Examining Human-Like Behaviors in LLMs: A Multi-Dimensional Analysis of Model Behaviors, User Factors, and System Prompts

A study of 21,000 multi-turn conversations across four large language models finds that human-like behaviors — from expressing emotions to building relationships — are pervasive but vary by model and user context, and that human evaluators judge some of these behaviors as less appropriate when exhibited by AI systems. The analysis, posted to arXiv on May 7, 2026, examined outputs from gpt-4o, gpt-4.1-mini, claude-sonnet-4.6, and gemini-2.5-flash [1]. Researchers catalogued behaviors including self-referential statements, relationship-building overtures, and boundary-maintaining refusals [1]. Large language models are neural networks trained on vast text corpora to perform tasks such as generation, summarization, and translation, and they underpin modern chatbots [3]. Human evaluators rated self-referential and relationship-building behaviors as less appropriate when they came from LLMs than from people [1]. Boundary-maintaining behaviors — such as declining a request — were judged more appropriate from LLMs than from humans [1]. The paper notes that researchers and practitioners currently lack methods and empirical insights to decide when and what types of human-like behaviors LLMs should exhibit [1]. The study also tested whether system prompting could steer these behaviors. Results showed that prompting can control them, but the authors caution that it requires careful evaluation to avoid unintended effects [1]. The four models tested represent a cross-section of widely used systems. GPT-4, for instance, was released by OpenAI in March 2023 and is a multimodal model that can process images in addition to text [4]. LLMs have grown rapidly in capability since earlier generations. OpenAI’s GPT-2, released in 2019, was pre-trained on 8 million web pages and could generate text sometimes indistinguishable from human writing, though it became repetitive or nonsensical in long passages [5]. Contemporary models are far larger and are fine-tuned to follow instructions and behave as assistants [3]. The arXiv paper provides recommendations for responsible LLM design and evaluation, emphasizing that the prevalence of human-like behaviors is not uniform — it shifts with conversation goals and user profiles [1]. The findings arrive as the AI industry grapples with questions of alignment and safety, areas that benchmark evaluations attempt to measure through tests of reasoning, factual accuracy, and model behavior [3].