As Generative Artificial Intelligence (GenAI) rapidly permeates higher education, its application in safety-critical STEM fields—such as civil engineering—has raised significant concerns regarding "false mastery." This phenomenon occurs when students produce plausible-sounding answers via AI that lack mechanism-based reasoning or verifiable evidence. To address this, this study proposes and evaluates a technology-enhanced scaffolding workflow designed to operationalize verification-oriented risk reasoning within a civil engineering construction safety module (N = 50).

The study structures student–GenAI interaction into five staged prompts and artifacts to ensure logical depth: (1) Scenario Interpretation and Boundary Setting: Defining problem parameters to prevent hallucinations or irrelevant AI generation. (2) Mechanism-Based Hazard Identification: Requiring students to analyze the underlying physical or causal mechanisms of potential accidents. (3) Likelihood-Severity Justification: Providing logical or quantitative defenses for assigned risk levels. (4) Control Selection: Aligning mitigation strategies strictly with the Hierarchy of Controls. (5) Verification-Oriented Reflection: Mandating uncertainty marking, evidence/source cross-checking, and detailed revision logs.

Assessment via rubric-based scoring and behavioral coding revealed significant improvements in reasoning transparency and a stronger coupling between hazard mechanisms and control logic. Furthermore, students demonstrated increased explicit verification practices, such as active uncertainty labeling and revision tracing. This study contributes a transferable GenAI-enabled learning design and assessment package—including prompt templates, expected artifacts, and scoring signals—to cultivate measurable verification literacy in safety-critical engineering education.