Generative Artificial Intelligence (GenAI) is reshaping education by introducing tools that enhance teaching methodologies, personalize learning, and streamline administrative tasks. However, adoption of these tools remains uneven, raising concerns about disparities in AI literacy and competency across geographic regions and educational contexts. Through the lens of Innovation Diffusion Theory, we investigated the adoption of GenAI tools among second language (L2) teachers in the United States, Colombia, Germany, and Macau. Using survey data, we assessed four areas: accessibility of GenAI tools, teacher knowledge of potential applications, integration in teaching practices, and the nature of professional development provided. Our results indicated substantial intra- and inter-context variance, with U.S. and Colombian educators reporting higher familiarity and usage compared to those from Germany and Macau. Additionally, university and high school teachers were more likely to access professional development and leverage GenAI for tasks like assessment and differentiation than elementary or middle school educators, regardless of geographic setting. These disparities align with broader trends in AI adoption, reflecting heterogeneity in cultural attitudes, systemic barriers, and institutional support. These findings contribute to Innovation Diffusion Theory by illustrating how accessibility, knowledge, and institutional support interact across diverse contexts. They offer practical guidance for policymakers and educational leaders to develop targeted interventions that promote equitable GenAI integration in language education worldwide. • We find an AI divide among L2 teachers associated with training and support gaps. • Disparities highlight emerging geographic divides in AI literacy and competency. • Government policies and institutional priorities influence GenAI adoption. • Targeted strategies are critical to mitigate gaps in adoption for L2 education.

For learners who find reading easier than listening and struggle to keep up with the speed of native speakers, listening practice that emphasizes linking sounds to written words and adapting to speech rate is essential. This study investigates the effects of deferred multimodal presentation , an input method that temporally separates reading and listening. Unlike simultaneous input (e.g., reading-while-listening), this sequential approach reduces cognitive load by preventing split attention and facilitates auditory decoding by leveraging prior information from another modality. Ninety-nine Japanese EFL learners, including a control group, engaged in five days of structured practice under one of four conditions: (1) normal-rate audio only, (2) fast-rate audio only, (3) transcript reading followed by normal-rate audio, or (4) transcript reading followed by fast-rate audio. Posttest results after the five-day practice sessions showed that transcript-supported listening—particularly when followed by fast-rate audio—led to greater improvements in both listening proficiency and adaptation to fast speech compared to audio-only conditions. These findings suggest that deferred multimodal input enhances predictive coding—the brain's mechanism for inferring incoming information based on prior knowledge—and supports the consolidation of meaning retrieval. Through such practice, learners can form sound–meaning integrated representations, ultimately improving their listening proficiency. • Deferred multimodal input promotes retrieval and sound–meaning mapping during listening. • Prior knowledge makes challenging auditory input more learnable and predictable. • Pre-listening transcript reading supports top-down processing in L2 listening. • Fast-rate listening with transcript reading improves L2 listening and processing speed. • Predictive coding helps store accurate, meaning-linked speech representations for better L2 listening.

Although second language (L2) speaking assessment usually seeks to cover interactional aspects, in practice, it is often limited to assessing L2 learners' individual audio-recorded performance or focusing merely on linguistic criteria. This study experiments with a new method for assessing interactional competence. The computer-joystick method allows observing both verbal and nonverbal features while rating an individual's behavior from video. The present study seeks to find out whether the method can be used to assess speakers' interpersonal behaviors reliably in the context of L2 learners' dialogue speaking tasks. The interpersonal behavior of academic L2 Finnish learners (N = 44) was examined during three spoken interaction tasks designed for A2–B2 level speakers. The study is part of a larger project which develops ways to measure and assess oral language skills automatically. Each learner's interpersonal behaviors (using two variables of dominance and affiliation, rating both linguistic and nonverbal features) were continually coded over time while watching audiovisual recordings of their speaking performance. The reliability of the ratings ranged from moderate to good. The resulting time series data were used to analyze changes in the variables and their variation according to task-level, dyad-level and participant-level predictors. Nonindependence metrics, actor and partner effects of language proficiency, and measures of synchronization between participants were investigated. In sum, it was found that the coding system can be used reliably for assessing interpersonal behaviors in the context of language assessment. Finally, potential implications for assessing spoken interaction in a second language using both humans and algorithms are discussed. • Joystick coding was used to rate interpersonal behaviors from video recordings. • Video recordings enable assessment of nonverbal and verbal features. • We analyzed language learners' dominance and affiliation in dialogue performances. • Time series data enables changes in the variables and their interplay to be analyzed. • The findings can be used to assess spoken interaction between humans and algorithms.