The growing popularity of large language models (LLMs) in education has spurred their use in a range of tasks, including creating assessments, analyzing data, and evaluating student responses. However, due to the potential for AI to generate misleading outputs, the European Commission (2024) has proposed guidelines emphasizing responsible usage of generative AI in research, focusing on reliability, honesty, respect, and accountability. These guidelines call for verifying and reproducing information produced by AI, ensuring transparency regarding AI use, recognizing the technology’s limitations, and taking overall responsibility for the outputs. Building on these principles, the present study aimed to demonstrate a process for applying these guidelines in research, centering on evaluating how accurately LLMs reproduce a qualitative data analysis conducted by human researchers and how consistent their outputs are across multiple LLMs.

The study posed three core research questions:

1. How accurately can AI models replicate a qualitative data analysis performed by human researchers?
2. How reliable are the outputs of each AI model when they are prompted repeatedly in the same manner?
3. To what degree do AI models vary in accuracy and reliability relative to the original human-led data analysis?

The study was carried out in four stages. First, the data were gathered from 66 high-school students through an online questionnaire where the students were asked questions regarding how they believed that AI would be integrated into math and science in schools in the future. Second, the team performed a thematic analysis on the students’ responses, which produced three themes: (1) creation of exercises for practice, (2) interactive learning (including games, experiments, and simulations), and (3) demonstrations of different problem-solving approaches to support students’ understanding. In the third stage, the researchers tested whether various LLMs could recover and categorize the same themes. To account for potential variability across models and across repeated prompting of the same model, the team ran these prompts 10 times on 13 LLMs, including ChatGPT o1 Pro, ChatGPT o3, ChatGPT o3 mini (H), ChatGPT o1, Claude 3.7 Sonnet, ChatGPT o1 mini, ChatGPT 4.5 preview, Gemini 2.o Pro, Gemini 2.0 flash, Deepseek-R1, ChatGPT 4o, Llama 3.1, and Llama 3.3. Finally, the researchers objectively measured each model’s fidelity to the human-established categories by checking whether the LLMs’ categories corresponded to those generated by researchers, and whether the reported number of matched instances aligned with the actual matches.

Most LLMs demonstrated strong consistency. ChatGPT o1 Pro, ChatGPT o1, ChatGPT o3, Claude 3.7 Sonnet, ChatGPT o1 mini, ChatGPT 4.5 preview, Gemini 2.0 flash, and ChatGPT 4o perfectly replicated the three categories in all 10 iterations. Some models produced minor divergences: ChatGPT o3 mini (H) struggled slightly with “Demonstrations of multiple problem-solving methods,” reaching 7 out of 10 correct categorizations. Llama 3.1 aligned with the human-defined categories for “Creation of exercises for practice” 8 times out of 10, but only 6 times for the other two themes. Deepseek-R1 and Llama 3.3 completed just three runs, even though prompted for 10, yet each correctly reproduced the categories in all three attempts. Gemini 2.0 Pro presented a separate issue. While it produced 10 sets of 3 categories each (totaling 30 categories), its summary table listed more than 30 instances. Specifically, it recorded 12 instances of “Creation of exercises for practice,” 11 of “Interactive Learning,” and 12 of “Demonstrations,” suggesting redundant categorization or a misunderstanding of the prompt. The average reliability within each LLM equaled 0.95, while Cohen’s Kappa ranged from 0.92-0.96 according to the themes that were analyzed. Further analyses will also be performed to compare the LLM outcomes from a validity perspective.

European Commission (2024). Living guidelines on the responsible use of generative AI in research (ERA Forum Report, March 20, 2024). Brussels: European Commission.