Experimental Evidence on the Learning Impact of Generative AI
Zara Contractor, Germán Reyes
Let students use AI during learning—it works. The gains stick if they use it to understand, not to generate. Train for augmentation (explaining concepts) not automation (text generation). Optimize for time reallocation: less drafting, more reading.
Generative AI might boost essay quality during use but undermine actual learning—the classic performance-competence gap. Does letting students use ChatGPT while learning teach them anything, or just make them dependent?
Method: Proctored experiment with undergraduates writing analytical essays. AI access raised immediate test scores by 0.27 standard deviations, and gains persisted one week later. Essay quality showed little change during AI use but improved in style and relevance one week later when writing unaided—but only for augmentation users who used AI to explain concepts. Automation users' quality gains vanished once AI was removed. Students shifted time from drafting toward reading and searching.
Caveats: Single topic, proctored setting. Real-world use without supervision may differ. Augmentation vs automation distinction relies on self-report.
Reflections: Does the augmentation advantage hold across disciplines with different knowledge structures (STEM vs humanities)? · What interface design patterns encourage augmentation over automation use? · Do the learning gains compound over a semester, or do students eventually regress to automation?