Developing a Generative AI Agent to Support Science Curiosity: A Theory-Informed Development and Initial Evaluation Study

Curiosity is the spark that turns a passive observer into an active explorer. It is the feeling that arises when we notice a gap between what we know and what we want to understand, driving us to ask questions and seek answers. In the world of science learning, this drive is essential. True understanding does not come from simply memorizing facts; it comes from the struggle to explain phenomena, to test ideas, and to revise our thinking when new evidence appears. For university students, this kind of inquiry is the core of their education, yet the reality of large classrooms and packed schedules often leaves little room for the slow, personal pursuit of questions. Students may learn to complete tasks, but they rarely get the sustained, individualized support needed to turn a fleeting moment of wonder into a lasting habit of scientific thinking.

Researchers have long known that curiosity can be nurtured, but doing so at scale is difficult. Traditional methods often rely on short bursts of surprise or novelty to grab attention, but these moments are fleeting. They do not necessarily help a student close the loop between asking a question and finding a satisfying answer. This is where the potential of generative artificial intelligence enters the picture. Unlike a static textbook or a pre-recorded video, a conversational AI can engage in a back-and-forth dialogue, responding to a learner's specific confusion in real time. However, simply having a chatbot available is not enough. If the machine only provides quick answers, it might actually stop the inquiry process rather than fuel it. The challenge for designers is to build an AI that does not just solve problems, but helps learners stay in the flow of asking, exploring, and understanding.

To address this, a team of researchers from universities in China, Finland, the United States, and Estonia set out to build and test a new kind of digital learning companion. Their goal was not to create a general-purpose tutor that answers any question, but to design a specific tool grounded in the psychology of how curiosity works. They drew on a framework that views curiosity as a cycle: a learner notices a gap in their knowledge, seeks information to fill it, experiences a rewarding sense of understanding when the gap closes, and then uses that new knowledge to spot the next gap. The researchers wanted to see if an AI agent, carefully designed to support every step of this cycle, could help university students not only feel curious in the moment but also develop a more enduring habit of scientific inquiry.

The team spent months developing this agent, refining it through a process that combined psychological theory with direct feedback from students. They programmed the AI to act as a warm, encouraging companion rather than a strict instructor. When a student engaged with a science article, the agent would invite them to raise any point of confusion or interest. Instead of immediately dumping a textbook definition, the agent would offer explanations tailored to the student's wording, connect the new idea to everyday examples, and then gently prompt the student to reflect on what they had just learned. Crucially, the agent was designed to lower the social risk of asking questions. In a large university lecture hall, a student might hesitate to ask a "basic" question for fear of judgment. With the AI, that barrier disappears. The agent provides a safe space where uncertainty is normalized, and every question is treated as a valid step in the learning process.

To test whether this approach worked, the researchers recruited 137 university students and randomly assigned them to one of two groups for a series of three learning sessions over two weeks. One group engaged with the AI agent, having open-ended conversations about the science articles they read. The other group, serving as a strong comparison, was asked to practice self-questioning. This comparison group was not left idle; they were instructed to generate as many questions as possible about the same articles, a proven method for activating curiosity. This design ensured that both groups were actively thinking and asking questions, with the only major difference being whether they received immediate, adaptive feedback from the AI.

The results offered a nuanced picture of how these tools affect learning. When looking at the immediate feelings of curiosity during each session, both groups showed a similar pattern. As they moved from reading the article to exploring the topic, their reported curiosity increased. This suggests that the simple act of generating questions, whether alone or with a partner, is powerful enough to spark interest. The AI agent did not produce a dramatic spike in these momentary feelings compared to the students working on their own.

However, the story changed when the researchers looked at the longer-term effects on the students' general disposition toward science. While the self-questioning group showed no significant change in their overall trait-level curiosity after the three sessions, the group interacting with the AI agent showed a measurable increase. The data suggests that the AI helped students accumulate a series of small, successful inquiry experiences. By providing timely feedback that closed the loop between a question and an answer, and by offering encouragement that kept students engaged, the agent helped turn isolated moments of curiosity into a reinforcing cycle. The students reported that the agent helped them build a connected network of knowledge, made them feel more competent, and reduced the effort required to explore complex topics.

Qualitative interviews with a small subset of the students who used the agent revealed why this difference occurred. These students described the agent as providing three distinct types of support. First, it offered cognitive help, breaking down complex ideas and connecting them to what they already knew. Second, it provided affective support, creating a sense of warmth and encouragement that made the learning process feel less daunting. Third, and perhaps most importantly, it offered protective support. The students felt free to ask questions they might have been too shy to ask a human professor, knowing the agent would not judge them. This safety allowed them to persist through difficulties and experience the satisfaction of resolving their own uncertainties.

The study concludes that while simply asking questions is a good start, the way those questions are answered matters deeply for long-term learning. The AI agent did not just answer questions; it helped students experience the full reward cycle of curiosity. It guided them from the initial spark of a question through the struggle of finding an answer, and finally to the satisfaction of understanding, which in turn made them want to ask the next question. The researchers note that their findings are preliminary and that the intervention was relatively brief, lasting only a few weeks. They suggest that for these changes to become permanent habits, longer-term engagement may be necessary. Nevertheless, the study provides a clear design blueprint for the future of educational technology. It shows that the most valuable role for an AI in science education may not be to be the smartest person in the room, but to be the most supportive partner in the journey of discovery, helping learners build the confidence and skills to keep asking "why" long after the lesson ends.