Generative AI Chatbots for self-regulated learning while balancing cognitive load: perceptions, interaction patterns, and instructional designs in K-12 learning

Learners today have access to a wide range of tools that facilitate self-regulated learning (SRL), a complex and often effortful process. SRL can lead to high levels of cognitive load, but it can also help to balance cognitive load. Generative Artificial Intelligence (genAI) chatbots have become a recent, versatile addition to the learners’ media environments. Students already make use of them regularly for information tasks like research. There is little research on how students perceive and interact with genAI chatbots during research. Without specific instruction, they may not perceive all the affordances for learning, like adaptation prompts (e.g., “Make it easier”, “Give me a summary”) which can help learners balance cognitive load. Learners may not make use of the affordance of conversation by which they can monitor and deepen their comprehension. Rather than being implemented with their default configuration and interface, chatbots can be designed to support students in making use of these affordances. Three studies were conducted iteratively to provide insights into K-12 students’ perceptions of chatbots, their interaction with them, and supportive chatbot designs. The Pilot Study investigated how eight students used a chatbot for research in an authentic classroom setting. It found a wide variety of interaction styles but also recurring patterns, for example, that the use of adaptation prompts was rare. In Study 1, adaptation guidance was therefore explored as a value-added design. 106 students in 9th and 10th grade performed 20-minute research with a chatbot (ChatGPT-3.5). The experimental group (EG) received tips on adaption prompts and their chatbot interface came with prompt suggestions and language adaptation options. The EG indeed used twice as many adaptation prompts whereas one third of the control group (CG) did not use any adaptation prompt. The EG showed more SRL behavior in terms of more adaptations and (follow-up) questions and less off-topic chat. Both groups perceived the chatbot as easy to use and judged their self-efficacy as high despite moderate prior experience. As this could imply that they invested too little effort and offloaded too much of the learning process onto the chatbot, in Study 2, the chatbot for the EG was prompted to ask questions to encourage the students to monitor their comprehension. 71 students conducted 15-minute research with a chatbot (ChatGPT-4o). The results show that the EG indeed answered questions in many cases – even though this was optional. As hypothesized, the CG used significantly more adaptation prompts. This is interpreted as more coregulation in the EG and more SRL in the CG. Overall, students perceived the chatbots as easy to use, and interaction patterns emerged: they did not transfer search engine strategies, rarely engaged in conversation, underutilized adaptability, and mostly used the chatbot as a content generator. Lightweight interventions significantly increased self- and coregulated learning but to significantly impact cognitive aspects like load and effort or learning outcomes, alternative chatbot designs should be explored.