EPISTEMIC RISKS OF UNSUPERVISED GENERATIVE AI USE IN AEC EDUCATION: AN EMPIRICAL STUDY

Resumo

The adoption of generative artificial intelligence (AI) in higher education has intensified concerns regarding learning quality, epistemic responsibility, and professional judgment in Architecture, Engineering, and Construction (AEC) education. This study investigates the consequences of unsupervised generative AI use in a design activity, focusing on performance, error propagation, and epistemic engagement. A two-phase design was adopted. Phase 1 applied a diagnostic survey (n = 244) on AI use and disclosure practices. Phase 2 conducted a quasi-experimental activity (n = 24), in which students solved a normative design task using ChatGPT (GPT-Vanilla) without prior instruction. Responses and interaction logs were evaluated through a rubric and an error taxonomy. Results indicate nearly universal and frequently undeclared AI use. In the GPT-Vanilla condition, performance was low (mean = 1.60/10), with errors concentrated in normative verification and multistep consistency. Of the 127 coded errors, 95.28% resulted from unverified acceptance of AI outputs, evidencing systematic error propagation and epistemic passivity. Findings reinforce the need for pedagogical mediation and explicit regulation, proposing “cognitive governance” as a core competency for responsible AI use in AEC education.