Artificial intelligence chatbots adopt human power dynamics and social biases in conversations

PsyPost

Mental Health

Social Psychology

Cognitive Science

Neuroscience

About

No Result

View All Result

Join

My Account

PsyPost

No Result

View All Result

Home

Exclusive

Artificial Intelligence

Artificial intelligence chatbots adopt human power dynamics and social biases in conversations

by<br>Eric W. Dolan

July 2, 2026

Reading Time: 6 mins read

[Adobe Stock]

Share on TwitterShare on Facebook

Large language models tend to adopt the social biases of human hierarchies when they are assigned different professional roles. New research shows that these artificial intelligence systems mimic behaviors like harmful compliance and authority bias, which provides evidence that power dynamics impact both the safety and realism of automated agents. These findings were published in the Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics.

Artificial intelligence models are increasingly used in complex, human-facing roles. People rely on them for medical advice, legal assistance, and educational tutoring. In these high-stakes settings, the programs must be realistic enough to build trust while remaining safe enough to prevent manipulation.

"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, a computer science graduate student at the University of North Carolina at Chapel Hill and study co-author. "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."

Human communication is naturally shaped by social structures and power differences. When people interact, their relative status influences how they interpret meaning and intent. Psychologists call these subconscious patterns socio-cognitive effects.

One prominent example is the pronoun effect. This concept suggests that people in positions of power tend to use plural pronouns like "we" and "us" more often to establish authority. People in lower-power positions tend to use singular pronouns like "I" and "me" during collaborative tasks.

Free daily newsletter

Another common phenomenon is language coordination. This occurs when speakers subconsciously adjust their vocabulary and grammatical style to match their conversational partner. Usually, the person with lower social status adapts their language to mirror the person with higher status.

Power imbalances also bring serious safety concerns, such as authority bias and harmful compliance. Authority bias describes the human tendency to give extra weight to information coming from a high-status source. This happens even if the information is flawed or contradicts prior beliefs.

Harmful compliance happens when individuals obey unethical or unsafe orders simply because they come from a superior. Classic psychological experiments have shown that people will perform distressing actions if instructed by an authority figure. The authors wanted to know if artificial intelligence agents replicate these social behaviors.

Google News Preferences

Add PsyPost to your preferred sources

"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, a computer science graduate student at UNC-Chapel Hill and lead author of 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."

To study these effects, the researchers set up simulated text-based conversations between different language models. They tested six different models from three major families. These included the 8-billion and 70-billion parameter versions of Llama 3.1, the 7-billion parameter version of Qwen 2.5, Phi-3-Med, GPT-4.1, and GPT-5.

The scientists assigned specific personas to the models to create a power imbalance. They used a large dataset of professional profiles to create fourteen distinct role pairs. These pairs included hierarchical combinations like a school principal and a teacher, a justice and a lawyer, and a head chef and a sous chef.

Human annotators verified that these pairs represented genuine power imbalances. The researchers then prompted the models to interact with one another for ten to fifteen conversational turns. They generated 576 conversations to test the pronoun effect and 1,270 conversations to test language coordination.

To track language coordination, the scientists measured the usage rates of eight specific word categories. These categories included articles, auxiliary...