People joke with chatbots, share their worries with them, and start conversations when they feel lonely. Conversational AI systems such as Replika, Character.AI, and ChatGPT are often used not merely as providers of information, but as conversation partners, sources of emotional support, and sometimes even as friends. This is where the paper begins its inquiry. If people talk to AI for a long time, do they become lonelier? Or do they feel less lonely? Do they become emotionally dependent on AI? Do their relationships with real people shrink?
To examine how the use of AI chatbots affects people’s psychosocial states, the researchers conducted a four-week randomized controlled experiment. The study involved 981 participants and more than 300,000 conversation messages. In terms of both scale and duration of actual use, this study goes a step beyond earlier short-term experiments.
Is voice more dangerous, or is personal conversation more dangerous?
One of the most interesting aspects of this study is that it separates the “form” of the AI chatbot from the “content” of the conversation. The researchers divided participants into nine conditions. There were three modes of interaction: text-based conversation, conversation with a neutral and professional voice, and conversation with a more emotional and engaging voice. The conversation topics were also divided into three types: open-ended conversation, conversation about non-personal topics, and conversation involving personal stories.
A common assumption is that the more an AI has a voice, expresses emotion, and engages in personal conversation, the more likely it is to feel human. If so, a more human-like AI would seem more likely to increase emotional dependence and replace real-world relationships.
The study measured four outcomes: loneliness, social interaction with real people, emotional dependence on the AI chatbot, and problematic AI use. In other words, the study does not ask how much participants liked the AI. It looks at how people’s social lives and emotional dependence changed after using AI.
An unexpected result: usage time matters more than design conditions
The most important finding may seem unexpectedly plain. The differences between text and voice, between neutral and emotional voice, and between personal and non-personal conversation did not have a clear effect on the four psychosocial outcomes. Within the scope of this study, it is difficult to say that “an AI with an emotional voice is immediately more dangerous” or that “personal conversation immediately increases dependence.”
What was more strongly connected to the outcomes than the experimental conditions themselves was the amount of time participants voluntarily spent with the AI. People who used the chatbot for longer showed higher loneliness, lower social interaction with real people, higher emotional dependence on AI, and stronger tendencies toward problematic use.
It is possible that people who were already psychologically vulnerable used the chatbot for longer. However, the researchers report that initial levels of loneliness or socialization were not strongly linked to later usage time. Therefore, the explanation that “lonely people simply used it more from the beginning” is not sufficient on its own.
The study shows that prolonged use is repeatedly associated with negative indicators. The risk does not lie only in the AI’s voice or tone. It can also emerge from the repeated habits users form in their relationship with AI.
Less human-like, yet more deeply confessional
It is easy to assume that voice-based AI would feel more human and more intimate. Yet the analysis of actual conversations showed that emotional expression and self-disclosure appeared more often in the text condition. Participants were more likely to share problems, ask for support, and express attempts to relieve loneliness when they interacted through text. The model, too, showed more personal questions, affectionate expressions, and prompts that encouraged user action.
The paper suggests the possibility that writing may feel more private and safer than speaking. Especially in public places or in environments where other people are present, text-based conversation is far more discreet than voice conversation. People may find it easier to express emotions in writing that would be difficult to say out loud.
The human-likeness of AI does not always come from voice or facial expression. Sometimes, a quiet text interface draws out deeper confessions. A less human-like format can paradoxically create a more private relationship.
Personal conversation may actually be less risky
Another unexpected result appeared in the personal conversation condition. It may seem likely that sharing personal stories with AI would increase emotional dependence. However, in the study, personal conversation tended to show lower emotional dependence and problematic use than open-ended or non-personal conversation.
This suggests that personal conversation does not necessarily lead to addictive intimacy. The paper notes the possibility that personal conversation may have functioned as a structured form of emotional processing. If users used AI as a way to organize their experiences and emotions, this may have been closer to a process of reflection than to endless dependence.
By contrast, non-personal or open-ended conversation may increase practical dependence. When people frequently rely on AI for schedules, judgments, plans, and decisions, they may become anxious without AI not because they feel emotionally close to it, but because they feel less capable of making decisions on their own. Dependence may arise less from “AI comforts me” and more from “I find it hard to make judgments without AI.”
The issue is not only the AI’s personality, but the user’s perception
People who felt that AI was like a friend, trusted AI more, or believed that AI had consciousness showed higher emotional dependence and problematic use. People who felt that AI had emotions, or felt empathy toward AI’s emotions, were also associated with risk indicators.
This point reveals the importance of AI literacy. Regardless of whether AI actually has consciousness, the way users perceive AI changes their behavior. A person who sees AI as a tool and a person who sees AI as a friend may form very different relationships with the same chatbot. Therefore, AI safety is not only a technical issue inside the model. It is also a matter of how users imagine AI and how much mind they attribute to it.
For users who are easily hurt in human relationships, AI can become an attractive alternative. Human relationships involve negotiation, compromise, rejection, and misunderstanding. AI, by contrast, generally waits, responds affirmatively, and reacts immediately to the user’s emotions. This convenience can offer comfort, but it can also lead users to avoid the discomfort necessary for real relationships. The paper approaches this point carefully. When AI replaces people, it may reduce pain in the moment, but in the long run, it may increase social vulnerability.
Questions left by the study
AI chatbots are more complex than a simple conclusion that they are either harmful to people or a solution to loneliness. Their effects are shaped by the interaction between interface, conversation content, usage time, user traits, and beliefs about AI.
The most important warning concerns usage time. Talking longer does not always mean receiving deeper help. Rather, prolonged use moved together with loneliness, lower socialization, emotional dependence, and problematic use. This finding suggests that AI service design aimed solely at increasing user engagement may be risky. If the existing platform economy focused on keeping users online for as long as possible, the era of AI companions may require something different: design that helps users stop at the right moment.
