Using AI, Doubting AI: Developer Mental Models and Trust in AI-Assisted Software Work

Generative AI has entered software work with remarkable speed, yet trust in its output remains unsettled. This study examines how developers and learners navigate AI-assisted software work through mental models and trust. Analyzing 40,052 responses from the 2025 Stack Overflow Developer Survey, the findings reveal a durable paradox: professionals use AI more intensively than learners but trust it less, evaluating its capability for complex tasks more cautiously. The primary threat to trust is corrective labor, such as debugging time, rather than code legibility. Cluster analysis identifies four mental-model profiles—open but light users, skeptical dependents, resistant/disengaged respondents, and enthusiastic adopters—with professionals disproportionately concentrated in skeptical dependence. Ultimately, adoption does not settle the trust question; instead, mental models act as interpretive filters through which developers decide when to rely, verify, and pull back.