Generative AI Can Harm Teaching

Teachers are rapidly adopting generative AI, yet there is little causal evidence on whether AI-assisted teaching improves or harms student outcomes. We conducted a randomized field experiment across a chain of middle and high schools in Turkey, assigning teachers to receive a generative AI teaching support tool or business-as-usual control. The majority of teachers used AI for lecture material, syllabus, and exercise creation. Providing teachers with AI reduced student intrinsic motivation by 0.11 standard deviations, with students rating AI-assisted courses as less enjoyable, less interesting, and less important. The decline was more pronounced among teachers who reported heavier AI use at baseline. AI did not affect average academic performance, but among students of lower-performing teachers, it significantly reduced both academic performance and students' confidence. On average, teachers did not change their beliefs about whether AI helps or harms student learning, but teachers who had used AI more prior to the experiment became more pessimistic, whereas those who previously had used it less became more optimistic. One possibility is that the initial awe of discovering the power of AI to instantaneously respond to any request later gives way to an awareness of its unintended negative effects. These findings challenge the assumption that productivity-enhancing AI tools automatically benefit instruction and highlight a principal-agent problem in AI-assisted education.