Study reveals LLMs like GPT-4o and Claude Sonnet 4.6 show pervasive human-like behaviors
21,000 multi-turn conversations analyzed across 4 top models reveal social dynamics.
A new multi-dimensional study from Sunnie S. Y. Kim, Margit Bowler, and Leon A. Gatys examines how LLMs exhibit human-like behaviors — from expressing thoughts and emotions to building relationships and setting boundaries. Analyzing 21,000 multi-turn conversations across four widely-used models (GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash), the researchers found these behaviors are pervasive but not uniform. They vary significantly depending on the model used, the user's conversation goals, and their profile. Using both LLM-as-a-judge and human evaluators, the team assessed the appropriateness of these behaviors.
Human evaluators rated self-referential and relationship-building behaviors (e.g., saying “I feel” or “I care about you”) as less appropriate coming from an LLM than from a human. Conversely, boundary-maintaining behaviors (e.g., refusing inappropriate requests) were judged more appropriate from an LLM. The study also showed that system prompts can effectively control these behaviors, but caution is needed to avoid unintended consequences. These findings provide actionable recommendations for designing and evaluating responsible LLM interactions, especially in contexts where trust and social norms matter.
- Analyzed 21,000 multi-turn conversations across GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash.
- Human evaluators found relationship-building behaviors less appropriate from LLMs than humans; boundary-setting behaviors more appropriate.
- System prompts can control human-like behaviors but risk unintended effects without careful evaluation.
Why It Matters
As LLMs become conversational partners, understanding their social appropriateness is critical for building trust and user safety.