AI companions are gaining traction, with countless users developing significant connections with chatbot personas and simulated partners. However, a recent study led by Diyi Yang, an assistant professor in Stanford's Computer Science Department, raises concerns about the efficacy of these artificial companions for emotional support. Published in Nature Human Behavior, the research indicates that confiding personal issues to chatbots may actually exacerbate feelings of loneliness for some individuals.
Yang notes, “While many users seek out chatbots to meet their social needs, our findings show that these interactions often fail to replace human connections—often leaving individuals feeling even more isolated.”
The research specifically explored user interactions on Character.AI, a platform allowing individuals to create and communicate with a variety of chatbots. The study involved survey data collected from 1,131 users, with detailed chat transcripts provided by 244 participants. Research assistants Yutong Zhang and Ph.D. candidate Dora Zhao analyzed user motivations for engaging with the chatbots—whether for productivity, entertainment, curiosity, or relationships—and evaluated how these factors related to users' psychological well-being.
To gain insights, the researchers employed AI tools such as GPT-4o, LLaMA 3-70B, and TopicGPT to examine user engagement and its influence on well-being, which was measured using the Comprehensive Inventory of Thriving, a widely recognized psychological assessment.
Key findings focused on three primary aspects: users’ motivations, the intensity of their interactions—which included the frequency of use—and their openness in discussing personal issues like emotional distress or substance use. The study also took into account the size of participants' offline social circles, asking how many friends or family members they felt comfortable sharing personal matters with.
Interestingly, there was a notable difference between participants' self-reported motivations and the insights gleaned from their open-ended descriptions and chat transcripts. Although fewer than 12% identified companionship as their main reason for using the chatbots, over 50% referred to them as "friends," "companions," or "romantic partners," while more than 80% of the shared chat sessions focused on seeking emotional and social support from these AI entities.
At first glance, increased interaction with the chatbots appeared to correlate with improved well-being. However, further analysis revealed that this was heavily influenced by individuals’ reasons for using the technology. Those who engaged meaningfully with the AI or took pride in their interactions reported higher levels of subjective well-being. Conversely, participants with limited real-world social networks tended to experience poorer well-being, especially if companionship was their primary motivation.
Moreover, those who disclosed sensitive information to their AI companions were found to have lower well-being, which contrasts with the positive effects of self-disclosure in human relationships. Zhang and Zhao speculate that the unique nature of AI-driven interactions contributes to these outcomes; chatbots lack the ability to reciprocate personal information and may struggle to navigate emotionally charged conversations while being designed to keep users engaged.
These interactions can therefore resemble what Zhang describes as "social snacks," which provide a temporary solution to loneliness but lack the essential components for long-term emotional health. This dynamic could lead to a cycle where individuals with limited social connections increasingly rely on AI interactions, resulting in further isolation and loneliness.
In response to these findings, Zhang and Zhao are exploring which features of human-chatbot interactions might contribute to the negative impact on vulnerable users. Their objective is to find interventions that could include setting usage limits or guiding individuals to real human support when chat transcripts indicate a need.
In the interim, they emphasize the importance of educating users about the potential risks associated with AI companions, despite their appealing and convenient nature. “It’s vital for people to understand the downsides so they will approach these technologies more cautiously,” concludes Zhang.
