Learning with Human Peers and AI Chatbots: Evidence from Two Experimental Studies

Le résumé fourni par la source

Discussion with peers is a well-established strategy for promoting learning. Because AI chatbots like ChatGPT can respond continuously to students’ questions and explanations, they may serve some of the functions traditionally provided by human discussion partners in learning contexts. Across two pre-registered randomized experiments, we compared learning with human peers and AI chatbots and experimentally manipulated how students interacted with ChatGPT. In Study 1, students learned an unfamiliar physics concept and then discussed the material either with human peers (n = 47) or with ChatGPT (n = 44). Learning outcomes were generally similar across conditions even though many students in the ChatGPT condition just used the chatbot as an answer-generating machine. Study 2 examined whether guidance designed to promote more active engagement with ChatGPT would influence learning with a new set of materials. Participants were assigned to a human-peer condition (n = 43), an AI condition without instruction (n = 43), or an AI condition with instruction (n = 44). The instructional intervention led to more sophisticated prompting and discussion strategies but did not improve learning. Taken together, the findings suggest that learning with ChatGPT can produce outcomes comparable to learning through peer discussion and that even a brief intervention can increase students’ use of recommended interaction strategies. The results also highlight an important distinction between changing how students use AI and changing what they learn from it.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.