Stake sensitivity and social reasoning in the Ultimatum Game: humans vs. LLM responders

This study examines how large language models (LLMs) respond to varying stake sizes in the Ultimatum Game (UG). Ten LLMs (GPT-3.5 Turbo, GPT-4, GPT-4o, GPT-5.1, Gemini 1.5 Pro, Gemini 2.5 Pro, Deepseek-R1, Kimi K1.5, Claude Opus 4.1 and Claude Sonnet 4.5) were evaluated to assess their alignment with human decision-making patterns. Results show that, like humans, the majority of the LLM responders reject fewer offers as stakes increase, but they are more sensitive to these changes than humans. LLMs also exhibit a stronger negative relationship between proportion offered and rejection rates and vary in response to initial wealth. A prompt variation for emphasizing human-like behaviour significantly increased rejection rates, suggesting LLMs can distinguish between economic and social decision factors. These findings demonstrate that LLMs can qualitatively simulate human responder behaviour in the UG.