AI-Driven Visual Preference Biases: Exploring Future Challenges in Urban Planning and Building Design

This contribution explores the topic of window view preference using text-to-image generative AI tools and ChatGPT's large language model (LLM). It begins by framing the concept of visual preference within the fields of envi-ronmental psychology and urban studies. The study then employs generative AI tools to create a dataset of 9000 pictures and 3000 textual prompts. The datasets, which focus on the duality of pleasant and unpleasant visual condi-tions in a residential context, have been subsequently analysed to identify recurrent patterns and trends. The analysis results revealed a stark preference within the queried AI tools to associate pleasant visual conditions with natu-ral motifs, while associating unpleasant visual conditions with urban motifs. These observations may seem to align with the principles of biophilic design. The overall study was conducted to probe the potential biases of commonly used AI tools in generating residential scenarios. It is possible that studies on visual preference biases within AI models will become a relevant subject soon. Any shortcomings of AI platforms in handling such topics could sig-nificantly impact physical construction, especially as these tools are going to be used more frequently within real world applications.