Defensive adoption and the intermediate coding threshold in generative AI usage among medical postgraduates: a cross-sectional mixed-methods study

The integration of generative artificial intelligence (GenAI) in medical research presents both methodological opportunities and pedagogical challenges. While traditional technology acceptance models characterize anxiety as a barrier to adoption, high-pressure academic environments may instead foster a “defensive adoption” mechanism. This study investigates the structural relationships among AI anxiety, agentic literacy, and usage intentions, and identifies distinct user profiles among medical postgraduates. This cross-sectional mixed-methods study analyzed 414 valid responses from medical postgraduates in China. The self-developed Agentic Literacy scale underwent expert content validation, cognitive interviews, and pilot testing prior to the main survey. The analytical framework comprised two stages. First, structural equation modeling (SEM) tested the hypothesized causal pathways, while controlling for prior programming experience. Second, latent profile analysis (LPA) identified heterogeneous subgroups based on user patterns. Quantitative findings were triangulated with a qualitative thematic analysis of open-ended responses to capture the cognitive mechanisms underlying different adoption profiles. Participants primarily engaged with GenAI for basic text processing, with 17.4% utilizing tools for deep methodological tasks. SEM analysis indicated a pattern consistent with defensive adoption, where higher AI anxiety was associated with greater intention to use GenAI ( \(\:\beta\:\) = 0.254, P < 0.001). Furthermore, deeper use was associated with higher agentic literacy, which emerged as the strongest observed correlate for crossing the intermediate coding threshold. LPA identified three distinct profiles: Agentic Innovators (high literacy, high anxiety), Moderate Adopters, and Detached Users (low literacy, low anxiety). Qualitative analysis revealed that “Detached Users” focused on safety and hallucination auditing, whereas “Agentic Innovators” leveraged AI for complex code orchestration, confirming that anxiety was associated with adoption only when paired with competency. In this cross-sectional sample, higher academic AI anxiety was associated with stronger intention to use GenAI, while deeper use was more closely associated with agentic literacy. Surpassing the intermediate coding threshold may require specific training in AI orchestration rather than mere technological access.