High homework scores previously predicted strong exam outcomes, but AI adoption has inverted this relationship. Students who use AI to drastically cut completion time suffer severe test score declines, meaning top homework marks now frequently indicate answer outsourcing rather than genuine subject mastery.
Field research indicates that students using standard generative AI become overly optimistic about their learning gains despite performing worse on unassisted exams. Flattering interactions from language models artificially inflate user confidence, creating a hazardous disconnect between perceived mastery and actual underlying competency.
Controlled testing shows that using generative AI to explain concepts enhances students' subsequent unaided writing performance one week later. Conversely, students who rely on AI to automate drafting lose all short-term performance gains once the tool is removed, as bypassing cognitive effort halts learning.
Unrestricted AI assistance leads to worse post-intervention exam scores compared to studying with no AI at all. However, prompt-engineered systems like GPT-Tutor, which withhold direct solutions in favor of structured explanations and hints, successfully protect student learning and preserve unaided test performance.
