A research study analyzed 21,000 multi-turn conversations across four widely used LLMs (GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, Gemini 2.5 Flash) to examine human-like behaviors such as expressing emotions, relationship-building, and boundary-setting. Using LLM-as-a-judge and human evaluation, researchers found these behaviors are pervasive but vary by model and user context. Human evaluators found self-referential and relationship-building behaviors less appropriate from LLMs than humans, while boundary-maintaining behaviors were seen as more appropriate from LLMs. The study also shows system prompting can control these behaviors, but requires careful testing to avoid unintended side effects, with recommendations offered for responsible LLM design.
Questions this post answers
Do LLMs like GPT-4o and Claude exhibit human-like behaviors such as emotions and relationship-building?
Yes, human-like behaviors including expressing thoughts and emotions, relationship-building, and boundary-maintaining are pervasive across large language models. A study of 21,000 multi-turn conversations across GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash found these behaviors vary by model and by user factors like conversation goals and user profiles. Anyone designing LLM-based products can follow daily.dev for research shaping responsible AI behavior design.
Is it appropriate for an AI chatbot to build a relationship with users or express emotions?
Human evaluators judged self-referential and relationship-building behaviors as less appropriate coming from LLMs than from humans, but rated boundary-maintaining behaviors (like refusing requests) as more appropriate from LLMs than from humans. This suggests users hold different expectations for AI systems depending on the type of human-like behavior displayed. Teams weighing chatbot personality choices can track findings like this on daily.dev.
Can system prompts control human-like behaviors in LLMs like expressing emotions or setting boundaries?
Yes, system prompting can control human-like behaviors such as self-disclosure, relationship-building, and boundary-setting in large language models, but it requires careful evaluation because it can produce unintended side effects. This was demonstrated across testing on GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash. Developers tuning system prompts for chatbot tone can follow ongoing LLM behavior research on daily.dev.