Exploring the application of AI chatbot tools in higher education: Evidence from the Duke University student survey

With the rapid advancement of AI technologies, generative tools such as ChatGPT and DeepSeek are being increasingly integrated into higher education to support student learning, cognitive development, and domain-specific training. This study explores the impact of AI chatbot usage on students’ academic outcomes on the basis of survey data from Duke University. The key variables include AI usage frequency, study duration, and academic performance. The results revealed a positive association between AI use frequency and perceived academic improvement but a nonlinear, inverted U-shaped relationship with actual academic performance, suggesting that moderate AI use enhances efficiency, whereas excessive reliance may hinder independent learning. Notably, students who studied 4–6 hours daily with moderate AI use achieved the best outcomes, whereas long study durations without AI were linked to lower performance. Satisfaction also varied by task type—AI was more effective for structured tasks (e.g., coding) than for creative or analytical tasks. Students’ primary concerns—information accuracy (67.7%), overreliance (22.6%), and ethical/privacy risks (6.5%)—reflect key challenges in digital safety and academic integrity. Random forest analysis revealed that AI usage frequency was the strongest predictor of learning outcomes. These findings contribute to ongoing discussions on the pedagogical use of AI in safety science education, highlight the educational potential of AI chatbots and emphasize the need for responsible, evidence-based integration in higher education.