From generation to co-creation: agency and collaboration in architectural design via natural language interaction

Large language models (LLMs) introduce Natural Language Interaction (NLI) to architectural design, enabling conversational engagement with AI in early-stage work. This study examines how agency and collaboration emerge through NLI in site analysis and massing tasks involving practising architects. Twenty-four experimental design sessions (n = 48 participants) were analyzed using a coding scheme developed for early-stage architectural dialogue. Sequential analysis shows a predominantly human-led flow: architects initiate, refine, and redirect, while LLMs contribute answers, generative content, elaborations, and occasional clarifications or confirmation. The sessions also reveal short intervals of shared initiative in which human and AI jointly steer direction and content, indicating emerging co-creative reasoning. The findings highlight the differences between request-answer sequences and generation-elaboration-modification cycles, which have broader implications for the development of design creativity through repeated interactions. Although AI-led moves are infrequent, they can prompt reformulation, renewed specification, and short intervals of shared initiative. These results position LLMs not as autonomous teammates, but as dialogic contributors whose interactional role depends on task structure and human steering. The study offers empirical evidence of how agency and collaboration unfold through NLI in early architectural design, and presents a task-specific coding scheme for analyzing human-AI dialogue in site analysis and massing.