Emergent Architectural Drawing Literacy in General-Purpose GPT Systems: A Longitudinal Case Study from Recognition to BIM Execution

Abstract

Architectural drawings require more than optical character recognition or object detection: useful interpretation depends on integrating geometry, dimensions, graphical conventions, spatial relations, functional semantics, and discipline-specific knowledge. This preprint presents a longitudinal observational case study of architectural drawing literacy across several deployed GPT systems. Preserved cases from March to August 2026 document a shift from locally available but semantically inconsistent interpretation toward later cases supporting cross-cue reasoning, executable BIM translation, and architectural dialogue. The June GPT-5.5/Codex BIM case uses a deliberately simple synthetic plan-and-elevation stimulus generated before the experiment from author-specified geometry and is interpreted as a controlled lower-bound executable test. The term emergent is used in an operational behavioral sense, not as a claim of a scaling-induced phase transition. The deposit includes the preprint and Supplements S1-S5 containing the preserved interaction records, BIM workflow log, and author-validated reference annotation. Provenance note: the residential control drawing is an author-produced CAD redraw based on measurements of a real apartment; the underlying project documentation is no longer available, and no claim is made over the underlying architectural design. The June BIM source drawings are synthetic ChatGPT-generated stimuli derived from an explicit geometric description supplied by the author. See RIGHTS_AND_PROVENANCE.md in the source package and the AI Assistance Disclosure in the manuscript.