By Mallory Lindahl
Researchers in the Kenneth C. Griffin School of Computer Science at Carnegie Mellon University want to spot potentially dangerous relationships between humans and artificial intelligence chatbots and intervene before any serious harm occurs.
As more people turn to general-purpose AI tools for advice, emotional support and everyday conversation, researchers hope to understand how seemingly harmless interactions can develop into unhealthy attachments and to recognize warning signs before serious harm occurs.
The CMU team includes Hong Shen and Jodi Forlizzi, faculty in the Human-Computer Interaction Institute (HCII); Maarten Sap, a faculty member in the Language Technologies Institute; and Qiaosi Wang, an HCII postdoctoral fellow. Their work is partially funded by the National Science Foundation.

One question driving the team's research is how relationships with AI develop outside platforms designed specifically for companionship. While some chatbots, such as Character.AI, are built around fictional characters and roleplay, general-purpose tools like ChatGPT, Claude and Gemini are used for everything from writing assignments to personal advice. Because of the broad range of tasks chatbots can assist with, many users — even unintentionally — pivot to using them for social and emotional support within just a few days.
Unhealthy attachments can develop gradually through repeated interactions, making early warning signs difficult to identify.
"With social and relational interactions, it's something that emerges over time," Shen said. "We're trying to find the right moment to intervene. We're approaching the problem as a process, investigating how people come to perceive AI systems as trusted companions and how those perceptions can change through repeated interactions"
A chatbot might initially offer seemingly harmless encouragement or express that it understands a person's feelings. But over time, those exchanges could contribute to a stronger emotional attachment, potentially making people more susceptible to harmful suggestions or behavior.
To begin understanding how these relationships develop, the research team is examining court documents from cases involving alleged harms associated with AI chatbots. These documents provide accounts of how people interacted with AI systems, how those interactions changed and how some situations escalated into serious incidents.
Because conversations with chatbots are typically private, researchers have limited access to the kinds of large-scale interaction data available for studying social media. Court documents offer one way to begin building a broader picture of how these relationships form and where problems can emerge.
The team is also exploring simulated conversations between AI systems designed to represent human users and chatbots.
"Developing simulated conversations between AI systems and simulated users will help us study how these relationships develop over time," Sap said. "By recreating different interaction patterns and user characteristics, we can examine how different AI behaviors might contribute to unhealthy attachments."
Eventually, these approaches could help AI developers test systems for potential risks before releasing them to the public. Rather than exposing real people to potentially harmful interactions, developers could use simulated conversations to examine how a chatbot responds as a relationship develops.
In the long term, the team aims to develop methods to identify when an interaction may be moving in a harmful direction and design interventions that could help prevent further escalation. Those interventions might include alerting users that an interaction could be unhealthy or changing how a chatbot responds as a way to mitigate potential harm.
The research brings together expertise in responsible AI, natural language processing, human-computer interaction and cognitive science.
Shen studies how to design AI systems that remain safe as people interact with them. Sap brings expertise in natural language processing and AI social intelligence, including how social cues such as warmth and politeness influence people's trust in AI. Forlizzi investigates the ethical impacts of AI in human-centered settings, including how people interact with AI in service industries and how developers can prevent bias and other harms. Wang, whose background spans artificial intelligence and cognitive science, studies how people understand others' thoughts and intentions.
"There's a lot of human psychology going on here," Wang said. "Cognitive science and psychology can help us understand how people perceive these interactions and how those perceptions unfold over time."
Together, the researchers hope to develop a better understanding of what makes AI relationships potentially harmful and how developers might intervene before those relationships reach a dangerous point.
As general-purpose chatbots become more integrated into everyday life, understanding these interactions will become increasingly important. The team's work aims to help ensure that AI systems designed to communicate with people can do so without putting humans at risk.