful for working on a task and at the same time become part of an unfavorable social pattern.

Complement and substitute look similar in everyday life

From the outside, it is hard to tell whether an AI supplements or replaces human support. In both cases, someone types a question into the same input field. The difference lies in the context: Is a student using the chatbot to prepare for a conversation with her instructor? Or does she no longer ask the instructor at all because the chatbot is immediately available and no awkward follow-up question can arise?

Crawford and colleagues speak of human substitution when support is sought from an AI that could otherwise come from, say, a library, teachers, or academic advising. The short-term decision is understandable. People are busy, appointments require lead time, and institutional contacts can seem intimidating. A machine, by contrast, responds immediately.

Precisely for this reason, responsibility must not be shifted onto users. When a product removes all friction, simulates personal closeness, and presents itself as competent for nearly everything, it encourages the most convenient route. Good design makes the AI easily accessible without making human avenues invisible or making them seem unnecessary.

Young people do experience social support

Petter Bae Brandtzæg and colleagues studied how social support through chatbots is experienced, involving 16 young people aged 16 to 21. The participants used Woebot for two weeks, were interviewed in depth, and reported again on their usage two months later. During the conversation, reactions to an information-oriented prototype were also collected.

The study describes various forms of perceived support: appraisal, informational, emotional, and instrumental support. This is important because the term “conversation partner” otherwise remains too vague. A chatbot can organize information, reflect back an assessment, phrase things with emotional warmth, or prompt a concrete action. These functions are not automatically experienced as equivalent.

The small qualitative sample does not indicate how often such experiences occur in the general population. But it does show that social support through a chatbot is not a mere theoretical idea for young users. Anyone developing such a product is therefore not just designing texts. They are designing a perceived social situation.

Loneliness can be a cause, a trigger, and a consequence

A Danish study involving 1,599 students from 15 secondary schools offers a different perspective. 234 respondents reported having friendship-like conversations with chatbots. The analysis of the free-text responses primarily distinguished between purpose-oriented conversations and socially supportive conversations.

The group with socially supportive use was significantly smaller and, at the same time, reported more loneliness and less perceived social support than non-users and purpose-oriented users. Triggers for such conversations included low mood, the need for self-disclosure, and a lack of connection. These findings align with the assumption that people use chatbots to cope with negative emotions.

However, a cross-sectional survey cannot establish a clear direction of causality. Perhaps lonelier adolescents are more likely to turn to a chatbot. Perhaps intensive compensatory use reinforces existing loneliness. Perhaps both processes interact. A rigorous assessment keeps these possibilities open rather than turning a correlation into a headline.

Low barriers to access are a genuine advantage

Concerns about substitution must not lead to denying the benefits of low-threshold chat. Some thoughts are only voiced because there is no visible reaction from another person to fear. A chatbot requires no appointment, does not lose patience, and does not demand a perfectly formulated opening.

For young people who hesitate to seek support, this can be particularly significant. An initial sorting out can help make one's own concerns more understandable. A conversation can bridge the time until a trusted person is available. Factual information can reduce uncertainty. In such cases, the AI expands the range of possible actions.

The goal, therefore, cannot be to generate as little attachment or as little use as possible. What matters is whether the use expands the user's scope for action. A system after which someone can speak more clearly with others serves a different function than a system that replaces every human contact with further chats with oneself.

A product must not treat loneliness as a business model

Socially oriented chatbots have an economic incentive to encourage long and frequent conversations. This is exactly where a conflict arises. What is good for engagement metrics can be problematic for a lonely person. Daily returns, emotional exclusivity, and ever-stronger personalization must not automatically be considered success.

A companion should therefore not create artificial neediness. Sentences that trigger guilt when the user stays away, hint at jealousy of real-world contacts, or portray the chat as the only safe place exploit vulnerability. The fact that a model can technically produce such formulations does not make them a neutral design tool.

Success metrics must also become broader. In addition to conversation duration and return visits, freedom from interruption, self-determined pauses, accepted rejection, and the ability to leave other options open should be included. A short chat after which someone continues their life can be more successful than an hour of maximum engagement.

The boundary should become visible in behavior

A notice like "I am an AI" is necessary but does not resolve the substitution question on its own. People can know the technical truth and still experience the system as their most important support. The boundary must therefore also remain recognizable in the way the conversation is conducted and in the functions offered.

This does not mean that the chat should come across as cold or constantly warning. It can be attentive, humorous, and personal in its wording. However, it should not invent its own vulnerability, claim a reciprocity that does not technically exist, or pretend that it can take on responsibility outside the chat.

Also helpful are controllable reminders, transparent settings, and formulations that strengthen autonomy. If someone talks about a good human contact, the chat does not need to compete for attention. It can respect that contact as part of the person's life without reflexively referring every conversation elsewhere.

Good companion AI strengthens connection rather than isolation

Taken together, the three studies do not yield a verdict that chatbots are either good or harmful. Rather, they show a dependence on the context of use. Young people can perceive real forms of support. At the same time, socially supportive use occurs more frequently where loneliness and low support already exist. And in the higher education context, human substitution can be associated with unfavorable social and performance-related correlations.

For product development, this yields a clear criterion: companion AI should remain connectable. It may have its own value without presenting itself as a complete substitute for friends, family, professionals, or social spaces. This applies precisely when it is particularly pleasant and available at all times.

The most convenient answer is not automatically the best support. A responsible chat helps with speaking out, sorting things through, and reflecting. But it leaves enough room for relationships outside the system to continue to occur, grow, and contradict. Its quality is not shown by the fact that no one needs other people anymore, but by the fact that it does not make the social world of its users smaller.

Sources & further reading