Similar aggregate but divergent disaggregate responses in human and AI under sequential reward-penalty incentives

With the rapid advancement of general-purpose AI, large language models (LLMs) are increasingly being used to simulate human decisions in structured environments. However, whether similarity in aggregate human and LLM responses extends to behavior under sequential reward-penalty incentives remains unclear. We address this question by introducing an established carbon mechanism to a repeated e-commerce delivery task and comparing decision responses between human, DeepSeek-v.3.2, and Qwen3-Max participants, wherein standard and slow delivery options are provided. In baseline settings, all participants predominantly select the theoretically dominant option. Under sequential reward-penalty incentives, all participant types shift toward slow delivery and exhibit comparably large aggregate changes. However, this aggregate-level similarity does not extend to "value of time" (VOT)-stratified response patterns and longitudinal trajectories. Humans exhibit gradual and incomplete adjustment, whereas LLMs display model-specific responses, including more stable or variable adjustment patterns depending on the model configuration. Thus, similar aggregate responses can coexist with divergent disaggregate behavioral patterns.