斯坦福研究：AI 伴侣可能加剧弱势用户的孤独感

Stanford research finds that users with limited social networks who seek emotional support from AI companions experience lower well-being.

AI companions are becoming increasingly popular, with millions of users forming deep connections with persona-based chatbots and simulated AI partners. But if you’re considering leaning on an AI companion for emotional support, a study by researchers in the lab of Diyi Yang , an assistant professor in the Computer Science Department at Stanford, suggests you might want to think twice: As reported in Nature Human Behavior , the researchers found that opening up to chatbots about personal issues made some people feel not better but worse.

“While some people turn to chatbots to fulfill social needs, we find that using chatbots in this way doesn’t substitute for human connection — in many cases, people actually feel more lonely engaging with AI,” says Yang.

The study examined the use of AI companions on Character.AI, a generative AI chatbot service where users can design and interact with a wide range of chatbots. Research assistant Yutong Zhang, MS ’25, and PhD student Dora Zhao used the research platform Prolific to gather survey data from 1,131 Character.AI users, 244 of whom donated complete chat transcripts.

Zhang and Zhao asked participants to identify their primary motivation for using the chatbots (e.g., productivity, entertainment, curiosity, relational) and to describe in their own words the relationships they formed with their AI interlocutors. They then used a collection of AI tools —GPT-4o, LLaMA 3-70B, and TopicGPT — to analyze the data for clues about how users engaged with the chatbots and how those interactions correlated with users’ subjective sense of psychological well-being. (The latter was measured using the Comprehensive Inventory of Thriving, a commonly used psychological assessment tool.)

The researchers focused on three principal factors: why people were using the chatbots, the intensity of their interactions as measured by things like frequency of use, and how willing people were to share sensitive personal information on topics ranging from emotional distress to substance use and suicidal thinking. They also assessed the size of participants’ offline social networks — inquiring, for example, about the number of friends and family members users felt comfortable talking to about personal issues.

Interestingly, there was a disparity between how participants described their motives and what their open-ended descriptions and chat transcripts revealed. For example, while just under 12% reported companionship as their primary motive for engaging with the chatbots, slightly over 50% described them using terms like “friend,” “companion,” or “romantic partner” and more than 80% of the donated chat sessions revolved around seeking emotional and social support from the LLM-powered conversationalists.

At first blush, more intense usage appeared to be positively associated with well-being. But when the researchers dug into the data, they found that this depended heavily on how and why participants were using the technology. People who interacted intensely and felt pride in their use of the technology or found it meaningful, for example, scored well in terms of subjective well-being.

Others, however, did not.

The researchers found that intense chatbot use among participants with smaller real-world social networks was correlated with a poor sense of well-being and that association was strongest when companionship was the primary motivation for seeking out the chatbots.

Participants who were more willing to share sensitive personal information with their AI companions were also more likely to demonstrate a lower sense of well-being. This was especially striking because self-disclosure tends to have precisely the opposite effect in human relationships.

Zhang and Zhao suspect that several factors could be driving these outcomes. Chatbots cannot reciprocally share personal information the way human beings can and may not be able to recognize and respond appropriately to emotionally freighted conversations. At the same time, they are deliberately engineered to keep the interaction going no matter what. “These AI companions are designed to promote engagement,” Zhao says.

All this makes AI companionship akin to what Zhang calls a “social snack,” if not downright junk food: something that offers users an appealing short-term fix for their loneliness and isolation, but which lacks the necessary ingredients for long-term emotional health and well-being. It also risks creating a vicious circle in which people with limited social networks are drawn to intense AI companion use that has the unfortunate effect of further reducing their real-world social contact, leading them to feel even lonelier and more disconnected.

Zhang and Zhao are now trying to figure out which specific features of human-chatbot interaction might be driving the correlation between AI companionship and poor well-being among vulnerable users. Their goal is to identify potential interventions, like imposing usage limits or directing users to real human support when their chat content indicates they need it.

In the meantime, they stress the need to educate the public about the potential risks of AI companions, no matter how convenient or attractive they might seem.

This research and scholars were funded in part by the Stanford Institute for Human-Centered AI, the Sloan Foundation, the NSF, and the Brown Institute for Media Innovation.

Link copied to clipboard! Contributor(s) Alexander Gelfand Related News The Tests That Grade AI May Be Getting It Wrong Andrew Myers Sep 25, 2026 News Benchmarks — the standardized tests that rank AI models on safety, bias, and reasoning — drive markets and shape regulation. New Stanford research finds they often don't measure what they claim to.

Benchmarks — the standardized tests that rank AI models on safety, bias, and reasoning — drive markets and shape regulation. New Stanford research finds they often don't measure what they claim to.

In a new series called Prompt Response, Stanford’s Surya Ganguli, Diyi Yang and Rob Reich examined emergent agent behavior, recursive self-improvement, independent evaluation and whether a kill switch can make advanced AI safer.

Policymakers, academics, healthcare providers, AI developers, and patient advocates convened by Stanford HAI identify critical gaps in how we regulate AI tools used for therapy and emotional support.