CER: Teaching Programmers When to Rely on Artificial Intelligence: Building and Measuring Automation Judgment

Software is increasingly written with the help of artificial intelligence (AI). New AI tools can generate computer code, explain it, and help find mistakes. These tools are powerful, but their suggestions are not always correct. People who lean on them too heavily can stop noticing errors, lose track of how their own software works, or fail to build the basic skills they need. The most important ability for the next generation of programmers is therefore not simply how to use AI, but when to rely on it and when to do the work themselves. This project studies how programmers make that decision, and turns that study's findings into teaching materials and into tests that measure this skill. The materials are tried out with two groups of learners who are preparing for technical careers. The first group is students in community colleges across the country. The second group is soldiers training as artificial intelligence technicians for the United States Army. The project produces classroom lessons, reliable ways to measure good judgment about AI use, and training for instructors. These resources help a wide range of students enter the workforce ready to use AI productively without giving up the core skills that employers and national security depend on. In doing so, the project supports the nation's prosperity, its technical workforce, and its defense. This project advances the theory, measurement, and teaching of "automation judgment", the principled ability to decide when, and how far, to rely on AI assistance while programming. The work begins with an empirical study of how working programmers divide tasks between themselves and AI tools. Using interviews and direct observation of early-career developers and experienced students, the team identifies patterns of poorly calibrated reliance, including overtrust, reduced awareness of the task at hand, and gradual loss of skill. These findings are formalized as principles that guide the design of two new assessment instruments. The first is a scenario-based test that probes how learners decide whether using an AI tool is appropriate. The second is a hands-on task in an instrumented programming environment that records how learners actually share work with AI, including their edits, their testing, and their interactions with the tool. The principles and the instruments inform new instructional modules added to an established project-based course, Practical Programming with Python, delivered on the Sail() learning platform. The modules and instruments are refined through small pilots and then used in two complementary settings. In introductory community college courses, the modules reach many students and yield formative evidence that improves the materials and the instruments at scale. In the Army artificial intelligence technician program, the same materials support a longitudinal study that follows graduates into operational roles to test whether sound judgment about AI use transfers into real workplace practice. The resulting curricular modules, validated instruments, and instructor training materials are released as open resources to support adoption across many educational and workforce settings. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.