Spatial agency beyond the prompt: an in-the-wild evaluation of generative AI literacy in architectural education

The integration of Generative AI (GenAI) into architectural education frequently relies on text-to-image models that lack the deterministic spatial agency required for rigorous design thinking process. This paper reports on an interdisciplinary studioArchitectural Design in the Age of AI taught at Politecnico di Torino, merging Architectural Design, Computer Engineering, and Representation to move third-year students ($n=51$) beyond prompting toward multimodal, human-in-the-loop co-creation. We present the studio articulation and a pedagogical protocol based on the ARCS workflow, integrating established open-weight models and off-the-shelf components for 3D massing, geometric conditioning (ControlNet), and stylistic fine-tuning (LoRA). To circumvent the "GPU divide" and ensure equitable access, we deployed a custom server-side virtual machine infrastructure, running open-weight models (Flux.1-dev) via ComfyUI. Students’ AI literacy-related perceptions were examined through an in-the-wild study, while their performance at the end of the course was assessed using an AI Literacy Test. Finally, we report actionable guidelines for educators and faculty, discussing the findings of such deployment of the first edition of this studio. The findings from this study suggest that GenAI literacy may benefit from an integrated, transversal approach within the same course rather than being ad-dressed exclusively in technical courses. Our work provides a foundational blueprint for developing equitable, domain-specific AI literacy in higher architectural education.