Examining Computer-Science Students’ Metacognitive Beliefs and Reported Practices in Verifying AI-Generated Code: A Concurrent Mixed-Methods Study

Generative artificial intelligence (GenAI) accelerates programming, but its educational value depends on whether students check generated code before accepting it. This concurrent mixed-methods study examined which beliefs accompany such checking among computer-science students at one Chinese university. Survey responses from 294 students were analyzed with an exploratory structural equation model treating three practices as separate outcomes: checking code line by line before use, practicing without AI, and declining in advance to delegate. Robustness checks included ordinal estimation, acquiescence modeling, split-half replication, and multiverse analysis; eleven students were interviewed. Verification-dispensability belief, endorsing reasons for treating further checking as unnecessary, was the only construct negatively associated with all three practices in every robustness check. Other beliefs were narrower or inert: perceived AI literacy accompanied line-by-line checking alone, generic metacognitive self-regulation none. Delegation breadth, the number of activities handed to GenAI, was unrelated to all three: volume of delegation said little about checking. Interviews exposed a mismatch between self-reported categories and actual workflows: students allocated checking by task purpose, accepted process-level understanding, and delegated verification to other AI models, a practice unmeasured by the survey. The data are cross-sectional and self-reported, so these are associations requiring behavioral and longitudinal confirmation.