Across the Years, Closer to Us: Temporal Changes in Moral Caring by Large Language Models

This study aims to explore how moral caring behaviors in large language models (LLMs) change over time and under different emotional priming scenarios, such as the positive friend prompt and the negative snake prompt. We assessed three majors which are LLMs ChatGLM4, ChatGPT-40, and Kimi using a standardized moral dilemma task, where the suggested donation amount served as the behavioral measure. The results showed that emotional priming had no significant effect on moral decisions. However, both the year and model type had significant impacts: models tested in 2025 generally suggested higher donation amounts than those in 2024, although the degree of change varied among the models. In an additional experiment, we examined six versions of Chatbox AI (2.0, 2.5, 3.0, 3.5, 4.0, and 4.1) to trace their moral decision patterns over a longer span of development. This allowed us to observe how behavioral tendencies shifted more granularly over time. The results showed that moral development wasn't simply about becoming “more generous”: while Kimi and ChatGPT grew more caring, ChatGLM4 and later Chatbox versions moved toward more rational, human like decisions. These findings suggest that temporary emotional cues play a minor role, while long term changes in structure, alignment strategies, and training goals shape how models make moral decisions. Importantly, moral adjustment does not follow a single direction model that were initially more self-centered became more caring, while those that were overly generous became more rational. This dynamic calibration supports the theory in machine psychology that moral minds in AI emerge through structural evolution and alignment refinement, and it offers real world insights for designing ethically aligned intelligent systems.