Artificial intelligence as the cognitive filter of design: the transition from natural language to structured data representation in architectural production

Abstract

This study examines the Sapir–Whorf hypothesis, the concept that language shapes perception, within the intersection of modern architecture and artificial intelligence (AI). It discusses how architectural jargon and digital tools act as cognitive filters that shape design approaches. Based on this framework, the research analyzes how using JSON-based structured data representations in communication with AI systems affects the quality and predictability of AI-generated architectural outputs. The methodology involved parametrically decomposing a reference architectural image using Gemini and ChatGPT, manipulating the resulting JSON data, and conducting cross-system testing. Findings from this limited experimental setup suggest that ChatGPT produces relatively well-organized results in semantic interpretation, efficiently decomposing images into conceptual components and ordered data structures. Conversely, Gemini appears to produce relatively more consistent results in this example when visualizing this structured data while maintaining aesthetic continuity and atmospheric integrity. These results suggest that although AI systems appear to work through shared concepts, they may exhibit different output patterns, which can be discussed analogically within the framework of Digital Whorfianism. The study suggests that translating human intent into representation languages compatible with the system’s processing logic may be beneficial for optimizing human-machine interaction in AI-assisted design, and that this process allows architecture to be approached as an activity of representation and translation.