Simulated Art World: Artist Simulation Based on Generative Models
“模拟艺术界”：基于生成式人工智能的艺术家智能体仿真模拟
Process of Artist Simulation Based on Generative AI
Wang, Yangyu. Under Review. Journal of Intelligent Society (Chinese).
王杨聿. 审稿中. 《智能社会研究》.
Abstract: Building on symbolic interactionist theory, sociologists have long argued whether interaction shapes artistic creation, yet empirical tests of the interaction–creation link have been hampered by data constraints. We bridge the gap by using generative models to simulate both artists’ interactions and their creative outputs. Agents based on Large Language Model (LLM) are paired with a conditional Generative Adversarial Net (GAN) to construct “semantic portraits” of artists from their Wikipedia entries and to simulate style debates and relationship formation in a virtual setting. Text generated during the simulation is analysed by BERTopic to trace the emergence of shared meanings; the resulting style‑ and relationship‑based variables then condition an image‑generation model that emulates how paintings evolve after interaction. Results show that interaction itself produces shared meanings among artists and gradually aligns their artistic styles, while the pathways of construction vary across simulation scenarios. The article not only addresses the theoretical question of how interaction and art are linked within the sociology of art, but also advances a methodological framework for analysing social processes for the elite classes.
中文摘要：沿袭着符号互动论的理论观点，艺术社会学家往往认为互动建构了艺术创作。然而，由于数据的限制，过往研究无法直接检验互动过程与创作过程的联系。面对这一研究困境，作者基于生成式人工智能，模拟了艺术家智能体的互动与创作过程。通过将大语言模型与条件生成对抗网络相结合，作者采用艺术家的维基百科词条构建艺术家语义画像，并驱动其在模拟情境中进行风格辩论与关系构建。随后，通过BERTopic主题模型分析仿真模拟过程中的文本数据，识别共享意义的形成路径；并进一步利用风格与关系变量作为条件，训练图像生成模型模拟互动后的绘画变化。结果显示，互动过程本身会建构艺术家之间的共享意义，同时也会使得艺术家的艺术风格逐渐趋同；同时，在不同的智能体模拟情境下，互动的建构过程有所差异。本文的结论不仅回应了艺术社会学中“互动-艺术”关联的理论问题，同时也提出了一种针对精英阶层社会过程分析的方法论框架。