生成式AI模型人智共创中的对齐问题研究

目的 以人工智能对齐的视角,探讨在人智共创中生成式AI模型如何对齐设计师的意图。方法 通过要素分析,以人工智能对齐问题中的可解释性与可控性为研究视角,探讨生成式AI技术作为辅助设计工具如何对齐设计求解过程中“探索-创新-评估”三个设计阶段的具体意图与需求,并分析对齐阶段中需要解决的对齐问题。结果 根据对齐阶段的任务构建基于人智设计概念表征交互式对齐方法、表征拓展方法与表征评估方法。结论 在三个对齐阶段中分别构建意图对齐、设计空间拓展和设计规则匹配这三种方法,帮助设计师构建可控、可解释的人智共创方法,从而构建可控、可信的人智共创。

Alignment Issues in Human-AI Co-creation Using Generative AI Models

The work aims to explore how generative AI models align with designers'' intentions in human-AI co-creation from the perspective of artificial intelligence alignment. Methodologically, the interpretability and controllability issues in AI alignment were explored through a feature analysis approach. The research examined how generative AI technologies, as auxiliary design tools, align with the specific intents and needs of the three stages of the design process:"exploration, innovation, and evaluation". The alignment challenges that needed to be addressed in each stage were analyzed. Technologically, the study proposed an interactive alignment method, representation expansion method and representation evaluation method based on human intelligence design concept representation. In conclusion, the study constructs three alignment methods:intent alignment, design space expansion, and design rule matching, in the three stages. These methods aim to assist designers in building controllable and interpretable human-AI co-creation methods to contribute to controllable and trustworthy human-AI co-creation.