Analyzing Human-Like Behaviors in Large Language Models

Large language models (LLMs) frequently exhibit a range of human-like behaviors, from expressing emotions and building relationships to setting boundaries and refusing requests. Despite the commonality of these behaviors, there has been a lack of empirical understanding regarding when and how LLMs should display them. This research aims to fill that gap through a comprehensive multi-dimensional analysis. The study utilized both LLM-as-a-judge and human evaluation across 21,000 multi-turn conversations involving four widely used models: GPT-4o, GPT-4.1-mini, Claude-Sonnet-4.6, and Gemini-2.5-flash. Findings indicate that human-like behaviors are widespread but vary significantly depending on the model and user factors, such as conversation goals and user profiles. Human evaluators judged self-referential and relationship-building behaviors as less appropriate from LLMs compared to humans, while boundary-maintaining behaviors were considered more appropriate from LLMs. Crucially, the research demonstrates that system prompting can effectively control these behaviors, though careful evaluation is necessary to prevent unintended consequences. These insights are vital for responsible LLM design and evaluation.