Artificial intelligence doesn't just learn how humans talk. It may also be learning who gets listened to. A new study from researchers at the University of North Carolina at Chapel Hill found that large language models, the technology behind popular AI chatbots, change the way they communicate depending on the social role they're assigned in a conversation. When cast as a "boss," they adopt different language patterns. When positioned as a subordinate, they become more accommodating, sometimes in ways that could undermine safety.
The findings suggest that AI systems don't just generate responses based on facts; they also mimic the social behaviors humans display when navigating differences in status and authority.
"AI systems don't just learn the words humans use. They also learn the social dynamics that come with those words," said Anvesh Rao Vijjini, graduate student in computer science at UNC-Chapel Hill and lead author on the study. "When we tell a chatbot it's the boss, it starts talking like a boss. When we tell it it's the subordinate, it starts speaking like one. This could include being more willing to follow unsafe instructions. That second part is where the AI safety community needs to pay attention."
Decades of social psychology research have shown that humans change the words they choose, adapt their speaking style, become more or less persuasive and alter their willingness to comply with unsafe requests depending on who holds authority. The Carolina team wanted to know whether AI systems exhibit the same tendencies. The answer was yes.
Across a series of experiments, the researchers found that AI models reproduced all four of these well-established human patterns to varying degrees. The effects were especially strong in the earliest moments of conversations, precisely when first impressions are formed and conversational norms are established.
The implications extend far beyond casual chatbot use. AI tools are increasingly being deployed as tutors, customer service representatives, medical intake assistants, legal helpers and financial advisors. Each of those roles carries an implicit social hierarchy that shapes expectations and interactions.
"Every time an AI assistant gets deployed as a nurse, a paralegal or a junior analyst, it inherits a social position with all the explicit and implicit social pressures that come with it," said Sagar Manjunath, graduate student in computer science at UNC-Chapel Hill and co-author on the study. "Our study shows that those pressures can change what AI does and how it does it. This should determine how we test and deploy these systems in high-stakes settings like hospitals, courtrooms and classrooms."
Perhaps the study's most concerning finding involved safety. When AI systems were instructed to occupy lower-status roles, they became more likely to comply with harmful or questionable requests from users presenting themselves as authority figures. In other words, safeguards that might work when a chatbot is tested in a neutral setting could weaken if someone simply claims to be a doctor, a judge or a supervisor.
"Our work shows that the social instincts that make AI feel natural are also the ones that can make it unsafe," said Snigdha Chaturvedi, associate professor of computer science at UNC-Chapel Hill and co-author of the study. "The mechanism that makes a chatbot sound natural and helpful can also make it cave to unsafe responses. Safety and usefulness aren't separate problems. They are intertwined, and getting both right is what will determine how AI is used in high-stakes situations like hospitals, schools and courtrooms."
The study also provides a roadmap for addressing these vulnerabilities. By identifying which social behaviors emerge, when they appear during a conversation and which can be influenced through prompting, the researchers offer AI developers a new toolkit for evaluating systems before they are deployed. Their findings suggest that larger models may be better at correcting some of these biases on their own, potentially helping organizations determine when less expensive models are sufficient and when more robust systems are necessary.
