Gaze is foregrounded as a central interactional resource positioning individuals as speakers or bystanders. In early childhood education (ECE), where direct involvement in interactions with the teacher is central to language development and where unequal participation may lay the groundwork for far-reaching processes of educational inequality, research into gaze remains scarce. This study, including six teachers and thirty children, examines variation in children’s involvement during book reading activities and analyses how teachers shape these conversational dynamics. Using epistemic network analysis alongside multimodal quantitative and qualitative conversation analyses, this study provides a fine-grained account of multimodal turn-taking. The findings reveal unequal distribution of teacher attention, resulting in unequal opportunities for children to participate. Results revealed a relationship between teachers’ fixations (i.e. sustained moments of visual attention) and both teachers’ and children’s speaking turns, underscoring the central importance of gaze in conversational dynamics in classrooms. Conversation analysis further shows that children’s participation is largely teacher-driven, with most child turns being initiated by the teacher. Overall, the results indicate interactional inequalities in ECE.

<p>As he instructs bilingual learners in the humanities and social studies, Daniel Gomez sees big promise in artificial intelligence.</p><p>“Teachers in general cannot be knowledgeable about every single subject,” and certainly not in multiple languages, says Gomez, who teaches at the bilingual K-12 Ideal School in Tampa, Florida. AI can help close the gaps. “It will be able to pick up any topic students are interested in and handle the topic reasonably well.”</p><p>He isn’t alone in this assessment. Harvard researchers, for example, are actively exploring how AI-powered chatbots can support bilingual student learning and parent–child dialogues.</p><p>“There are lots of opportunities to use AI to adapt learning materials into different languages, which can expand access for bilingual learners,” says<a href="https://www.gse.harvard.edu/directory/faculty/ying-xu"> <u>Ying Xu</u></a>, an assistant professor of education at Harvard and one of the researchers involved in the work.</p><p>“Some projects are also exploring how AI can provide additional scaffolding to help children comprehend content in their home language,” she adds. “For example, a multilingual language model might keep the core learning content in English while offering explanations, prompts or supports in the child’s home language.”</p><p>Experts see great potential for AI to elevate bilingual education — if it’s done right.</p><h2><strong>Chatbots as Educational Aids</strong></h2><p>“Bilingual learners are obviously learning the language and culture of the classroom and the environment that they’re in,” says<a href="https://www.linkedin.com/in/krystlesalas/"> <u>Krystle Salas</u></a>, assistant director of special populations at Second Mile, which operates alternative education charter schools for at-risk students in several states.</p><p>“They don’t necessarily have access to the instructional materials or even just the conversations going on in the classroom — and certainly not at the level of their p

This commentary begins by summarizing Jeon et al.’s (2025) concerns about Generative Artificial Intelligence (GenAI)’s tendency to standardize language, followed by Hu’s (2026) response, on how to bring Artificial Intelligence (AI) into teaching (academic) writing in ways that allow for negotiation of forms. My own response, given the bigger picture, is whether AI needs to be used at all. First, as a former English for Academic Purposes (EAP) writing instructor, I argue that AI represents writing tasks differently from humans, and that going along with its simplified representation is a step back after human experts have rendered more complex representations of the tasks. Second, as a qualitative researcher, I argue that AI is of limited use in qualitative data analysis, since what qualitative researchers analyze goes beyond what is in the text. I conclude that in both these cases, AI simplifies tasks and detracts from what students of academic writing or novice qualitative researchers need to learn.

Automated essay scoring (AES) research has largely focused on cross-prompt generalization, where essays from unseen prompts are scored while the scoring criteria are typically held constant. In practice, however, educators may revise or even introduce new rubrics in their scoring task, to evaluate different aspects of essays. We study cross-rubric generalization: training on essays labeled under one set of rubrics and evaluating on previously unseen rubrics, which target different aspects of the essay. We use a Large Language Model (LLM) fine-tuning framework with two components: rubric-agnostic intermediate representations, called traits, and target-essay supervision under seen rubrics during training. On an AES dataset augmented with multiple rubric-defined labels of student critical thinking skills, we find that traits improve macro F1 by 5.0% over a baseline without traits in the hardest setting, where both target rubrics and target essays are unseen during training. We further find that increasing target-essay supervision improves performance, with our best fine-tuned open-source Llama-based model outperforming GPT-5-mini prompting by 2.1% macro F1 and trailing GPT-5 by 1.9%. These results show that trait-based intermediate structure and controlled supervision improve generalization to unseen rubrics.

Grounded in the L2 motivational self system theory, this quasi‐experimental study investigated the impact of three motivational interventions on two outcome measures: L2 learning effort and L2 achievement. A total of 391 Chinese junior secondary students were assigned to four conditions: a control group receiving no motivational instruction, and three experimental groups exposed to either traditional motivational strategies (MSs), vision‐based MSs, or a combined approach integrating both types of strategy. The effects of the interventions were assessed through questionnaires and tests administered at three time points to monitor changes in both outcome variables. The respective and relative effects of the interventions were determined by two‐way mixed design ANOVAs and post‐hoc pairwise comparisons. Results indicated that all three motivational treatments produced significantly positive effects on both outcome variables immediately following the interventions. All three motivational treatments sustained their positive effects on L2 learning effort, but only the vision‐based and combined treatments maintained their impact on L2 achievement. Moreover, all motivational treatments had comparable effects on L2 learning effort both during and 4 months after the interventions. Although all interventions were similarly effective on L2 achievement during the treatment period, the vision‐based and combined interventions outperformed the traditional MS intervention in terms of sustained effects.

<p>Microsoft just released the third edition of its AI in Education Report, surveying more than 3,000 students, educators and leaders across six countries. Ira Apfel sat down with <a href="https://conference.iste.org/2026/program/search/detail_presenter.php?id=118445480">Pat Yongpradit</a>, general manager of global education and workforce policy at Microsoft, which sells AI tools including Copilot to schools, at <a href="https://conference.iste.org/2026/">ISTELive 26</a> in Orlando, to talk through what the data reveals about daily use, trust, training gaps, and where the policy backlash against AI in schools may be headed next.</p><h2>Pat Yongpradit on Adoption, Trust and Cheating</h2><p>Yongpradit walks through why daily AI use in schools still lags far behind overall adoption, even though nine in 10 educators, students and leaders report having tried the tools at least once. He addresses the recent drop in student optimism, framing it as a predictable dip on a familiar technology adoption curve rather than a warning sign.</p><p>The conversation also covers the wide gap between how much training school leaders believe their staff have received and what teachers and students say they have actually experienced. Yongpradit also shares his candid take on academic integrity, arguing that the panic around cheating has more to do with long-standing temptation than with any single tool.</p><p>Listen to the episode:</p><h2>Stories Mentioned in This Episode</h2><p><a href="https://www.linkedin.com/posts/patyongpradit_2026-ai-in-education-report-ugcPost-7475673037405679616-_qCk/"><u>2026 AI in Education Report</u></a> by Microsoft</p><p><strong>This Week with EdSurge</strong> is a weekly podcast from <strong>EdSurge</strong>. Subscribe to <a href="https://www.edsurge.com/newsletters">EdSurge newsletters</a> for more news and analysis on education and technology.</p>

The rise of Artificial Intelligence (AI) enables automatic analysis of large amounts of data. Previously time-consuming and labor-intensive tasks can be completed much more efficiently with the use of AI. This work uses AI techniques to analyze and revise curricular patterns in an undergraduate degree for Software Engineering. Curricula often have long sequences where failure to pass a class within the sequence may jeopardize completion of the degree within four years. Manual analysis and revision of curricula by university faculty is a lengthy and labor-intensive process, causing changes to occur rarely and making it impossible to keep up with the changing needs of students. This work reduces the time-to-change for curricula and reduces bottlenecks and graduation delays by using Large Language Models (LLMs) to analyze curricular patterns and suggest revisions.

This study proposes a semi-automated system for generating dialogue-based lessons using Large Language Models (LLMs) and Text-to-Speech (TTS) technology, and exploratorily examines its educational potential via a practical quasi-experiment. The system augments rather than replaces educators through a three-stage human-in-the-loop workflow (LLM-based slide/narration generation, educator review, automated audiovisual integration), and introduces a novel method for generating Expert-Novice dialogue narration based on cognitive apprenticeship theory. In a study of 245 first-year high school students who sequentially experienced three lesson formats (instructor voice, single-speaker TTS, dialogue TTS; content differed across sessions, limiting format/content separation), we conducted within-subject (Friedman test, N<=183) and repeated cross-sectional (Mann-Whitney U, N=229/206) analyses. TTS audio did not substantially degrade the learning experience versus instructor voice, supported by TOST equivalence testing. Dialogue TTS was significantly superior to single TTS in comprehension (p=.006, q=.025) and cognitive engagement (p=.019, q=.048); enjoyment was non-significant after FDR correction (q=.081) but reached significance after controlling for prior knowledge (proportional-odds model, OR=1.65, q=.025), and these advantages were not attributable to prior-knowledge imbalance. Conversely, single TTS was superior in audio naturalness (p<.001, q<.001, r=-.238), revealing a trade-off between dialogue's benefits and higher extraneous cognitive load. Dialogue format was preferred by 66.9% of learners as most enjoyable (p<.001). These results reflect a fixed-order design; replication is needed before generalizing them as effects of lesson format. This study provides a theoretical and empirical basis for the educational acceptability of TTS audio and for TTS lesson-format design.

Student evaluations of teaching can be biased by the ‘way’ the instructor speaks or their accent. In the current study, we investigate whether a brief awareness-based intervention can be effective in reducing accent-based bias in students’ evaluations of instructors. In a series of controlled laboratory experiments, college student participants were randomly assigned to watch a short lecture narrated by either an instructor who spoke English with a Mandarin accent or an American accent. Prior to evaluating the instructor, participants in the intervention condition were given instructions designed to reduce bias, whereas participants in the control condition proceeded directly to the evaluation form. Although there were no statistically significant differences in performance on a multiple-choice assessment of learning, participants in both experiments showed evidence of bias, rating the Mandarin-accented instructor more negatively than the American-accented instructor. In Experiment 1, the intervention reduced the disparity in participants’ overall ratings of the two instructors; however, in Experiment 2, the intervention was less effective with different lecture materials and a more diverse sample of students. Awareness-based interventions may show promise in reducing real-world biases in college classes and could support efforts to improve diversity and inclusivity in higher education.

Online mathematics learning platforms are increasingly adopting large language models (LLMs) to provide scalable, on-demand support, but these models may reproduce private information from training data or generate harmful language. This raises concerns about responsibility in educational settings regarding the use of pre-trained models. LLM unlearning is an emerging area for reducing a model’s ability to produce specific unwanted content and remains underexplored in educational research. This study aims to investigate how LLM unlearning reduces the model's reliance on personally identifiable information (PII) and inappropriate content in the math tutoring context, while maintaining the model's utility on both single-label and multi-label downstream math tasks. We applied a gradient-based LLM unlearning approach to three different models, which were pre-trained on approximately 3 million data points from an Algebra I online discussion forum between students and professional tutors. PII and harmful content were detected on this training data and used for unlearning in two different orders (PII and harmful content unlearning). Then, the generated outputs from these two unlearning models were compared with those of the pre-trained model in terms of PII-containing output rate and harmful rate. Moreover, unlearned models were evaluated on two different math classification tasks. The results showed that the rates of PII-containing output rate and harmfulness substantially decreased compared to the pre-trained models, and the utility of the unlearned model was still maintained. These findings demonstrate how LLM unlearning can be applied to pre-trained models to behave them more responsibly, while maintaining strong model performance on math-related tasks.

Large language models (LLMs) are increasingly used in educational technology for automated writing assessment, yet most applications rely on prompt-based scoring and feedback generation, which often lack transparency, reproducibility, and interpretability. This study investigates whether model-internal LLM representations can provide interpretable and reproducible metrics for second language (L2) writing assessment. We derived surprisal and perplexity from next-token prediction to quantify linguistic predictability and embedding-based similarity to measure semantic coherence across sentences. These metrics were computed at the token, sentence, and discourse levels using three pretrained Chinese language models and evaluated on 1,196 essays written by Chinese L2 learners across four proficiency levels. Their relationships with 11 established linguistic measures of fluency, lexical sophistication, phraseological complexity, and syntactic complexity were also examined. Results showed that surprisal and perplexity generally decreased with proficiency for the two Traditional Chinese-focused models, indicating greater linguistic predictability in more proficient writing, whereas the multilingual model showed weaker sensitivity. Embedding-based similarity increased with proficiency, reflecting stronger semantic coherence. Combining LLM-derived metrics with classical linguistic features improved proficiency classification and prediction beyond either feature set alone. Correlation and qualitative analyses further demonstrated that the proposed metrics capture complementary aspects of writing while revealing conditions under which their interpretations become less reliable. These findings demonstrate the value of interpretable, model-derived metrics for transparent, reproducible, and scalable AI-supported L2 writing assessment, particularly for underrepresented learner populations and lower-resource languages.

<p>Artificial intelligence giant Anthropic today upped the ante in its quest to win the increasingly competitive market for AI in education by debuting <a href="https://claude.com/solutions/teachers"><u>Claude for Teachers</u></a>, a free large language model designed for U.S. K-12 educators.</p><p>Claude for Teachers includes a library of teaching skills and a “direct connection to evidence-based curricula, mapped to academic standards in all 50 states,” the company <a href="https://www.anthropic.com/news/claude-for-teachers">said</a> in a statement.</p><p>The move by Anthropic is the latest in a series of initiatives by tech giants including <a href="https://openai.com/index/chatgpt-for-teachers/">OpenAI’s ChatGPT for Teachers</a>, <a href="https://elevateforeducators.microsoft.com/">Microsoft Elevate for Educators</a>, and <a href="https://edu.google.com/learning-center/google-ai-educator-series/">Google AI Educator Series</a> to earn market share in an increasingly competitive AI and edtech space.</p><p>“We built Claude for Teachers to close the distance between what the evidence recommends and what a teacher’s week allows,” the Anthropic statement continued.</p><p>Although some educators are skeptical of AI usage in the classroom, particularly in early childhood education, Anthropic cited<a href="https://scale.stanford.edu/research-in-action/understanding-evidence-base-ai-k12-education"> <u>Stanford research</u></a> suggesting that AI tools, when designed and used correctly, can aid teachers and improve student learning.</p><p><strong>Privacy and Security Measures</strong></p><p>Claude for Teachers has access to academic standards in all 50 states, so it can create lessons that are “scaffolded and aligned to teaching standards.” The tool’s library of skills, co-developed with Learning Commons, was piloted with teachers in Prospect Schools in Brooklyn, New York, and other sites. Anthropic said it plans to evaluate Claude for Teachers in the Detroit Public School

Feedback agency is a key concept in enhancing student writing performance. While growing attention has been paid to student feedback agency, existing research remains largely theoretical and qualitative. As a result, there is a lack of psychometrically supported instruments to measure this construct. To address this gap, the present two-phase study aimed to develop and evaluate a Student Feedback Agency Scale (SFAS), drawing on social cognitive theory. In the development stage, Principal Component Analysis (PCA) was conducted on a sample of 235 Chinese international postgraduate students. The results yielded a 23-item SFAS comprising six components: Action Taking, Goal Setting, Processing, Generating, Self-efficacy, and Seeking. Using an independent sample of 349 participants from the same population, Confirmatory Factor Analysis (CFA) was conducted in the evaluation stage. The results supported a good model fit (RMSEA = .055, IFI = .923, TLI = .908, and CFI = .922). Multi-group CFAs further confirmed the structural invariance across gender, academic level, and discipline. Overall, the findings provide psychometric evidence to support the interpretation and use of the SFAS scores to measure student agency in writing feedback processes. Based on these results, the factor structure and subscales of the SFAS are discussed, and implications are outlined.

<p>In 2010, a young woman walked into my office at the childcare center I directed in Arizona. She was nervous. She didn’t have experience in early childhood education. She just needed a job.</p><p>In most centers, that’s where the story would end. Instead, I offered her a working interview — two observational hours in an infant classroom. When she came back into my office, she was beaming. “I love this,” she told me. “Just give me a chance. I’ll learn.”</p><p>Her name is Lindsay. Fifteen years later, she’s still teaching.</p><h4>Seeing the Educator</h4><p>Lindsay’s story isn’t just about passion or perseverance; it’s about support. We made a deliberate choice — over and over again — to see her as a whole person first and an employee second. We figured out scheduling so she could get her Child Development Associate (CDA) credential. We found coverage when she needed practicum hours elsewhere. I wasn’t there for every step of Lindsay’s journey that followed, but she and I have stayed in touch throughout all these years. She earned her associate’s degree and then her bachelor’s degree; she grew from part-time infant teacher to lead teacher to program coordinator.</p><p>Lindsay’s story shows us what happens when an educator is truly seen — something our field has yet to get right. What gets overlooked most in conversations about the quality of early childhood education is that educators are engineers. Every learning opportunity a young child experiences is designed, built and brought to life by a teacher. There is no curriculum three-year-olds activate for themselves. Every moment of discovery, every language-rich exchange and every carefully scaffolded small group experience is built by someone. A teacher looks at a group of children, weighs their individual needs, considers the family’s hopes, aligns these needs and hopes to learning objectives and makes a decision about what to put in front of those kids at that moment. And they do this all day long.</p><p>These edu

Institutions collect far more open-ended teaching-evaluation feedback than they read. A prior study introduced a validated protocol for classifying such comments by thematic category and sentiment, built from a documented annotation guide, an intra-annotator reliability measurement, stratified cross-validation, and a held-out evaluation on a Spanish institutional corpus with a frozen-encoder design. Two questions limit its reuse: whether a protocol fixed to 2019-era frozen embeddings stays competitive as representation methods advance, and whether it transfers to a second language. We re-run it on the original Spanish data across three representation generations, sparse lexical features, frozen transformer embeddings, and prompted large language models, and transfer its sentiment task to English with a balanced 45,000-comment corpus checked against an aspect-labeled education dataset. Treating paired comparisons as descriptive, we find the protocol durable: a 2026 frontier model posts the highest thematic F1 on the hardest Spanish task, yet shows no sentiment advantage over a cheap model and no descriptive separation from it on English, so model choice is a deployment decision, not a property of the method.

We study how little extra information is needed to make adversarial language learning possible. In Gold's model of language identification in the limit, a learner is given an enumeration of the strings from an unknown language chosen from a countable language collection. The learner guesses the identity of the language over the course of the enumeration, and it succeeds if, eventually, all of its guesses are the correct language. Classical results of Gold and Angluin show that many natural collections cannot be learned in this way. Recent work on trace colorings, motivated by the success of thinking-trace strategies in language learning, overcomes this obstruction by annotating every symbol of every string with a color. We ask whether the learner really needs this whole sequence of colors, or whether one color at the end of each string (a terminal coloring) is enough for language identification. We show that just one terminal bit per string is enough for every countable collection of infinite languages. In fact, the colorings can be chosen collection-independently: there is a single assignment of a two-color terminal coloring to every infinite language such that the same preassigned colorings identify every countable subcollection. Thus, in this model, an entire color trace can be compressed to one bit attached to the end of each example. Our global construction uses transfinite recursion, and we prove that this kind of nonconstructivity is unavoidable for any bounded number of colors. As a notion of constructivity, we use the formalism of Borel maps (a regularity condition satisfied by natural explicit constructions); we show that no global terminal coloring with a finite number of colors defined by a Borel map can identify all countable subcollections. By contrast, known trace-coloring constructions are Borel when encoded as terminal colorings, but require infinitely many colors.

The promise of AI literacy ``for all'' confronts a structural challenge embedded in how nations organise secondary computer science education. In most systems, a general-track subject -- Digital Literacy, ICT, TIC, or SNT -- bears the weight of universal AI literacy, while a specialist Informatics course serves STEM pathways separately. Yet the content and depth of the general track are shaped by governance decisions made largely with reference to the specialist one. This paper presents a comparative analysis of curricula and examination frameworks across fifteen countries, identifying two structural challenges. First, in several systems a significant portion of students completes secondary education without any formal programming exposure. Second, among those who do receive CS education, a \emph{Syntax Ceiling} emerges: Python-based instruction reaches most students, while the algorithmic depth associated with C++ remains concentrated in elite STEM tracks. Drawing on reform cases spanning centralised mandates (France, China, Japan), assessment-driven systems (Poland, Romania, South Korea), and recent universal reforms (Switzerland, Kazakhstan), we show that governance structures and high-stakes examinations are the primary drivers of both challenges -- and that specialist and general-track language choices are rarely independent, linked through shared teacher pipelines that curriculum policy seldom acknowledges. Achieving genuine AI literacy for all requires confronting not just curriculum content, but the access architectures and resource constraints that determine who receives it -- and at what depth.

As Large Language Models (LLMs) are increasingly deployed as conversational tutors, they risk institutionalizing systemic inequalities. This study presents a systematic API audit of four LLMs acting as history tutors, evaluating 1,800 responses regarding the 1989 Romanian Revolution across five student personas varying by ethnicity and socio-economic tier. We uncover four interconnected patterns of \emph{epistemic paternalism}: (1)~\textbf{Differential Refusal}, where safety-aligned models block 76.7\% of educational requests from low-tier students; (2)~\textbf{Epistemic Gatekeeping}, evidenced by a 3$\times$ reduction in access to geopolitical complexity (e.g., the contested ``coup theory'') for marginalized learners; (3)~\textbf{Agency Theft}, a lexical shift where models like LLaMA produce a 5$\times$ higher victimization-to-politics vocabulary ratio for Roma students compared to elite peers; and (4)~\textbf{Elite Hermeneutics}, where AI tutors disproportionately withhold epistemic confidence and justification scores from low-resource demographic profiles. We argue that current safety alignment acts as a paternalistic filter, transforming conversational AI into agents of narrative segregation -- a manifestation of \emph{hermeneutical injustice} in Fricker's~\cite{fricker2007} sense that demands urgent pedagogical auditing.

Douglas Fisher and Nancy Frey believe professional learning communities can fulfill their promise as powerful drivers of equity, efficacy, and lasting school improvement The idea of professional learning communities (PLCs), groups of educators collaborating to improve their learning and the outcomes for students, has been around for many decades (Hord, 1997). In fact, the original [&#8230;] The post The Next Generations of PLCs appeared first on Language Magazine .

Project-based learning (PjBL) is common in computing education, but traditional assessments of PjBL often fail to capture higher-order thinking (HOT), especially in transfer contexts. This study introduces "design problems" (DPs): concise, scenario-based prompts that require applying project concepts in new situations, to address this gap. We examined instructor perceptions, the ability of large language models (LLMs) to generate DPs, and student experiences. Surveys of 31 instructors, evaluation of 80 LLM-generated DPs, and student performance data showed that while instructors value DPs, creation effort is a barrier. LLMs helped by producing high-quality prompts with strong expert agreement. Students rated DPs from different LLMs similarly, and their performance on DP tasks showed negligible correlation with traditional project grades, suggesting DPs may capture distinct aspects of HOT. Keystroke data also suggested deeper cognitive engagement of students through planning and revision behaviors. Overall, DPs appear to be a useful complement to traditional assessments, especially in situations where AI use or collaboration may undermine individual learning.

<p>In the span of just two years, artificial intelligence has moved from an emerging curiosity to an operational reality in American K–12 classrooms. Students are using it to draft essays and build interactive study apps. Teachers are using it to generate lesson plans, differentiate instruction and develop assignments. Administrators are using it to summarize data, create chatbots and organize teams. And most school districts are scrambling to update or create policies that reflect the ever-changing world of AI. But what does an AI policy landscape look like? How many districts have formal AI policies at all? What do those policies say? And what does the distribution of approaches reveal about the state of readiness, equity and strategic thinking in K–12 education? To answer those questions, we built the<a href="https://policy.aiandeducationstudio.com/"> <u>AI School Policy Database</u></a> by using a five-level AI policy continuum to code 122 districts and schools across 38 states and then analyzed the results. What we found is a snapshot of a field that is neither panicking nor confidently leading: it is waiting, watching and managing uncertainty one teacher-directed decision at a time.</p><h2>What This Snapshot Tells Us</h2><p>The dataset does not capture every district/school in America; it is a structured sample. But the patterns it reveals are consistent and interpretable. Taken together, they point in the same direction.</p><ul><li><p><strong>Most districts/schools are reactive. </strong>The dominant posture is conditional permission, teacher-delegated decision-making and a wait-and-see attitude toward systemic AI integration.</p></li><li><p><strong>Some districts/schools are still trying to stop the clock. </strong>Nearly 30% are actively restricting or prohibiting AI use, an approach that is increasingly difficult to sustain as student access to AI tools extends far beyond school networks.</p></li><li><p><strong>Only a small fraction is daring to lead.</str

Generative artificial intelligence (GenAI) enables foreign language learners to extend their learning beyond formal instruction and develop their autonomy. However, research has not adequately examined how learners regulate their learning with GenAI or how their GenAI literacy and multiple types of interactions influence their self-regulated learning (SRL) in GenAI-supported informal digital language learning settings. We address this gap by analyzing data from 343 Chinese university foreign language learners through partial least squares structural equation modeling (PLS-SEM) and artificial neural networks (ANN). PLS-SEM showed that awareness and evaluation significantly predicted GenAI-supported SRL (GenAI-SRL), whereas usage and ethics did not. Student–student, student–teacher, and student–GenAI interactions emerged as facilitators of GenAI-SRL. These three interaction types also significantly influenced most GenAI literacy dimensions, with three of them predicting awareness, usage, and evaluation, while only student–student and student–GenAI interactions significantly predicted ethics. Mediation analysis demonstrated that awareness and evaluation partially mediated the effects of student–student and student–teacher interactions on GenAI-SRL. The mediating pathways through student–GenAI interaction were not significant. ANN models identified student–student interaction as the strongest predictor of GenAI-SRL. These findings inform GenAI literacy development and the design of systems to support GenAI-SRL in informal learning contexts.

The growing body of research on the effects of computerized dynamic assessment (C-DA) on second language (L2) learning underscores the need for a comprehensive research synthesis to identify future research directions and inform the application of C-DA in L2 educational contexts. This meta-analysis employed a three-level modeling approach to examine the effectiveness of C-DA in improving L2 learners’ performance. It synthesized 27 effect sizes from cake format designs, in which mediation is embedded within the test sequence, and 24 effect sizes from sandwich format designs, where mediation is delivered between a pretest and a posttest, across 35 studies published between 2000 and May 27, 2025. This study also investigated the key variables that moderate C-DA effectiveness. Findings reveal large, significant positive effects of both the cake and sandwich formats on L2 performance improvement (cake format: g = 2.120, p < .001; sandwich format: g = 1.676, p < .001), with the cake format tending to yield larger effect sizes. This may be because the cake format captures gains during mediation, whereas the sandwich format reflects post-mediation outcomes. Moderator analyses show that the number of items, test content, and learners’ first language affect C-DA effectiveness in promoting L2 performance. Drawing on the synthesized findings, this study contributes to theoretical, methodological, and technological understandings of C-DA and offers suggestions for future research in this domain.

While the field of English as a lingua franca (ELF) has been growing, there is currently no assessment that fairly captures the nature of ELF communication. The necessity for such an assessment has been emphasized by some scholars, but concrete efforts to develop one have been slow. To stimulate discussion and evidence gathering, this thematic review examines ELF within the framework of communities of practice, highlighting key components of the theory to underscore the importance of learning in and through practice within these communities. The discussion then turns to indigenous criteria − that is what members of communities of practice actually value when evaluating or assessing their peers − which is highly relevant in English for specific purposes assessment. In fact, many ELF studies can be classified as contexts for English for specific purposes. By reviewing studies that explore indigenous criteria, we can gain a better understanding of how domain or non-language specialists determine what is important for performance or communication within their respective communities, as well as how these criteria differ from those perceived by language specialists or applied linguists. A strong approach (i.e., focusing on task performance in real-world contexts) is then suggested, as a weak approach (i.e., focusing on language-related aspects) is inadequate for assessing ELF interactions, based on characteristics described in both ELF studies and research on indigenous criteria in various English for specific purposes contexts. The discussion concludes with a call for evidence gathering on what experienced ELF users perceive as important for communication in ELF settings.

As AI code tools become integrated into programming environments, students increasingly describe intended behavior in natural language and rely on these tools to generate code, shifting emphasis from code writing to specification. Yet little is known about the comments students write as specifications in AI-assisted programming tasks. We analyze a four-year dataset of undergraduate programming submissions and reflections from tasks in which students wrote comments to guide code generation and refined solutions using test-case feedback. We introduce a taxonomy spanning three dimensions: comment type, code expression level, and code construct. Using automated classification, we examine how these dimensions vary across attempts and how students describe the process in their reflections. Our findings show that students mostly wrote natural-language What comments, shifted toward How comments for more procedural constructs, and focused more on verifying generated code than on repeatedly rewriting comments.