As schools and educator preparation programs increasingly turn to artificial intelligence (AI) for support, one question looms large: Can AI realistically simulate classroom interactions—and help educators get better at their craft? Our recent study explores whether AI-powered teaching simulations can give elementary educators meaningful opportunities to practice instructional skills—especially in core subjects like mathematics. Study findings show strong potential but also highlight areas where additional refinement can enhance the impact on educator learning.  

Study Focus: Exploring the Potential of GenAI Teaching Simulations 

Our team tested a digital platform where educators interact with a simulated elementary student powered by generative AI (GenAI). The goal was to give educators a safe space to develop their ability to engage in one high-leverage teaching practice: engaging students with subject matter. Two key instructional strategies for doing so involve educators in eliciting and responding to student thinking and making content explicit in instruction.  

In this study, we used a GenAI teaching simulation where the teacher interacted with a third grade GenAI student named Neil to help him learn how to use area models to compare fractions with the same numerators or same denominators. As shown in Figure 1, Neil started the conversation with a typical misunderstanding—thinking that larger denominators mean larger fractions when comparing fractions with the same numerators. During their conversation with Neil, each educator’s goal was to use productive teaching moves to uncover what Neil already knew—and where he was confused—and to use area models to help Neil learn how to compare fractions correctly. This design mirrors key aspects of real classroom teaching, where educators must continuously assess and respond to student thinking. After engaging in the simulation, each educator received a feedback report about the nature of the teaching moves they used and completed a survey about their simulation experience and perceptions.  

Twenty-eight educators participated in this study including elementary paraeducators, in-service teachers, and administrators. Our research team analyzed chat transcripts from the GenAI teaching simulation (one per participant) using an evaluation rubric (Mikeska et al., 2026), calculated descriptive frequencies for Likert scale survey items, and identified themes in open-ended survey responses to determine: (a) how well the GenAI student responses aligned with their pre-designed learner profile and (b) how educators experienced the simulation in terms of its authenticity and usability. Our team analyzed the GenAI student responses across all transcripts for whether each one was responsive to the educator's questions and prompts, included adequate information, was age/grade appropriate, had an adequate emotional response, was coherent, and aligned with the GenAI student’s pre-designed understandings and misunderstandings. For perceived authenticity and usability, participants rated how realistic the interactions with the GenAI student felt, how easy the platform was to use, and whether the experience was engaging. This combination of performance data and user feedback provided a fuller picture of how well the simulation worked. 

Key Findings: What Worked Well and Improvement Areas for Future Work 

Findings showed strong response quality from Neil, the third grade GenAI student, in these conversations. Most of Neil’s responses were relevant and appropriate (94.3%), showed coherence and consistency across the conversation (95.8%), were age appropriate (99.6%), and were aligned with Neil’s intended knowledge profile (93.3%). These results suggest that the GenAI student was generally believable and stable in its behavior. In addition, most participants found the platform easy to use, agreeing that the platform was well-designed and the instructions and preparation materials were clear. In terms of engaging elements, some participants reported that the GenAI student’s voice and responsiveness made the interaction feel “very close to a real conversation.” These findings highlight the simulation’s potential to create immersive practice experiences. 

Despite strong overall performance, several limitations emerged. Sometimes the GenAI student responses were inconsistent, especially in terms of information provided. About 20% of Neil’s responses included either too much or too little information relative to the educator’s prompt or question. This can disrupt the flow of instruction and make it harder for educators to assess student understanding. In addition, participants identified a few barriers that reduced engagement with the most common one being response delays from the GenAI student. Participants also noted that the online interaction did not have any manipulatives for use, which limited the student’s opportunity to engage in hands-on learning. Such challenges can affect how realistic the experience feels.  

Looking Ahead: Enhancing the Impact of GenAI Teaching Simulations 

Ensuring that educators have access to scalable tools to practice and develop their instructional interactions in safe, low risk environments is a continuing challenge in teacher education. This study suggests that GenAI simulations can provide flexible, on-demand practice spaces where educators have opportunities to try out various instructional strategies without the need for real students. The findings highlight the importance of the accuracy of the GenAI student responses and the usability and authenticity of the simulation experience. In future development, it would be important to reduce some of the technical friction participants experienced by exploring approaches to decrease the GenAI student response delays and to expand GenAI student diversity by including varied knowledge profiles and emotional expressions. These improvements are useful in making simulations more reflective of real classrooms and more effective for teacher preparation and professional development. With continued refinement, these tools can become a powerful complement to current professional development—helping educators better understand student thinking and improve their practice before stepping into the classroom. 

References 

Mikeska, J.N., Beigman Klebanov, B., Halder, S., Jorgenson, H., Behl, K., & Bhatia, A. (2026, April 8-11). Evaluation criteria for measuring the accuracy of generative AI chatbots for teacher learning. [Paper presentation]. National Council for Measurement in Education Annual Meeting, Los Angeles, CA. 

Disclaimer 

This post was drafted with the support of Microsoft Copilot and was based on original research to be presented at the ISTE 2026 conference. AI assistance was used to help summarize key points and outline the post’s structure and main ideas. The final content, interpretations, and accuracy were reviewed and verified by the first author.