Frontiers | An experimental study of structured generative AI integration to mitigate pedagogical, cognitive, and ethical barriers in programming education
Frontiers | An experimental study of structured generative AI integration to mitigate pedagogical, cognitive, and ethical barriers in programming education
At Covenant University in Nigeria, 62 students per arm studied introductory Java over seven weeks. Everyone had a one-off AI-literacy session; the structured arm additionally practiced task-aligned prompts, checking and revising outputs, peer exchange, and gradually less instructor help. A 28-criterion rubric evaluated baseline and endline problem-solving, critical thinking, creativity and programming logic; blinded raters scored anonymized work. Adjusted post-test differences favored the guided arm by 0.29 on the 0–4 higher-order-thinking scale (p<.001) and 0.21 for programming logic (p=.047). Chat logs recorded more of the behaviors the guided arm was taught.
The comparison randomizes an instructional bundle, not prompt technique alone. Its sample is students learning Java, not engineers using coding agents on maintained software. Although the final course capstone allowed ChatGPT, the authors do not separately report comparative capstone correctness, and they do not clearly state whether the rubric post-test allowed AI. Instructor hours, longer-term independent diagnosis and assisted delivery value remain unmeasured. A shorter TUM tutor trial found better AI-assisted exercise scores without larger measured knowledge gains, so dosage and endpoint matter.