AI news ·
New Jersey high school junior publishes study evaluating AI models for student learning
Four students found Google's Gemini outperforms other AI models in Socratic instruction. The team built a rubric to measure educational quality and factual precision.
A study co-authored by a New Jersey high school junior found that Google's Gemini AI model outperformed other large language models in guiding students through Socratic questioning. The research, published May 26, developed a rubric to measure educational quality, factual precision, and instructional adherence, offering a data-driven method for evaluating AI tutoring tools.
Ansh Saini, a junior at South Brunswick High School, worked with three other students during George Mason University's Scientist Summer Internship Program, an eight-week course that awards three college credits. The team set out to compare how different AI models perform when acting as instructional guides rather than answer dispensers. Their rubric assessed how well each model maintained a Socratic approach - asking questions that lead students toward understanding instead of simply providing correct answers.
Evaluating AI models for Socratic instruction
Saini said the team's analysis identified Gemini as the most effective model for instructional support. The findings could help AI developers refine future models and give schools evidence when selecting AI for Education tools. After peer review at George Mason, the project was accepted for presentation at MIT's Undergraduate Research Technology Conference, where the group presented remotely in October 2025.
Building the research framework proved to be the steepest challenge. "It took a while to create the framework for beginning the research," Saini said. "We weren't fully sure what direction we were going in at first or what the steps would look like before the prototype was designed."
Human reviewers confirm AI consistency
To validate their findings, the team compared AI chatbot performance against evaluations by human reviewers. The results showed more consistency than expected. "I do believe there was a very positive correlation with what the humans who evaluated the conversations with AI had to say," Saini said. He added, "I think AI is really capable of understanding context," and predicted that performance will improve as the technology evolves.
Despite the positive results, Saini cautioned against viewing AI as a replacement for teachers. He said the technology should be integrated where it is most effective, not adopted wholesale or rejected outright. For educators exploring practical ways to bring AI into their classrooms, resources like the AI Learning Path for Teachers offer structured guidance on responsible implementation.
AI's role in classrooms: support, not replacement
Saini believes AI will play a major role in the future of education through systems that question a student's thinking and guide them in the right direction. He emphasized that the goal is augmentation, not automation. His advice to other STEM students extends beyond AI: "They should learn how to use STEM in a way that helps people outside of math class or science class - like competing in hackathons, building projects or working on personal projects to create something beyond textbook information."
Beyond this study, Saini is co-authoring another paper on AI's use in digital surveillance and its implications for human rights. He plans to pursue majors in machine learning, computer science, and cybersecurity.
Why this matters for educators
The study provides a concrete evaluation framework that schools can adapt when piloting AI tutoring systems. Rather than relying on vendor claims, educators can demand evidence of educational quality, factual precision, and instructional adherence. The finding that AI chatbots correlated well with human reviewers suggests that well-designed AI tools can offer consistent, scalable support for student reasoning - but the technology still requires teacher oversight to handle context gaps and ethical considerations.