The study, published in the Q1/D1-ranked journal Computers and Education: Artificial Intelligence published by Elsevier, examined the factors that influence students’ intention to continue using ChatGPT for academic purposes. The research was conducted by Chaman Verma, a member of the Department of Media and Educational Informatics at the Faculty of Informatics, as part of the University Excellence Scholarship Program (EKÖP-24). The work arrives at a moment when universities around the world are still wrestling with how to position generative AI within teaching, assessment, and academic integrity frameworks, making evidence about the drivers of sustained student use particularly timely.

At the heart of the investigation was a dataset built from questionnaire responses provided by 166 university students. Questionnaire-based studies of technology acceptance have a long pedigree in educational research, but the ELTE study distinguished itself through its analytical approach. Instead of applying only conventional regression techniques, the researcher employed interpretable machine learning methods to examine which factors play a decisive role in students’ intention to continue using ChatGPT in their learning. Interpretability is a crucial consideration here: machine learning models can be powerful at detecting patterns, but if a model behaves as a black box, its findings are of limited use to educators and policymakers who need to understand why certain factors matter, not just that they do.

The results of the analysis point to a cluster of key factors that determine whether students intend to keep ChatGPT in their academic lives. These include learning support, accessibility, academic engagement, usability, and the perceived reliability of the tool. In practical terms, the findings suggest that students are more likely to incorporate ChatGPT into their learning processes if they perceive it as genuinely useful, an instrument that helps them understand complex concepts, supports them in completing assignments, and contributes to more effective learning. Usefulness, in other words, is not an abstract judgment but a concrete experience of the tool helping with real academic tasks.

Each of the identified factors tells its own story about the student experience of generative AI. Learning support captures the degree to which ChatGPT functions as a study companion, something that can explain a difficult theorem, draft an outline, or translate dense material into more accessible language. Accessibility reflects how easily students can reach the tool when they need it, a consideration that has grown less restrictive as free and widely available versions of large language models have proliferated. Academic engagement speaks to whether interacting with the tool deepens or dilutes a student’s involvement with their coursework, while usability concerns the friction of the interaction itself, from the clarity of prompts to the quality of responses. Perceived reliability rounds out the picture, capturing how much students trust what the system tells them.

Beyond usefulness, the research highlights that trust is also key. The accuracy of responses, the reliability of the tool, and academic integrity remain central issues in the use of generative artificial intelligence in higher education. Students’ attitudes towards ChatGPT are influenced not only by how effectively the tool can support their learning, but also by how reliable they perceive it to be and how they view the possibilities for using it responsibly in an academic environment. This dual lens, usefulness on one side and trustworthiness on the other, mirrors a broader tension in the deployment of large language models: the same systems that can produce impressively fluent explanations can also generate confident but incorrect statements, a phenomenon that has become one of the most discussed challenges in applied artificial intelligence.

The academic integrity dimension adds a further layer of complexity. Universities have responded to the arrival of generative AI with a spectrum of policies, ranging from outright prohibition to structured integration into curricula, and students are acutely aware of where the boundaries lie. The ELTE findings suggest that a student’s willingness to continue using ChatGPT is bound up with their sense of whether such use is legitimate and responsible within their institution. A tool that students find helpful but ethically fraught may face very different adoption trajectories than one they perceive as both useful and sanctioned, and the interpretable machine learning approach allowed the study to weigh these considerations alongside more purely pragmatic factors.

Methodologically, the study contributes to a growing movement within educational data science that favors models whose reasoning can be inspected and communicated. Interpretable machine learning sits between the transparency of classical statistics and the predictive strength of more complex algorithms, offering researchers a way to capture nonlinear relationships and interactions among variables, for instance, how usability might amplify the effect of perceived usefulness, without sacrificing the ability to explain which inputs drove the predictions. For a question as policy-relevant as the determinants of continued AI use among students, that combination of predictive power and explainability is what makes the findings actionable rather than merely descriptive.

The study contributes new findings to research on the use of artificial intelligence in education and offers useful, evidence-based insights for educators, higher education institutions, and policymakers. As generative AI tools continue to evolve, understanding the conditions under which students choose to sustain their use, and the trust and integrity concerns that could temper it, will be essential for designing guidance, training, and institutional policy. The ELTE team’s work suggests that the path forward lies in ensuring that generative artificial intelligence tools can responsibly, consciously, and effectively support teaching and learning in higher education in the future, with usefulness and reliability advancing together rather than in competition.

For the wider conversation about AI in academia, the message from the 166 students behind this dataset is ultimately a measured one. They are prepared to embrace ChatGPT as a learning aid when it demonstrably helps them understand, engage, and perform, but their continued adoption depends on the tool earning and keeping their trust. As institutions refine their approaches to generative AI, studies of this kind, grounded in student data and analyzed with methods that make their conclusions visible, provide a template for evidence-based decision-making in a field where the technology is changing faster than the rules that govern it.

Subject of Research: Determinants of university students' continued use of ChatGPT for learning, analyzed with interpretable machine learning

Article Title: ChatGPT at university: What determines whether students will continue using it?

Article References: ChatGPT at university: What determines whether students will continue using it?. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: ChatGPT, generative artificial intelligence, higher education, interpretable machine learning, technology acceptance, learning support, academic integrity, perceived reliability, usability, student engagement, Eötvös Loránd University, educational data science

That gap is the starting point for a research article published on 22 September 2025 in Frontiers of Digital Education, a Springer journal, by Haoxuan Li of Beihang University, Jifan Yu of Tsinghua University and colleagues. The team reformulates knowledge tracing as a new task they call explainable few-shot knowledge tracing, and proposes a cognition-guided framework built on large language models that can track a student’s knowledge state from only a few exercise records while producing natural language explanations of its conclusions. Across three widely used benchmark datasets, the authors report that large language models perform comparably to, or better than, competitive deep knowledge tracing methods, despite operating under conditions that would cripple conventional models.

To appreciate why this matters, it helps to understand how traditional knowledge tracing works. The field traces its lineage to Bayesian knowledge tracing, introduced by Albert Corbett and John Anderson in 1994, which treats each skill as a binary state, learned or unlearned, and updates the probability of mastery each time a student answers a question. From 2015 onward, deep knowledge tracing replaced these hand-built probabilistic updates with recurrent neural networks that learn hidden representations of student ability directly from long sequences of responses. Later refinements added self-attention, graph neural networks that exploit relationships between exercises, and memory networks with dynamic key-value stores. These models are powerful, but they share two structural dependencies: they need extensive interaction data per student to converge, and their output is a bare number, a predicted probability of a correct answer, with no account of why.

Both dependencies clash with teaching practice. A teacher assessing a student in a tutoring session sees perhaps a dozen attempts, not thousands, and must nonetheless form a judgment. That judgment is then communicated as feedback: the student confuses the distributive property with the associative property, or consistently misapplies a sign rule when moving terms across an equation. The Beihang and Tsinghua authors describe this mismatch as current methods falling into the cracks between laboratory benchmarks and real-world pedagogy. Their response is to redefine the task itself: instead of predicting a numerical score from abundant data, the model should infer a student’s cognitive state from sparse evidence and articulate that inference in language a teacher or student can read.

Large language models are, in a technical sense, an unexpected but fitting instrument for this reformulation. Models in the lineage of GLM and LLaMA are pretrained on vast text corpora and exhibit emergent abilities in reasoning and generation, meaning they can perform tasks they were never explicitly trained for. Crucially, they carry prior knowledge about academic subjects themselves: what a quadratic equation is, what concept a fraction problem tests, which misconceptions typically arise. A conventional deep knowledge tracing model sees only anonymized question identifiers and correctness bits; a language model sees the actual content of the exercise and can reason about the cognitive skill it probes. This allows the framework to lean on semantic understanding rather than sheer volume of interaction history, which is precisely what the few-shot setting demands.

The architecture the researchers propose is cognition-guided, a design choice that distinguishes it from simply prompting a chatbot with a transcript. The framework structures the model’s reasoning around cognitive dimensions of learning, guiding the language model to decompose a student’s performance into mastery of specific knowledge components rather than emitting an undifferentiated guess. The student’s few exercise records are rendered into a structured prompt, the model reasons over which underlying skills each question engages, and it then generates both a prediction of future performance and a natural language explanation tracing the evidence for its judgment. In effect, the explanation is not bolted on after the fact, as with post-hoc interpretability techniques applied to neural networks, but emerges from the same reasoning chain that produces the prediction.

The empirical evaluation is where the claim becomes concrete. The authors tested their framework on three widely used knowledge tracing datasets, benchmarking against competitive deep learning baselines drawn from the field’s standard toolkit, including models assembled through the PYKT benchmarking library. Under few-shot conditions, where each student contributes only a small number of records, the language model-based approach achieved results comparable to or superior than the deep baselines, which typically require far more data to reach their reported performance. The significance is twofold. Practically, it suggests that meaningful student modeling may be possible in cold-start scenarios, new courses, new platforms, or individual tutoring, where deep models have historically floundered. Scientifically, it demonstrates that the semantic knowledge embedded in pretrained language models can substitute, at least in part, for the statistical signal that large datasets normally provide.

The study does not present itself as a finished solution. The authors explicitly discuss potential directions and call for future improvements, acknowledging that the field is at an early stage of understanding how language models should be adapted to educational measurement. Open questions loom large. Language models can hallucinate, producing confident but wrong explanations, and in an educational context an incorrect explanation of a student’s misconception could misdirect instruction. The cost of running large models at scale across millions of learners is nontrivial compared with compact neural networks. And the psychometric tradition, from item response theory to the Rasch model, has spent a century building rigorous measurement theory that these new generative approaches have not yet absorbed. The authors’ framing, grounded in the established literature on educational assessment, suggests they see their work as a bridge between that tradition and modern generative AI rather than a replacement for it.

Even so, the implications ripple outward. Explainability is becoming a regulatory and ethical requirement for AI systems that make consequential decisions about people, and few decisions are more consequential than those shaping a student’s education. A model that can say, in plain language, that a student has mastered linear equations but struggles with word problems gives teachers something actionable; a probability output gives them nothing. The work also joins a broader wave of research applying large language models to education, from computerized adaptive testing to automated feedback on mathematics responses, indicating that the generative AI era may reshape educational assessment as profoundly as the deep learning era did. If the trend holds, the next generation of tutoring systems may not merely predict what a student will get right, but explain, in a teacher’s own vocabulary, what the student knows, what they are missing, and what to do next. That, the authors suggest, is the standard real teaching has always demanded, and one that machine learning is only now beginning to meet.

Subject of Research: Using large language models for explainable few-shot knowledge tracing in educational assessment

Article Title: Explainable Few-Shot Knowledge Tracing

Article References: Explainable Few-Shot Knowledge Tracing. (n.d.). https://doi.org/10.1007/s44366-025-0071-x

Image Credits: AI Generated

DOI: 10.1007/s44366-025-0071-x

Keywords: knowledge tracing, large language models, explainability, educational assessment, few-shot learning, student modeling, deep learning, intelligent tutoring systems, natural language explanations, educational data mining, cognitive modeling, personalized education

The research team, led by Conrad Borchers of Carnegie Mellon University together with Fitore Morina, Lennart Klein, and Christian Fischer of the University of Tübingen, faced a fundamental methodological challenge: how do you systematically discover education-related communities on a platform that does not label them? Their solution was a novel, reproducible sampling pipeline. Starting from Germany’s most popular teacher community, the Twitterlehrerzimmer, known by its hashtag #twlz and translating roughly as Twitter’s teacher lounge, the team harvested co-occurring hashtags and expanded outward through naming conventions tied to academic subjects, German federal states, and broader education-related suffixes. This snowball procedure generated 21,893 candidate hashtags, of which 1,074 produced actual post volume, ultimately yielding 100 qualifying communities after strict exclusion criteria removed groups with fewer than 100 original tweets, apparent participation of minors, predominantly non-educational content, or predominantly non-German language.

The scale of the resulting dataset allowed the researchers to answer questions that interviews and surveys simply cannot address. Every tweet is time-stamped and traceable, capturing likes, reposts, replies, and mentions as behavioral trace data rather than potentially biased self-reports. To classify participants, the team trained a machine learning classifier on 1,000 hand-coded user profiles, achieving strong precision in distinguishing teachers from non-teachers. A conservative deterministic algorithm flagged 187 bot accounts, including one that automatically retweeted every #twlz post, which together were responsible for 261,160 tweets, nearly all of them reposts. After filtering, the human landscape emerged with striking clarity: teachers made up only 7.4 percent of users, yet they generated almost half of all tweets across the communities, and 47.6 percent of all posts in the teacher lounge itself.

The map revealed a distinctly German geography of professional learning. The #twlz community dwarfed everything else with over 2 million tweets and roughly 98,000 users. Subject-specific communities followed, led by German language teachers with about 146,000 tweets, then Religious Studies with roughly 85,000, and Computer Science with about 39,000, the latter including hashtags campaigning for computer science to become a compulsory school subject. Regional communities tied to federal states formed a third layer, with Bavaria dominating at over 208,000 tweets, followed by North Rhine-Westphalia and Baden-Württemberg. Interestingly, when adjusted for population, the smaller city-states of Berlin, Bremen, and Hamburg showed the highest proportional activity. This decentralized network of partially overlapping communities contrasts with the United States, where participation tends to concentrate around a few large, recurring hashtags such as #EdChat and #NGSSchat.

Temporal patterns added further texture. The teacher lounge grew exponentially beginning around 2017, expanding from 176 active users in 2016 to more than 22,500 in 2020, with annual tweet volume increasing nearly 96-fold in that window. The rise of #twlz coincided with a decline in EdchatDE, Germany’s weekly scheduled chat community, which peaked around 60,000 tweets in 2015. Posting activity tracked the rhythms of the school year, dipping during the summer holidays and peaking in February, March, October, and November. Teachers and non-teachers posted at similar times of day, but teachers remained notably active after regular working hours and on weekends, underscoring the argument that these communities offer professional learning at times convenient to teachers rather than dictated by institutional schedules.

Conversation dynamics differed sharply across community types. The asynchronous teacher lounge and state-based communities sustained discussions more than twice as long as the scheduled EdchatDE chats, with mean conversation lengths of roughly 10.5 tweets compared with 4.5. Subject-specific conversations had the longest half-life, the time for half of a thread’s tweets to accumulate, at over 52 hours, suggesting slower, more spaced-out exchanges. Sentiment analysis, validated against human-coded German tweets using the SentimentWortschatz dictionary, showed an overall positive environment: for every 100 positive or neutral tweets, only 32 were negative. The teacher lounge carried the highest negativity ratio, which the authors interpret not as toxicity but as a space where teachers feel safe voicing professional frustrations alongside requests for support.

The study’s theoretical centerpiece asked whether #twlz satisfies the three defining characteristics of a Community of Practice, the influential informal learning model developed by Jean Lave and Étienne Wenger: a shared domain, a genuine community, and a joint repertoire of practice. On the domain criterion, the evidence was quantitative and direct. Of 47 pre-validated teacher-specific German terms, ranging from inclusive words for students to terminology for distance learning, 42.3 percent of all teacher-lounge tweets contained at least one, and teachers used them more frequently than non-teachers. Among the 1,000 most engaged tweets, nearly 70 percent dealt with profession-specific situations or the sharing of teaching resources, precisely the kind of shared professional identity and expertise the model predicts.

The community criterion emerged from social network analysis of nearly 3.8 million directed interactions. Highly influential users, measured by betweenness centrality, teachers, and veterans, defined as the earliest half of members to join, all preferentially interacted within their peer groups, while newer and less central users gravitated toward them, reproducing the trajectory from legitimate peripheral to fuller central participation that Lave and Wenger described. Teachers were significantly overrepresented among both the most influential and the most veteran members. Perhaps most strikingly, among the most engaged tweets, more than half expressed emotional support, including a teacher seeking comfort after learning a student faced deportation. The practice criterion was corroborated by linguistic analysis showing rising social references concurrent with the community’s 2017 growth, and topic modeling revealing shared themes, such as digital tools and pandemic schooling, discussed with distinctly different vocabularies by teachers and non-teachers.

The researchers took unusual care with ethics and privacy, replacing all platform identifiers with anonymous ones, masking email addresses and phone numbers, and paraphrasing rather than quoting tweets to prevent search-engine re-identification. This methodological transparency extends to the sampling pipeline itself, which the authors have released as reproducible code. They argue it serves as a blueprint for mapping educational communities in other countries and languages, addressing a long-standing gap in a literature dominated by United States contexts. Limitations remain, including the snapshot nature of engagement counts, the restriction to a single platform, and the need to validate community definitions against users’ own self-perceptions, all of which the authors flag for future work.

The broader significance is considerable. Recent meta-analytic evidence indicates that online teacher professional development produces medium-sized effects on teacher learning and smaller but tangible effects on student outcomes. By demonstrating that a national teacher community on X exhibits the structural signatures of a genuine Community of Practice, shared language, sustained reciprocal interaction, and collaboratively developed practice, this study helps explain why such informal platforms work. For educational policymakers, the message is that teacher participation in social media communities deserves recognition as a legitimate and measurable form of professional development, one that operates on teachers’ own time, across geographic boundaries, and at a scale no workshop could ever match.

Subject of Research: Large-scale mapping of educational Twitter/X communities and informal teacher professional learning in Germany

Article Title: Mapping the landscape of educational use of X (Twitter) in Germany: informal teacher learning in online communities of practice

Article References: Borchers, C., Morina, F., Klein, L., & Fischer, C. (2025). Mapping the landscape of educational use of X (Twitter) in Germany: informal teacher learning in online communities of practice. Journal of New Approaches in Educational Research, 14(1), Article 27. https://doi.org/10.1007/s44322-025-00047-8

Image Credits: AI Generated

DOI: 10.1007/s44322-025-00047-8

Keywords: teacher professional development, Twitter, communities of practice, informal learning, social media, Germany, educational data science, teacher collaboration, online communities, sentiment analysis, social network analysis, Mapping