Generative AI technologies such as large language models show novel potential to enhance educational research. For example, generative large language models were shown to be capable of solving quantitative reasoning tasks in physics and concept tests such as the Force Concept Inventory (FCI). Given the importance of such concept inventories for physics education research, and the challenges in developing them such as field testing with representative populations, this study seeks to examine to what extent a generative large language model could be utilized to generate a synthetic dataset for the FCI that exhibits content-related variability in responses. We use the recently introduced ChatGPT based on the GPT 4 generative large language model and investigate to what extent ChatGPT could solve the FCI accurately (RQ1) and could be prompted to solve the FCI as if it were a student belonging to a different cohort (RQ2). Furthermore, we study, to what extent ChatGPT could be prompted to solve the FCI as if it were a student having a different force- and mechanics-related preconception (RQ3). In alignment with other research, we found that ChatGPT could accurately solve the FCI. We furthermore found that prompting ChatGPT to respond to the inventory as if it belonged to a different cohort yielded no variance in responses, however, responding as if it had a certain preconception introduced much variance in responses that approximate real human responses on the FCI in some regards.

[This paper is part of the Focused Collection on Qualitative Methods in PER: A Critical Examination.] Qualitative research methods have provided key insights in physics education research (PER) by drawing on non-numerical data such as text or video data. While different methods towards qualitative research exist, they share two essential steps: recognizing patterns in the data and interpreting these patterns. Although these methods have led to the development of rigorous theory, there are challenges: As such methods require a series of judgments by the analyst, they are difficult to validate and reproduce. Further, they are hard to scale so that they are unavailable to the analysis of large-scale data. In this way, important phenomena may remain inaccessible to qualitative analysis. Reacting to these challenges and leveraging the potential of emerging methods of artificial intelligence (AI) such as machine learning and natural language processing, sociologist Nelson has proposed the concept of computational grounded theory (CGT). CGT proceeds in a process of three consecutive steps: In the first step, one leverages the power of computational techniques, especially natural language processing and unsupervised machine learning techniques, for pattern detection in large datasets—those of a size and scope that may prohibit human-driven analysis from the outset. In the second step, one relies on the integrative and interpretative capabilities of human researchers to add quality and depth to the quantity and breadth of the first step. In the last step, one again uses computational techniques to test the extent to which the detected and refined patterns from the first and second step hold throughout the whole dataset under investigation. Interestingly, CGT does not aim at simply automating parts of the qualitative process by using AI, but rather aims at integrating AI into the human analyst’s workflow within a qualitative analysis. This leads to an analytical system that can do something that is quantitatively and qualitatively different from what a human or machine can do alone. In this way, CGT aims at addressing questions about validity, reproducibility, and scalability in qualitative research while preserving the theoretical sensitivity and unique inferencing capabilities of the human analyst. In this paper, we provide a primer on CGT, present how it can be used to investigate the physics problem-solving approaches of $N=417$ students based on textual data, and discuss CGT’s potentials and challenges in PER. In consequence, this paper can provide critical input to the discussion of how emerging AI technologies can provide new avenues in qualitative PER.

This study presents findings on the physics identity development of female students in the German Physics Olympiad who participated in an intervention designed to support their engagement in physics. Enrichment programs such as the Physics Olympiad have been found to positively impact students’ engagement and intent to pursue a career in science. However, many enrichment programs, including the Physics Olympiad, suffer from an underrepresentation of young women. The intervention investigated in this study capitalizes on gender-related research in physics education in order to explore ways in which gender equity can be raised in enrichment programs. To this end, we designed an identity-safe learning environment that facilitates participating young women’s physics identity development. For example, same-sex groupings and considerately adopted physics contents that particularly acknowledge young women’s interests (e.g., relation to medical issues and the human body) were utilized. Overall, 30 Olympians took part in a one-day intervention (13 females, 17 males). Positive effects in two important physics identity constructs, namely, interest and competence, for young women were found, while at the same time no effects were found for young men. Furthermore, the young women were more likely to participate in next year’s Physics Olympiad, compared to the overall female Physics Olympiad population. These results indicate that the careful design of an intervention based on gender research and science identity theory can support young women’s physics identity development.

Science education researchers have developed a refined understanding of the structure of science teachers’ pedagogical content knowledge (PCK), but how to develop applicable and situation-adequate PCK remains largely unclear. A potential problem lies in the diverse conceptualisations of the PCK used in PCK research. This study sought to systematize existing science education research on PCK through the lens of the recently proposed refined consensus model (RCM) of PCK. In this review, the studies’ approaches to investigating PCK and selected findings were characterised and synthesised as an overview comparing research before and after the publication of the RCM. We found that the studies largely employed a qualitative case-study methodology that included specific PCK models and tools. However, in recent years, the studies focused increasingly on quantitative aspects. Furthermore, results of the reviewed studies can mostly be integrated into the RCM. We argue that the RCM can function as a meaningful theoretical lens for conceptualizing links between teaching practice and PCK development by proposing pedagogical reasoning as a mechanism and/or explanation for PCK development in the context of teaching practice.

Constructing arguments is essential in science subjects like chemistry. For example, students in organic chemistry should learn to argue about the plausibility of competing chemical reactions by including various sources of evidence and justifying the derived information with reasoning. While doing so, students face significant challenges in coherently structuring their arguments and integrating chemical concepts. For this reason, a reliable assessment of students' argumentation is critical. However, as arguments are usually presented in open‐ended tasks, scoring assessments manually is resource‐consuming and conceptually difficult. To augment human diagnostic capabilities, artificial intelligence techniques such as machine learning or natural language processing offer novel possibilities for an in‐depth analysis of students' argumentation. In this study, we extensively evaluated students' written arguments about the plausibility of competing chemical reactions based on a methodological approach called computational grounded theory. By using an unsupervised clustering technique, we sought to evaluate students' argumentation patterns in detail, providing new insights into the modes of reasoning and levels of granularity applied in students' written accounts. Based on this analysis, we developed a holistic 20‐category rubric by combining the data‐driven clusters with a theory‐driven framework to automate the analysis of the identified argumentation patterns. Pre‐trained large language models in conjunction with deep neural networks provided almost perfect machine‐human score agreement and well‐interpretable results, which underpins the potential of the applied state‐of‐the‐art deep learning techniques in analyzing students' argument complexity. The findings demonstrate an approach to combining human and computer‐based analysis in uncovering written argumentation.

Computer-based analysis of preservice teachers’ written reflections could enable educational scholars to design personalized and scalable intervention measures to support reflective writing. Algorithms and technologies in the domain of research related to artificial intelligence have been found to be useful in many tasks related to reflective writing analytics such as classification of text segments. However, mostly shallow learning algorithms have been employed so far. This study explores to what extent deep learning approaches can improve classification performance for segments of written reflections. To do so, a pretrained language model (BERT) was utilized to classify segments of preservice physics teachers’ written reflections according to elements in a reflection-supporting model. Since BERT has been found to advance performance in many tasks, it was hypothesized to enhance classification performance for written reflections as well. We also compared the performance of BERT with other deep learning architectures and examined conditions for best performance. We found that BERT outperformed the other deep learning architectures and previously reported performances with shallow learning algorithms for classification of segments of reflective writing. BERT starts to outperform the other models when trained on about 20 to 30% of the training data. Furthermore, attribution analyses for inputs yielded insights into important features for BERT’s classification decisions. Our study indicates that pretrained language models such as BERT can boost performance for language-related tasks in educational contexts such as classification.