Mockingbird in Humanity: Data Fondness of LLM in Hosting Virtual Personalities

The intelligent development in building design, construction, and operation & maintenance is exceptionally rapid, which has become a trend that cannot be ignored in the current field of architecture. With the help of prompt engineering, architects can use generative AI to lay out building space designs and even generate 3D drawings. Artificial intelligence agents can act as designers and owners, representing all parties involved in the building life cycle. In this way, they simulate all parties involved in the building life cycle, providing a comprehensive perspective and solutions for the smooth progress of the building. However, this has led to a problem worthy of in-depth exploration: large models have tendencies when playing different roles. In this article, we aim to deeply explore the tendencies of large language models (LLM) when playing virtual personalities. Specifically, we will conduct extensive experiments to examine two important aspects. One aspect is the analytical ability of large models in terms of virtual personalities, which includes how they interpret requirements in different situations and how they conduct logical analysis according to different role positions. The other aspect is the performance of large models in terms of regions and ethnic groups when playing virtual personalities. Different regions have different architectural cultural and style requirements, and different ethnic groups also have unique architectural aesthetics and traditions. Although LLMs have shown a certain discriminative ability during the role-playing process and can distinguish different role requirements, we find that the content they generate still shows a specific content tendency. This research can deepen the understanding of LLM's performance in multiple aspects such as building design and operation & maintenance.