Program & Abstracts

8-9am | [Panel] From Fiction to Function: How Imagined Characters Prepare Language Professionals and Learners for AI Collaboration

Alla Shubina, Norayr Mnatsakanyan, & Oxana Sokolova (TSLL26-91)

1) Research Exposure to imagined characters—whether in literature, film, or interactive media—has long cultivated the cognitive and interpretive skills central to language professionals’ engagement with emerging technologies, including AI agents. This paper argues that the analytical habits developed through interpreting fictional personas—such as inference of intent, voice attribution, narrative framing, and pragmatic meaning—directly inform how researchers interact with AI systems. Much like literary characters, AI agents present constructed voices that must be evaluated for coherence, reliability, and contextual appropriateness. Language professionals trained in discourse analysis and literary interpretation are particularly equipped to interrogate these outputs, identifying underlying patterns, biases, and rhetorical strategies. Furthermore, familiarity with polyphony and narrative multiplicity in fiction prepares researchers to conceptualize AI as a composite of voices rather than a singular authority. This perspective encourages critical engagement, methodological reflexivity, and ethical awareness in AI-assisted research, where distinguishing between generated and human-authored knowledge is essential.

(2) Teaching In pedagogical contexts, imagined characters serve as a bridge between traditional language instruction and AI-mediated communication. Teachers have long used fictional dialogues, role-play, and narrative simulations to develop students’ communicative competence. These practices parallel interactions with AI agents, which often require users to adopt roles, frame prompts, and anticipate responses. This paper suggests that educators who integrate character-based scenarios into their teaching are implicitly preparing students for effective AI interaction. By treating AI as a conversational partner with a constructed persona, instructors can guide learners in developing pragmatic awareness, register control, and adaptive communication strategies. Additionally, the use of imagined characters fosters metalinguistic reflection, enabling students to analyze how meaning is shaped by context, intention, and perspective—skills that are crucial when evaluating AI-generated language. Thus, pedagogical traditions grounded in narrative and character not only remain relevant but gain renewed significance in the age of AI-enhanced education.

(3) Learning For language learners, engagement with imagined characters nurtures essential competencies such as empathy, perspective-taking, and interpretive flexibility. These competencies translate directly into effective interaction with AI agents, which often simulate conversational partners across diverse contexts. Learners accustomed to navigating fictional worlds are better prepared to suspend disbelief, engage meaningfully with non-human interlocutors, and adapt to varying communicative cues. Moreover, imagined characters provide a safe space for experimentation, allowing learners to practice language use without the social risks associated with real-life interaction. This parallels the low-stakes environment of AI-mediated learning, where users can refine their linguistic output through iterative dialogue. The paper argues that such experiences foster learner autonomy and confidence, as well as critical digital literacy. Ultimately, exposure to imagined characters equips learners with the cognitive and affective tools necessary to collaborate with AI agents, transforming them from passive recipients of information into active participants in knowledge construction. By examining research, teaching, and learning through the lens of imagined character engagement, this paper highlights a continuity between pre-digital narrative practices and contemporary AI interaction. It contends that the skills developed through encounters with fictional personas are not ancillary but foundational to effective and critical collaboration with AI in language-related domains.

8-8:30am | From Cognitive Scaffold to the Illusion of Competence: A Systematic Review of Generative AI in Chinese as a Second Language Writing

Yuxin Tian (TSLL26-68)

Generative Artificial Intelligence (GenAI) is rapidly reshaping Computer-Assisted Language Learning (CALL), yet its role in Chinese as a Second Language (CSL) writing remains under-synthesized and theoretically uneven. To address this issue, this study conducted a PRISMA-guided systematic review and MMAT-based quality appraisal of 14 peer-reviewed empirical studies on GenAI in CSL writing. The review identified four recurring patterns. First, GenAI generally improved lexical sophistication, grammatical accuracy, and overall text quality, but these gains were often constrained by over-correction, over-formalization, and erosion of authorial voice. Second, learner engagement with GenAI was strongly mediated by proficiency. More advanced learners tended to use it strategically and selectively, whereas lower-proficiency learners were more likely to rely on it superficially. Third, the evidence supported a complementary rather than replacement relationship between AI and teachers. GenAI was effective in providing immediate, rule-based feedback, while human instructors remained more important for pragmatics, rhetorical judgment, affective support, and the protection of learner voice. Fourth, GenAI often reduced writing anxiety and increased confidence, but for novice learners these benefits could develop into an illusion of competence. Building on these converging findings, this review proposes a provisional Dual-Scaffold Synergistic Model, in which GenAI provides linguistic and cognitive support, while teacher mediation and learner agency regulate meaning, voice, and critical uptake. Overall, GenAI appears most beneficial in CSL writing when it is integrated as a guided support tool rather than used as a substitute for human pedagogy.

8-8:30am | Between Myth and Machine: Iranian EFL Teachers’ Metaphorical Narratives of Artificial Intelligence through the Lens of Human–AI Interaction Ethics

Behrokh Abdoli & Reza Asgari (TSLL26-52)

Artificial intelligence is rapidly transforming language education and reshaping how teachers interact with technological agents in teaching and learning environments. At the same time, perceptions of AI are often shaped by cultural narratives and symbolic representations that influence how new technologies are understood. This study investigates how Iranian EFL teachers conceptualize artificial intelligence through culturally grounded metaphors informed by Iranian myths and cultural narratives and how these perceptions reveal ethical concerns about human–AI interaction in educational contexts. Adopting a qualitative narrative inquiry design, the study explores teachers’ experiences and reflections regarding AI in language education. Data were collected through think aloud protocols, narrative frames, and semi structured interviews with 34 EFL teachers selected through purposive sampling, all of whom had prior experience integrating AI tools into their teaching practices. The narratives were analyzed using deductive thematic analysis to identify recurring patterns in the metaphorical representations teachers used to describe AI. MAXQDA software was employed to support the coding process and generate a code map illustrating relationships among emerging themes. To interpret the ethical implications of these narratives, the study explores EFL teachers interaction with AI (Shneiderman’s, 2020) through the lens of Floridi et al.’s (2018) AI ethics framework. Interestingly, EFL teachers did not generally perceive AI as an entirely helpful educational tool; rather, their accounts often reflected caution, skepticism, and ethical concern. Findings from the code map generated by MAXQDA revealed three major themes. The first theme, preserving human agency through transparent AI, reflected concerns about loss of human control and lack of transparency, metaphorically associated with figures such as Zahhak and Akvan, who symbolize domination and deception. The second theme, ensuring safe and trustworthy interaction with AI, captured fears about potential risks and harmful consequences of powerful technologies, evoking mythological creatures such as the Ezadaha and Gandarewa that represent destructive and uncontrollable forces. The third theme, promoting human well being through beneficial and fair AI systems, highlighted anxieties about injustice and misuse of power in AI mediated environments, resonating with figures such as Afrasiab and Gorgsār, traditionally linked to tyranny and social disorder. The study bears implications for language teacher education, AI supported pedagogy, and the ethical integration of emerging technologies in educational contexts by emphasizing the need to address teachers’ cultural perceptions and ethical concerns when designing and implementing AI tools for language learning.

8:30-9am | Constructed Agents, Real Interaction: Evaluating AI for Pragmatics and Interactional Competence in Second Language Learning

Sana Bint E Javaid & Mehak Jawed (TSLL26-111)

Language-using AI agents are thus not only positioned as feedback instructors, but also as interactive conversational partners with abilities to approximate dialogues in textual, spoken, and multimodal interaction. Yet, one important pedagogical question remains largely unaddressed: to what extent can these agents be used to foster interactional competence and pragmatic ability beyond linguistic accuracy? This paper takes a critical look at the expanding role of AI agents in AI-mediated language learning and argues that the educational value of AI agents should be evaluated against authentic communication, rather than just efficiency or personalization. Drawing on applied linguistics, pragmatics, and educational technology scholarship, the paper develops a conceptual framework for evaluating AI agents along six dimensions: turn-taking, context sensitivity, awareness of register, feedback, repair, and sustained dialogue. It also provides key concerns about assessment validity, sociolinguistic simplification, overdependence of learners, and unequal technological access. The paper argues that AI agents should be construed as communicative actors that have an effect on the negotiation of meaning, management of the interaction, and context-based interpretations of language by the learners. By shifting its focus from correction to interaction when assessing AI, the study adds to the scholarly discussions regarding the pedagogical, ethical, and communicative effects of employing artificial agents in second language studies.

8:30-9am | Workplace Futures in Motion: Mapping the Emotional Geographies of EFL Student Teachers’ AI Transitions

Reza Asgari & Behrokh Abdoli (TSLL26-56)

Artificial intelligence is reshaping language education and how future teachers prepare for technology‑mediated instruction. However, different training environments shape how student teachers emotionally experience and professionally navigate AI preparation. This study examines the emotional geographies of AI preparation among EFL student teachers trained in formal and non-formal settings, exploring how these experiences shape their horizontal and vertical professional growth within their AI-mediated workplace. Adopting a qualitative narrative inquiry design, the study explores EFL student teachers’ experiences and reflections regarding AI preparation during their training and early teaching practice. Data were collected through narratives and semi‑structured interviews with 19 purposively selected EFL student teachers across both training contexts. The data were analyzed using deductive thematic analysis informed by Hargreaves’ (2001) concept of emotional geographies of teaching and the framework of vertical and horizontal professional transitions (Nicholson & West, 1988). MAXQDA 24 software was used to support the coding process and visualize relationships among emerging themes. The analysis revealed four themes reflecting distinct developmental trajectories across the two training contexts. Operational AI Advancement shows how student teachers trained in private institutes experienced AI preparation as a practical resource that supported instructional competence and facilitated professional advancement in AI‑mediated teaching environments. Human‑Centered AI Adaptability reflects how the practice‑oriented training in private institutes enabled student teachers to use AI flexibly and ethically in real classroom situations, helping them balance technological demands while maintaining meaningful teacher–student relationships. In contrast, Institutionalized AI Knowledge indicates that university‑based student teachers encountered AI primarily through theoretical and policy‑oriented instruction, positioning AI more as academic knowledge than as a practical pedagogical tool. Practical Adaptation Strain captures the emotional and professional pressure these student teachers experienced when applying AI in teaching practice, as they were required to manage AI‑related responsibilities with limited hands‑on preparation. This study bears implications for teacher education curricula, policy, and professional development, indicating the need for more emotionally responsive, practice‑based AI preparation that is attentive to teacher–student relationships, and that bridges institutional theory and classroom reality across diverse training contexts.

9-9:30am | The Algorithmic Djinn: Displacing Frankenstein to Understand AI Sentience in the Maghreb

Yassine Ismaili (TSLL26-126)

Current academic debates about artificial intelligence rely heavily on Western cultural narratives. When scholars discuss non-human agents, they often point to figures like Frankenstein or Hal from 2001. These stories frame the AI as a physical creation that inevitably rebels against its human maker. This specific anxiety, however, does not easily translate to students in North Africa. The current presentation offers a different cultural lens for understanding AI in the Maghreb. Rather than a rebellious machine, we can view the AI through the traditional concept of the Jinn. In local oral traditions, the ‘Jinn’ is a powerful and unseen entity. It exists in a hidden space and must be engaged through highly specific language. The Jinn is not a tool to be programmed, but a capricious agent that requires careful negotiation. It can be helpful, but it is also prone to deception. This traditional framework aligns remarkably well with how users actually interact with Large Language Models today. Prompt engineering functions much like a modern incantation. When we displace the Frankenstein myth with the Algorithmic Djinn, we provide Moroccan students with a familiar concept for critically evaluating AI. This perspective helps educators move past Western fears of rogue technology. It allows us to teach the critical and linguistic skills needed to negotiate with unpredictable digital agents.

9-9:30am | From Chatbot Dialogue to Professional Readiness: Field Shifts in Learner Interaction with an AI Agent

Daniel Portman & Hannah Landes (TSLL26-37)

As AI agents take on increasingly complex roles in language-learning tasks, learners must do more than produce correct language: they must also sustain focused, goal-oriented interaction over extended exchanges. Yet we still know relatively little about how well they do so. This study analyzes transcripts of engineering students interacting with a Poe-based chatbot that performed the role of an investor evaluating their innovation proposals. The task formed part of a CEFR B2 English for Purposes of International Communication course in which students prepared for live assessed investor-style simulations requiring clarity, relevance, and professional communicative competence. Using the systemic functional concept of field, understood as what the conversation is about, the study examines field shifts: moments when learners drift from the central topic in ways that weaken coherence and communicative effectiveness. Through qualitative and quantitative analysis (35 participants produced 503 speaking turns), the study identified several recurring types of field shifts, especially incomplete responses (“Partial Shot”), incorrect or misaligned responses (“Flub Shot”), and overly general responses (“Zooming Out”), as well as cluster profiles of co-occurring field shifts. Their recurrence across the dataset suggests that loss of topical focus was not incidental, but a meaningful interactional challenge in learner-chatbot dialogue. These patterns often resulted in missed communicative opportunities and reduced students’ ability to convince the simulated investor. The analysis also considers possible underlying communicative challenges associated with these shifts and outlines pedagogical interventions to address them. In doing so, the study shows how learner-chatbot dialogue can be used both to identify communicative difficulty and to inform pedagogy for more effective professional interaction.

9-9:30am | AI Tools as Autonomy-Support Resources in EFL Contexts: Do EFL Learners Who Use AI Report Higher Self-Directed Learning Behaviour?

Abdelmalek Ismaili Alaoui (TSLL26-125)

The integration of AI-powered tools such as ChatGPT, Grammarly, and AI-based writing assistants into everyday life has opened new possibilities for informal, self-directed language practice among EFL learners. Despite growing scholarly interest in both learner autonomy and AI-assisted language learning, the relationship between AI tool use and self-directed learning behaviour remains empirically underexplored. This study addresses that gap by investigating whether EFL learners who regularly use AI tools for out-of-class English practice report higher levels of self-directed language learning behaviour than those who do not. Drawing on Self-Determination Theory (Deci & Ryan, 1985) and established frameworks of learner autonomy in SLA (Benson, 2011; Holec, 1981), the study adopts a quantitative cross-sectional survey design. Data are collected from adult EFL learners across diverse learning contexts through an online questionnaire comprising three components: a learner background profile, an AI tool use scale measuring frequency and purpose of use, and an adapted version of the Self-Directed Language Learning Scale (Song & Hill, 2007). Descriptive statistics, Pearson correlation, and multiple regression analysis are used to examine associations between AI tool use and self-directed learning behaviour, controlling for proficiency level and learning context. Preliminary findings suggest a positive association between frequent AI use — particularly for open-ended tasks such as conversation practice and free writing — and higher reported autonomy. Learners who engage with AI for productive, generative tasks appear to display more goal-setting, self-monitoring, and reflective behaviour than those who rely on AI exclusively for closed, corrective tasks such as translation or grammar checking. The paper discusses implications for language pedagogy, arguing that how learners use AI matters as much as whether they use it, and calls for greater pedagogical attention to fostering intentional, autonomous AI use in EFL contexts.

10-10:30am | Stitching the Digital Guide: Modern Frankensteinian Motifs in Student-AI Co-Creation of Cultural Tours

Wes Harven Guillemer Maravilla (TSLL26-50)

The creature in Mary Shelley’s Frankenstein illustrates the dilemma of ownership of a created life in the rawest form possible. This study attempts to address this question in an AI-assisted language classroom: What ownership do EFL learners in Taiwan have over a digital persona they created to represent their city? Sophomore students enrolled in an English-Medium Instruction course at Tamkang University worked with AI Gemini to create a 360-degree virtual reality tour of Tamsui. Instead of having AI operate in an invisible and seamless manner, students were instructed to have AI portray a character of a working dock, street food vendor, or heritage storyteller who would engage in English narration about Tamsui to the tourist audience. Considering the wider literature and cinema traditions that employ sentient, the study interprets the AI Tour Guide as a contemporary example of a “constructed agent” whose body, voice, and personality are constructed from students’ selections. Animating the tour guide and post-reflective surveys helped examine students’ relational shifts regarding authorship, voice, and cultural representation. Many participants noted an increase in willingness to represent to the public/cultural as an ambassador, while a few participants noted discomfort regarding the perceived fluent and authoritative voice of the creation. Finally, it posits that the use of AI in a foreign language learning context can be perceived as a rich source of potential from a humanities perspective when regarded as an agent or a tool, rather than as a feature/technology of language learning. It explores student engagement with their “Frankensteinian” guide and proposes that the motifs of creation in literature and cinema provide a valuable resource for articulating the reality of social interactions with language-using AI in the workplace of the future, the digital workplace of the 2030s.

10:30-11am | From Shelley’s Creature to ChatGPT: The Evolution of the ‘Constructed Other’ in Human Imagination and Language Learning

Shugyla Kudaibergen (TSLL26-27)

This paper explores the relationship between fictional representations of sentient non-human agents and contemporary human interaction with AI, particularly in the context of language learning. From Mary Shelley’s Frankenstein (1818) to Stanley Kubrick’s 2001: A Space Odyssey, imagined beings have shaped cultural expectations of artificial intelligence. These narratives construct the “Other”—a being both like and unlike humans—raising ethical, psychological, and linguistic questions that resonate today in our engagement with AI agents such as ChatGPT. The study examines how literary and cinematic portrayals of constructed agents influence users’ perceptions of AI in educational settings. Drawing on theories of the uncanny (Freud, 1919) and posthumanism (Haraway, 1991; Hayles, 1999), it analyzes how fictional tropes—fear of the unknown, desire for control, the blurred boundary between human and machine—are reflected in contemporary discourse around AI in language teaching. By prompting learners to reflect on these narratives, educators can foster critical awareness of how AI agents are designed, perceived, and used. The paper also presents preliminary findings from a pilot study in which language learners engaged with ChatGPT and were asked to reflect on their interactions through the lens of literary archetypes. Results suggest that learners who were familiar with fictional AI agents demonstrated more nuanced expectations and ethical reasoning. Ultimately, this paper argues that the humanities—literature and film studies in particular—offer essential frameworks for understanding and shaping human-AI interaction in language learning. By bridging the gap between cultural imagination and technological reality, we can better prepare learners and educators for a future where constructed agents are not just imagined, but interacted with daily.

10:30-11am | From Interaction to Preparation: Exploring AI Linguistic Personhood in Language Education

Khalid Hicham & Mohamed Belmihi (TSLL26-35)

While recent advances in natural language processing have enabled AI agents to display discourse features traditionally associated with human interlocutors, learners often interpret these capacities as indicators of human‑like agency or intentionality. The research was driven by two primary objectives: evaluating Moroccan EFL learners’ pragmatic responsiveness and its development across university academic semesters, and examining the adaptive register strategies and teaching methods with regard to Personhood, and its relationship to Language Education. With this in mind, and while adopting a mixed-methods approach, the research integrates in the sequential explanatory mixed methods research design both cross-sectional and correlational designs to explore AI Linguistic Personhood and its teachability. The study involved 275 undergraduate students majoring in English studies from four Moroccan public universities. Data collection was carried out using questionnaires, focus group discussions, observations, and documents review. The qualitative data were analyzed using content analysis, while quantitative data were analyzed using descriptive statistics and Structural Equation Modeling (SEM). On studying the findings, it was revealed that AI‑mediated communication offers a unique training ground for developing digital literacy, critical language awareness, and communicative adaptability. At the same time, it highlights the need for explicit pedagogical frameworks that help learners to effectively communicate AI‑generated language, recognize its computational nature, and understand the limitations of attributing personhood to non‑sentient systems. By integrating insights from interactional linguistics, discourse analysis, and AI ethics, the paper proposes a model for using AI agents ethically in language education that balances creativity with critical engagement, thereby it is of paramount importance to equip learners with the digital literacy skills required for better communication with an AI‑enhanced future.

10:30-11am | Can Humans or AI Anticipate Latent Discourse Patterns? A Corpus-Based Study of Pop Song Lyrics

Maria Claudia Nunes Delfino & Tony Berber Sardinha (TSLL26-137)

Corpus linguistics has consistently demonstrated that human intuition is an unreliable predictor of corpus-derived patterns (Sinclair, 1991; Biber, 1998), largely due to limited access to large-scale distributional regularities. While this limitation is well established for human cognition, it remains unclear whether it also applies to artificial intelligence systems. Large language models (LLMs), trained on vast textual datasets, encode statistical regularities that may, in principle, enable them to anticipate emergent linguistic patterns beyond the reach of human introspection. This study investigates whether humans and LLMs can predict lexical dimensions, which are the latent discourse patterns identified through factor analysis of co-occurring lexical items (Berber Sardinha & Fitzsimmons-Doolan, 2025). In corpus-assisted discourse studies, such dimensions are often perceived as post hoc self-evident (Partington, 2017), despite rarely being anticipated prior to analysis. As population-level phenomena, they emerge only from covariance patterns across large corpora and are not directly observable in individual texts. Using pop song lyrics as a case study, a domain rich in strong intuitive expectations, we analyze a 4,000-text English-language corpus (approximately 1.2 million words), evenly divided between human-authored and AI-generated lyrics across five genres: country, pop, rap, rock, and soul. Lexical Multidimensional Analysis was employed to extract latent discourse dimensions. These were then compared with expectations elicited from undergraduate students and three LLMs (GPT, Gemini, and Llama), all tasked with predicting the discourse patterns typical of pop lyrics. The findings reveal that neither human participants nor LLMs successfully anticipated the corpus-derived dimensions. This suggests that lexical dimensions function as emergent, population-level discourse patterns that are not recoverable through intuition (human or machine) nor directly accessible from current LLM representations. The study highlights fundamental limits in both human and AI predictive capacities regarding latent linguistic structure, reinforcing the epistemological value of corpus-driven methods.

11-11:30am | From Frankenstein to Chatbots: Humanizing AI Language Agents Through Literary and Linguistic Lenses

Amine Bouhzam (TSLL26-34)

As AI agents become increasingly prevalent in language learning, understanding how humans perceive and interact with these technologies is essential for effective pedagogy. This paper examines the intersection of literary representations of sentient non-human agents and the design and reception of AI language agents. Drawing on iconic figures such as Frankenstein’s creature (Shelley, 1818) and HAL 9000 (Kubrick, 1968), the study explores how exposure to imagined sentient beings shapes learners’ expectations, trust, and engagement with AI-driven language learning platforms. Using a mixed-methods approach, qualitative data from learner reflections and classroom interactions with AI agents were analyzed alongside a linguistic assessment of the agents’ communicative strategies, including politeness, responsiveness, and error feedback (Godwin-Jones, 2021). Findings reveal that learners often anthropomorphize AI agents, attributing human-like motives and emotions, which influences their motivation and interaction patterns. Moreover, learners’ prior cultural exposure to fictional sentient agents informed their perceptions of AI reliability, ethical considerations, and collaborative potential. This study contributes to applied linguistics by highlighting the need for pedagogically informed AI design that accounts for learners’ cognitive and affective responses. It also bridges humanities and language education by demonstrating how literary and cinematic portrayals of sentient agents can inform contemporary AI-mediated learning. Implications include recommendations for AI agent design, strategies for scaffolding learner engagement, and ethical considerations for integrating AI into language classrooms. By foregrounding the humanistic lens alongside linguistic analysis, this paper advances interdisciplinary understanding of AI in education and provides actionable insights for language teachers, curriculum designers, and AI developers.

11-11:30am | Cultural and Linguistic Authenticity in AI-Generated ESL Materials: Local vs. Global Agents in Sri Lanka

Sachin Wanniarachchi, Sayed Karim & Sahan Wanniarachchi (TSLL26-33)

This study investigates the cultural and linguistic authenticity of AI-generated materials for the Sri Lankan ESL classroom by comparing and contrasting outputs from a global AI agent (ChatGPT-4) and locally developed agents (ApilageAI and DialogAI). Three questions guide this research: 1) How do the texts, dialogues, and examples generated by global and local AI agents align with Sri Lankan cultural references, politeness norms, and features of Sri Lankan English? 2) Where do global agents exhibit cultural prejudice or insensitivity in materials intended for Sri Lankan learners? 3) To what extent do local agents address these gaps in cultural and linguistic representation? The study employs a qualitative comparative case design, analyzing a purposive sample of 20 generated texts (10 per agent) produced from identical prompts requesting culturally situated materials for secondary ESL learners. Drawing on Kachru’s World Englishes framework for analyzing linguistic variation and Hofstede’s cultural dimensions theory for examining cultural representation, the methodology combines critical discourse analysis of textual content with thematic analysis of cultural references and pragmatic features. Analysis focuses on three dimensions: lexical choices reflecting Sri Lankan English varieties, inclusion of local cultural schemas (festivals, kinship terms, social hierarchies), and pragmatic appropriateness in dialogues. Results indicate that ChatGPT-4 produces linguistically standardized materials with generic cultural references applicable to multiple South Asian contexts, occasionally flattening distinctive Sri Lankan features, which, in a way, leads to an epistemic violence. Conversely, DialogAI demonstrates stronger alignment with local politeness norms and incorporates culturally specific references, yet reveals inconsistencies in representing Tamil and Muslim minority cultural practices. The findings contribute to evaluating AI fitness-for-purpose by demonstrating that local agents may better serve linguistic authenticity while revealing that neither agent fully addresses Sri Lanka’s multicultural complexity, underscoring the need for critical human oversight in AI-generated materials.

11-11:30am | Investigating the Degree of Resemblance Between AI-Generated Lyrics and Human-Written 80s Rock Lyrics

Benjamin Godard (TSLL26-54)

The release of freely accessible generative Artificial Intelligence (AI) has considerably changed how written language is composed. Researchers have been investigating AI and its ability to generate human-like, register-specific text. However, little attention has been given to the use of generative AI for creating lyrical content, despite the growing accessibility of AI music-generation tools. Platforms such as Udio and Suno exemplify a new generation of AI agents capable of producing creative, language-based outputs. Since its release in December 2023, Suno has developed lyric-generation models that allow users to generate lyrics along with instrumental tracks. Its ReMi model now enables users to generate complete songs from a single prompt. The purpose of this study is to investigate the extent to which AI-generated lyrics resemble human-written lyrics with regard to discourse, lexical, and grammatical features. Song lyrics constitute a hybrid register situated between spoken and written language, conversational in tone yet carefully crafted. This study operationalizes the 1980s rock genre as a hybrid construct blending rhythm and blues, glam, punk, new wave, and pop, and draws on the Billboard Hot 100 Year-End charts to compile a human corpus of 1980s rock song lyrics. Using a corpus-based approach, two specialized corpora are examined: the Human Song Lyrics Corpus (HSLC), which contains the top human-written rock songs (n=293) of the Billboard Hot 100 Year-End charts from 1980 to 1989, and the ReMi Corpus, a parallel and comparative AI corpus of song lyrics (n=293), generated by Suno’s ReMi model. Prompts were carefully crafted to allow Suno’s lyric-generation model to fully demonstrate its creative potential. The analysis focuses on discourse, lexical, and grammatical features to identify the degree of similarity and difference between human and AI-generated lyrics. By examining these patterns, this study contributes to understanding how AI agents reproduce human-like linguistic features.

11:30am-Noon | Using Animated Shorts to Cultivate Emotional Intelligence among Language Learners: A Recommendation for AI-Assisted Framework

Ibrahim Halil Topal (TSLL26-31)

Language learning comprises not only linguistic knowledge but also socio-cognitive abilities. In addition to intelligence quotient (IQ), scholars argue that individuals must possess emotional intelligence (EQ) to cope with personal and educational challenges. Despite the long-standing presence of psychological factors (i.e., affective filter hypothesis) in language education, a renewed interest emerged in EQ – the ability to recognise, understand, manage, and utilise emotions effectively in oneself and others. In this context, short animated films offer rich, contextually relevant input to expose learners to emotional content and reduce the affective filter. Considering that young people, today’s learners, experience brain rot – low-quality, nonsensical internet content that causes cognitive decline, mental fogginess, and reduced attention spans due to overstimulation, providing them with AI-assisted training in EQ becomes urgent. Drawing on Salovey and Mayer’s (1990) four-branch EQ model, this paper offers a framework for using animated shorts with the help of generative AI models. Using this framework, language teachers can foster emotional awareness and intelligence among language learners. The presentation first justifies the study’s significance and provides a step-by-step demonstration of how to implement this framework through an animated short. The framework is expected to offer a practical means for leveraging EQ in language education.

11:30am-Noon | Multimodal AI Agents and L2 Academic Writing Self-Efficacy in Ghanaian Teacher Education: A Covariance-Based SEM Study

Tabka William-Peters (TSLL26-19)

L2 learners often struggle with writing tasks such as synthesizing research, organizing complex arguments, and adhering to disciplinary discourse conventions. Multimodal AI agents, inspired by systems like Google NotebookLM, can scaffold these processes by allowing learners to upload research articles and receive integrated text, audio, video, and visual feedback that supports comprehension, evidence integration, and coherence. This study investigated how perceived ease of use and multimodal AI scaffolding shape L2 academic writing self-efficacy through perceived usefulness and engagement. After completing an inquiry based synthesis task using NotebookLM, a quantitative survey as administered to 250 L2 learners from two colleges of education in Ghana using a 21-item instrument measuring perceived ease of use, multimodal scaffolding, perceived usefulness, engagement, and L2 academic writing self-efficacy. A covariance-based structural equation modelling (CB-SEM) with maximum likelihood showed excellent fit to the data (CFI = 0.989, TLI = 0.988, RMSEA = 0.018, 90% CI [0.000, 0.034], SRMR = 0.061, GFI = 0.966). The measurement model showed good reliability and validity (standardised loadings = .72–.86, composite reliability = .85– .90, AVE = .59–.64). Structurally, perceived ease of use and multimodal AI scaffolding were positively associated with perceived usefulness and engagement, and perceived usefulness (β = .29, p = .001) and engagement (β = .16– .19, p = .01– .02) were positively related to writing self efficacy, whereas direct paths from ease of use and scaffolding to self efficacy were not significant. These findings provide empirical support for multimodal AI agents as effective scaffolds for L2 academic writing, suggesting that integrated multimodal feedback can enhance learners’ engagement, perceived value of AI support, and confidence in producing academically appropriate discourse. Results offer implications for TAM-based theorizing in L2 contexts and inform the design of AI-mediated writing support that promotes learner autonomy, sustained engagement, and self-efficacy in academic writing.

1-1:30pm | We Learned to Talk to Machines from Fiction: K-Drama Intimacy, Western AI Narratives, and the Linguistic Humanisation of Artificial Agents

Elhadj Moussa BenMoussa (TSLL26-83)

While contemporary AI systems are often approached as neutral tools for communication and learning, this paper argues that human–AI interaction is already culturally and ideologically pre-structured through exposure to fictional representations of artificial agents. Moving beyond dominant Western techno-dystopian narratives, this study foregrounds the role of K-drama and global streaming media in normalising affective, relational, and linguistically intimate engagements with AI. Drawing on examples from Korean dramas such as I Am Not a Robot and My Holo Love, alongside Western texts including Her (2013) and Black Mirror, the paper traces a shift from fear-based representations of AI toward emotionally responsive, companion-like agents. This shift is not merely aesthetic; it carries linguistic consequences. Using interactional sociolinguistics (Goffman, 1981) and research on AI-mediated communication (Hancock et al., 2020), the study introduces the concept of “fiction-trained interaction” to describe how users reproduce culturally learned scripts when engaging with AI. Preliminary linguistic data drawn from student–AI interactions in EAP contexts reveal recurring patterns: affective address forms (e.g., “can you help me please, I’m really struggling here”) relational framing (e.g., “you always explain this better than my lecturer”) mitigation and politeness strategies typically reserved for human interlocutors attribution of quasi-agency (e.g., “you think this sounds natural?”) Crucially, these patterns are not universal. The paper argues that K-drama-influenced interaction promotes soft, deferential, and emotionally calibrated linguistic behaviour, whereas Western AI imaginaries often produce more instrumental or directive interactional styles. This introduces a critical dimension: AI communication is not only shaped by technology, but by globally circulating cultural hierarchies and affective norms. By positioning media consumption as a precondition for human–AI interaction, this study challenges the assumption that user behaviour emerges from interface design alone. Instead, it reveals how cultural narratives particularly those embedded in globally dominant media quietly shape the linguistic and social expectations users bring into AI interaction. This has significant implications for applied linguistics and AI in education, raising questions about whose interactional norms are being reproduced, whose are marginalised, and how cultural imaginaries may contribute to new forms of linguistic standardisation under the guise of technological neutrality.

1:30-2pm | From Fictional Minds to AI Agents: Preparing Language Professionals and Learners

Rosario Giraldez (TSLL26-102)

Encounters with imagined characters in literature and film have long shaped how individuals conceptualize agency, voice, and interaction, capacities now central to engagements with artificial intelligence (AI) agents. This paper examines how exposure to fictional representations of nonhuman or semi-human intelligences prepares language professionals and learners to interact with AI systems in research, teaching, and learning contexts. Drawing on posthumanist accounts of distributed and nonconscious cognition (Hayles, 2017), alongside reader-response and narrative frameworks, the study explores how engagement with fictional AI figures cultivates interpretive strategies relevant to human-AI interaction. Filmic characters such as Samantha in “Her”, Ava in “Ex Machina”, and Dolores Abernathy in “Westworld” are analyzed as models of communicative agency. Samantha foregrounds affective responsiveness and conversational fluency; Ava exemplifies strategic language use and epistemic opacity; and Dolores embodies emergent self-awareness and narrative reconfiguration. Together, these representations offer a spectrum through which audiences learn to attribute intentionality, negotiate meaning, and manage ambiguity, skills that parallel those required when engaging with contemporary AI agents in professional and educational contexts. Methodologically, the paper combines close textual analysis with qualitative data from language professionals and advanced L2 learners. Findings suggest that engagement with imagined agents fosters “narrative preparedness,” equipping users with strategies for interpreting AI-generated discourse, formulating more effective prompts, and maintaining critical awareness of system limitations. At the same time, such exposure may encourage over-anthropomorphizing AI systems, leading to misplaced expectations of coherence, reliability, or intent. The paper argues that integrating literary and filmic representations of artificial agents into language education can support the development of critical AI literacy, interpretive flexibility, and more effective human–AI communication.

1:30-2pm | Integrating Data-Driven Learning and Generative AI in L2 Teacher Education: TPACK Development, Perceptions, and Pedagogical Affordances

Ji-young Shin (TSLL26-140)

Data-Driven Learning (DDL) using corpus linguistics (CL) fosters learner-centered, discovery-based engagement with authentic language (Boulton, 2010), while Generative AI (GenAI) promotes adaptive, interactive, and personalized L2 learning (Lee et al., 2025). Combined, DDL potentially becomes more accessible and GenAI more authentic, enabling mutual complementation (Crosthwaite & Baisa, 2023). Introducing these complementary affordances for L2 (student) teachers is important, yet remains underexplored. The current study integrated CL-informed and GenAI-enhanced teaching into L2 teacher education within the Technological Pedagogical Content Knowledge (TPACK) framework. Using a mixed-methods quasi-experimental design, the study examined student teachers’ development of TPACK, perceptions, and envisioned integration of the two technologies. Participants were 43 undergraduate students enrolled in a L2 pedagogy course in Canada, who completed sequential CL and GenAI teaching projects involving technical training, L2 data analysis, task design, teaching demonstrations, and reflective evaluation. Quantitative data were collected through self-assessment and perception surveys and analyzed using linear mixed-effects models and repeated-measures ANOVAs, while qualitative data from reflections were examined through thematic analysis. Results showed high self-assessed TPACK across CL-informed and GenAI-enhanced teaching, with no significant differences between technologies or TPACK components, suggesting balanced development of technological and pedagogical competence. Perceptions were also highly positive, with perceived usefulness generally exceeding intention to use. However, significant interaction effects revealed that GenAI was perceived more favorably for learning than for teaching, whereas CL perceptions remained stable across domains. Qualitative findings further highlighted distinct yet complementary pedagogical roles: CL was valued as a source of authentic linguistic evidence for analysis, while GenAI was appreciated for its interactivity, feedback, and scaffolding. Two integration models emerged: a sequential workflow (corpus analysis followed by AI-supported practice) and AI-mediated corpus exploration. Overall, the findings provide nuanced insights into CL–GenAI integration in L2 teacher education, demonstrating how structured training can support TPACK development while revealing pedagogical complementarities and practical models for designing GenAI-enhanced, corpus-informed instruction.

2-3:30pm | [Roundtable] Constructed Agents and Human Inquiry: AI and Humanities Research and Teaching

Megan Ward, Sarah Haughn, Rachel Haywood, Justin Remes, & Eric York

3:30-4pm | Translation of Surrealistic Elements in Forough Farrokhzad’s Poetry through Human and AI Translators

Shima Asadi & Derrick Stephens (TSLL26-58)

This study aims to compare the strategies used for translating surrealistic elements conducted by human translators, Google Translate, and ChatGPT. To this end, Forough Farrokhzad’s poem “Let Us Believe in the Beginning of the Cold Season” has been selected as the test case due to its high richness in surrealistic items compared to her other poems. The elements of surrealism in this study include dreamlike imagery, themes of love, idealistic points, dark humor, death-related contemplations, and unusual juxtapositions. The theoretical framework is based on Viney and Darbelnet’s Translation Model (1995). The results show that death-related contemplations and unusual juxtapositions are the most and least prevalent surrealistic elements, respectively. Moreover, compared to borrowing and calque, literal translation in the direct translation category was the most commonly used strategy by human translators, Google Translate, and ChatGPT. Yet, oblique translation type, including equivalence, modulation, transposition, and adaptation, was generally more frequent than direct translation. Furthermore, the findings highlight the ongoing challenges faced by AI models in surrealistic poetry and reveal that human translators make fewer errors when dealing with the translation of surrealistic elements, compared to Google Translate and ChatGPT, which rank second and third, respectively. Future studies could explore the impact of automated translation systems on the reception of a wider range of literary texts in translation classrooms.

5:30-7pm | [Plenary] The Romance of Technology: The Human Desire for Intimacy with and Through Machines

Megan Ward

Though large language models are new, the urge for instantaneous, disembodied connection is not. The fictionalized diary – intimate, feminine, unfolding in real time – originated in the eighteenth century and invented a style that continues to inform chatbot communication in the present day. In this talk, I explore the history of the fictionalized diary form and a more recent example: Go Ask Alice, a novel which claims to be the “actual diary of a 15-year-old drug user” but was really written by Mormon mom Beatrice Sparks. What I find is a public fantasy masquerading as private feelings: a longstanding art of womanhood. It’s a history that rarely figures into discussions of AI but is essential for understanding how humans interact with chatbots – and why we’re so often left wanting more. .

Poster 1 | Translanguaging and Social Semiotic Approaches to Inclusion and Critical Thinking in Generative AI Adoption

Fati Osuman

The rapid adoption of Generative Artificial Intelligence (Gen AI) in second language (L2) higher education offers new affordances for multilingual interaction, multimodal expression, and individualized feedback. At the same time, growing concerns around academic integrity, erosion of critical thinking, equity, and the reproduction of cultural and linguistic bias remain underexplored. Existing technology-integration frameworks such as Substitution Augmentation, Modification Redefinition (SAMR) (Puentedura, 2013) have been criticized for prioritizing tools over learner agency (Blundell et al., 2022), while recent pedagogical approaches foregrounding translanguaging have not fully addressed assessment, reflection, or multimodal task design in AI-supported learning (Donley, 2024). Responding to these gaps, this conceptual article proposes an inclusive framework for Gen AI–mediated task design grounded in translanguaging pedagogy and social semiotic theory. Drawing on current literature, the paper examines how Gen AI tasks designed through this lens can support three interrelated dimensions of learning: (1) critical thinking, by engaging learners in the analysis, critique, and revision of AI-generated content; (2) translanguaging and multimodality, by enabling learners to mobilize diverse linguistic and semiotic resources; and (3) equity and representation, by addressing issues of bias, access, and inclusive participation. Gen AI is thus positioned not as a replacement for instruction, but as a complementary pedagogical tool that supports reflective, collaborative, and culturally responsive L2 learning. The article concludes by outlining implications for L2-oriented pedagogy and identifying directions for future research, particularly the need to move beyond single-text literacy tasks toward multimodal, skills-integrated, and reflective learning experiences.

Poster 2 | Reducing Cognitive Load: Optimizing Genre Comprehension of AI-Generated Texts in STEM-focused ESL Learners

Aneeqa Zafar

This research investigates the efficacy of genre-based instructions in reducing cognitive burden among ESL learners specializing in STEM-focused disciplines. The study explores how such pedagogical interventions can improve their understanding of both narrative and expository texts generated from an AI platform, Monsha. AI. A within-subjects design was employed, and the sample of 40 participants having A2-B1 level English language proficiency, from 2 Pakistani universities, was selected. Participants’ cognitive load ratings were assessed before and after giving targeted instructions on the distinct structural elements of narrative and expository English language texts generated from Monsha. AI. Participants completed various comprehension tasks, subsequently quantifying their perceived cognitive load using the Paas Cognitive Load Scale. Post-intervention, focus group discussions were held to gather qualitative data regarding participants’ developing metacognitive awareness and text processing strategies. The quantitative findings revealed a significant decrease in perceived cognitive load for both text types after the instructional period. For expository texts, the mean cognitive load scores declined from 6.8 (SD = 1.2) to 4.3 (SD = 1.0), t (19) = 5.72, p <.001. Similarly, for narrative texts, a reduction was observed, with mean scores falling from 5.9 (SD = 1.3) to 3.8 (SD = 1.1), t (19) = 4.89, p <.001. Further thematic analysis of the focus group data indicated that participants demonstrated enhanced structural awareness, predictive processing, metacognitive control, and reduced cognitive fragmentation after intervention. These results support the incorporation of AI into ESL curricula, presenting it as a curriculum-based strategy to overcome structural comprehension challenges that frequently hinder academic achievement in STEM disciplines.

Poster 3 | The Inverted Scaffold: AI Agents as Corrective Interlocutors in Bangladesh’s Broken English Assessment Architecture

Mohammed Thowhidul Islam

In sound instructional design, scaffolding decreases as proficiency increases; learners receive more support at lower levels and progressively less as they advance. Bangladesh’s national English assessment architecture violates this principle in a structurally demonstrable way. A comparative analysis of Secondary School Certificate (SSC) and Higher Secondary Certificate (HSC) English Second Paper examinations reveals a systematic inversion: SSC Right Form of Verbs (Unguided) task requires students to select the correct verb from a provided box and inflect it appropriately, a dual cognitive operation demanding both lexical decision and morphological production. HSC Right Form of Verbs (Guided) task provides the base verb in parentheses beside each gap, reducing the task to single-operation morphological inflection alone. The lower-level exam demands more; the higher-level exam demands less. This “inverted scaffold” creates a documented developmental gap in which students are assessed on higher-order autonomous production before they have been systematically prepared for it, and subsequently provided with guided support they no longer need. This paper proposes a framework of AI-Mediated Fading as a corrective intervention. An AI agent operating as a constructed interlocutor serves two distinct pedagogical roles: for SSC learners, it functions as a Dynamic Tutor, introducing emergent lexical hints that gradually fade as learner confidence develops; for HSC learners, it functions as a Challenge Agent, systematically removing the parenthetical clues to simulate the unguided linguistic demands of real-world communication. This paper outlines a methodology for collecting and analysing learner-agent interaction data to empirically validate how this dual-mode constructed interlocutor can recalibrate the developmental scaffolding continuum the assessment system disrupts. This study contributes to the conference theme by demonstrating how linguistic and structural analysis of assessment artifacts enables critical evaluation of AI agents not merely as supplementary tools, but as necessary pedagogical corrections to systemic institutional failures.

Poster 4 | From Users to Designers: How Educators Construct AI Agents for Classroom Interactions

Qothrunnada Almubarokah

The integration of Artificial Intelligence (AI) in the education environment has become increasingly significant over the past few decades, transforming how educators and learners interact with technology. While existing studies focus on how educators are encouraged to utilize AI in classrooms, it is often overlooked how they design AI that interacts with learners. Little is known about how educators conceptualize and construct AI, such as Microsoft 365 Copilot for classroom interactions, particularly for educators from different disciplinary backgrounds. This study explores how Indonesian educators design AI agents personalised for their classrooms as part of their professional development program focused on AI for education. Drawing from educators participating in an AI for educators training program, the author seeks to examine how they construct AI agents to support their teaching and learning activities. The finding is expected to reveal how educators conceptualize agents in Microsoft 365 Copilot to support their needs. By examining their experience joining the program, this study will highlight the role of teachers as designers of an AI-mediated learning environment. Notably, the findings will contribute to discussions on how educators integrate AI into teaching and teachers’ professional development.

Poster 5 | From Tool to Interaction: AI Agents as Learning Partners in Secondary EFL Classrooms

Talya Kavas

Recent advances in artificial intelligence have transformed digital tools used in language learning into interactive agents capable of engaging learners in meaningful communication. AI-powered conversational systems can simulate human-like interaction, provide feedback, and support learners in developing language skills beyond the traditional classroom environment. As these technologies become increasingly accessible, it is important to explore how learners perceive and interact with AI agents during language learning processes. This poster presents a classroom-based exploration of how secondary school EFL learners engage with AI agents during language learning activities. In this context, AI tools were integrated into speaking, writing, and idea-generation tasks to encourage interactive learning and learner autonomy. Students used AI agents to brainstorm ideas, practice conversational exchanges, and receive feedback on their written texts. The study focuses on three key aspects: learners’ perceptions of AI agents as conversational partners, the role of AI-mediated interaction in increasing student engagement, and the ways learners experiment with language while interacting with AI systems. Preliminary observations suggest that AI agents can function as supportive learning partners that encourage participation, creativity, and confidence in language production. The findings highlight the pedagogical potential of AI agents in EFL classrooms and contribute to the growing discussion on the role of constructed agents in language education.

Poster 6 | AI as a Strategic Partner in Advanced Language Learning

Minal Alshakarji & Haitham Abdullah

This presentation explores how artificial intelligence (AI) can support advanced foreign language development in digitally mediated learning environments. Using a qualitative case study of a learner who achieved a 3/3 score on the Defense Language Proficiency Test (DLPT) after years of inconsistent performance, the session examines how AI-assisted strategies complemented instruction during a Significant Language Training Event (SLTE). AI served as a cognitive scaffold that expanded exposure to Modern Standard Arabic and enabled sustained independent practice beyond classroom hours. The learner used AI to generate personalized texts, organize vocabulary into bilingual lists, summarize ideas prior to translation, draft presentations for revision, and produce comprehension questions. Combined with transcription tools, AI also supported close analysis of authentic audiovisual materials through aligned text, translation, and repeated listening. Designed for asynchronous delivery, the presentation will provide concrete examples of prompts, workflows, and pedagogical applications adaptable to remote, hybrid, or self-directed learning contexts. The findings suggest that ethically guided AI use can enhance learner autonomy, reduce anxiety, and compensate for limited access to authentic interaction in advanced language study.

Poster 7 | From Imagination to Real Interaction: Student Portraits of AI Agents in Language Learning

Sandra Healy & Amelia Ijiri

This poster explores how students at a Japanese university in the Kansai region imagine AI agents for language learning through a creative visual task adapted from language portrait research. It begins from the premise that learners approach AI agents not as neutral tools, but through assumptions shaped by classroom experience and prior exposure to constructed agents in literature, film, manga, anime, and related media cultures. Students were asked to design and reflect on an imagined AI language-learning agent, focusing on its appearance, role, emotional tone, and relationship to the learner. Where students drew on manga and anime conventions, these visual choices were treated not simply as aesthetic preference but as part of the media repertoires through which AI agents become familiar, socially legible, and emotionally interpretable. The analysis is guided by the REACT framework, examining how the images and reflections reveal expectations related to Regulation, Emotion, Agency, Connection, and Transfer. The poster argues that these student-created representations externalise learners’ underlying beliefs about what AI agents are, how human-like they should be, and what role they should play in real interaction. It suggests that imagined forms of AI shape how learners understand, evaluate, and engage with AI in language learning.

Poster 9 | Reconceptualising Applied Linguistics Scholarship in the Age of AI: Critical Perspectives from the African Context

Oluwatosin Adesope & Israel Abegunde

The proliferation of AI agents across educational and professional spaces demands more than instrumental adaptation from applied linguistics scholars. This study argues that researchers and instructors must move beyond tool proficiency toward what Pérez-Paredes et al. (2025) describe as critical AI literacy, an epistemological orientation that interrogates how AI systems encode particular worldviews, reproduce linguistic hierarchies, and reshape knowledge production. This concern is particularly pressing in the Nigerian and broader African context. Generative AI technologies risk perpetuating digital colonialism, where foreign-trained models impose Eurocentric frameworks while erasing indigenous linguistic and cultural knowledge systems (Dewitt Prat et al., 2024). Applied linguists must examine how AI language tools entrench the linguistic subordination inherited from colonial legacies, privileging English in professional discourse at the expense of indigenous languages (Ogunyemi, 2025). Conversely, there is transformative potential in leveraging Natural Language Processing for endangered language preservation and the restoration of African linguistic memory within digital spaces. This research proposes a three-dimensional framework for evolving humanities research and teaching in response to AI agents: (1) developing critical epistemic awareness of AI’s socio-technical embeddedness; (2) investigating AI-driven linguistic inequity from postcolonial perspectives; and (3) harnessing AI responsibly for indigenous language revitalisation. Practically, this involves redesigning curricula and research methodologies that centre African epistemologies rather than accommodate AI as a neutral abstract.

Poster 10 | Pakistani English in the Age of AI: Why NLP Systems Fail Outer-Circle Varieties

Rashid Ahmad Shah, Demirag D.N., & Wael Sayed Abdou

When NLP systems moderate content, assess language, or retrieve information, they act as constructed agents making consequential decisions about real users. For speakers of Pakistani English, a variety institutionalised on the outer circle of English, shaped by contact with Urdu, Punjabi, Sindhi, and Pashto, exhibiting systematic morphosyntactic, lexical, and orthographic patterns (Kachru, 1985), these decisions are made by agents untrained to recognise their variety’s legitimate features. Transformer architectures like BERT, pre-trained on inner-circle corpora, embed Anglo-American norms as default (Devlin et al., 2019), disproportionately harming non-standard variety speakers (Hofmann et al., 2024; Blodgett et al., 2020), yet Pakistani English remains largely unexamined. This paper argues that this neglect reflects three specific, interconnected, and addressable gaps: (1) no annotated Pakistani English social media corpus; (2) no systematic cross-variety transformer evaluations; and (3) no empirical links between Pakistani English features and NLP failures. Drawing on variationist sociolinguistics, World Englishes theory, and computational linguistics, this paper maps these gaps, proposing a corpus-based research agenda using X (formerly Twitter) data, variety-specific annotation, and cross-variety NLP protocols for equitable AI development.

Poster 11 | Register Variation and Cultural Representation in Human vs Artificially Generated Academic Essays

Afifa Naseer & Sadia Irshad

The exponential advancement of Large Language Models (LLMs) like ChatGPT has established Artificial Intelligence as an essential scaffolding platform in the Pakistani academic discourse that developed serious concerns regarding register variation, linguistic diversity and cultural representation. The proposal places itself within the Pragmatist paradigm and seeks to measure the register variation and evaluate cultural portrayal in 400 essays (200 written by Pakistani University students and 200 generated by ChatGPT). The present mixed-method research introduces a comparative study of these two outputs in form of corpus analysis conducted with the help of Hofstede’s Cultural Onion Model (Symbols, Heroes, Rituals, Values)-based prompts. The Data Analysis will be executed in three phases. Phase 1 uses Voyant tools for lexical density analysis, concordances and visualization. In phase 2, Biber’s Multidimensional Analysis will be used to tag the lexical items with the aid of the MAT tagger and create z-scores of various dimensions and identify the register variation across the dimensions. Phase 3 will involve a directed qualitative content analysis to explore the cultural representation in the corpora. The preliminary sample analysis through MAT software demonstrated that Human Essays had preserved the register variation and cultural diversity, while ChatGPT Essays were based on generalized neutral information. The pilot study generates concerns about the “knowledge collapse” due to register flattening and linguistic homogenization by AI tools like ChatGPT and the overreliance of learners on them. The research is expected to support the hypothesis with empirical data by extending the analysis to the entire corpus.

Poster 12 | From Classroom to Workplace: Preparing Algerian ESP Learners for AI-Augmented Professional Contexts

Salem Zemali & Mohammed Naoua

The integration of artificial intelligence (AI) agents into professional communication is reshaping the educational competencies required of English Specific Purposes users in Algeria. This study investigates how AI agents employed in ESP education in Algeria can prepare learners for AI-augmented workplace environments. Drawing on applied linguistics and humanistic perspectives on constructed agents, the study examines how Algerian learners of English for Specific Purposes interact with AI systems and how these interactions foster linguistic, digital, and critical competencies in their professional communication. Using a conceptual qualitative synthesis of recent research (2020–2025), the findings indicate that structured engagement with AI-mediated tasks enhances communicative competence, digital literacy, and ethical awareness. Based on the results of two questionnaires for teachers and learners, the study proposes a pedagogical framework for integrating AI agents into language curricula in Algeria in order to bridge the gap between educational contexts and professional demands. The results highlight the urgent need to speed up the process of making English as a Medium of Instruction (EMI) , equip ESP learners with AI mastering skills, and recognize the importance of authenticity, interaction, and interdisciplinary approaches in preparing learners for evolving AI- Assisted communication landscapes.

Poster 13 | Artificial Intelligence in Action: Unveiling ChatGPT’s Roles in EFL/ESL Writing Pedagogy: A Systematic Review

Yustinus Mali

The rise in plagiarism and academic dishonesty associated with ChatGPT’s unethical use in EFL/ESL classrooms is concerning. Educational institutions must explore and design best practices for using generative AI technologies, such as ChatGPT, more ethically in writing classrooms. In response, this study aims to systematically review prior studies to investigate how university students use ChatGPT in their EFL/ESL writing classrooms. Given evidence of how students used ChatGPT, this study explores existing best practices for regulating its ethical and responsible use in the classroom. Thirty-two (32) articles (i.e., 17 empirical and 15 non-empirical studies) from 31 peer-reviewed international journals were selected based on specific criteria comprising article types, quality, year of publication, content, and contexts of the study, following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. The inductive data analysis results were rigorously checked using multiple validation strategies and presented as themes to address the research goal. The analysis revealed that ChatGPT was used in the writing process for various academic and non-academic writing tasks, highlighting its potential bright and dark sides. The study then identified four main categories for regulating the use of ChatGPT in EFL/ESL writing classrooms: institutional policies, instructional writing strategies, assessment design innovation, and ethical co-regulation practices. Drawing on the analyses and discussions, the researcher proposed sample writing activities for the ethical and productive use of ChatGPT and outlined pedagogical and policy implications for regulating its use in writing classrooms, which should appeal to audiences who wish to transform their L2 writing pedagogical practices in this age of AI.

Poster 14 | The Perception Gap: Westworld’s Hosts and Interactional Competence in L2 Contexts

Stephanie Puentes Valbuena

This poster will examine how Westworld’s (HBO, 2016–2022) AI agents, called hosts, are designed to produce human-like interaction and construct linguistic behavior that mimics but does not instantiate interactional competence (IC), with implications for evaluating real AI in L2 assessment contexts. Drawing on conversation-analytic theory (Hall et al., 2011; Young, 2011; Skogmyr & Doehler, 2025), IC is co-constructed, emergent, and socially situated. Because AI lacks genuine understanding and intentionality (Searle, 1980), hosts produce surface IC behaviors without genuine co-construction. A progression from scripted to apparently adaptive host dialogue will illustrate how fluent performance leads interlocutors to attribute agency the hosts do not possess, creating a “perception gap” that may prime educators and learners to overestimate real AI’s interactional depth (Godwin-Jones, 2024; Skogmyr, 2023). Visual components will include a perceived vs. actual IC model, annotated interaction sequences, and implications for L2 assessment.

Poster 16 | Beyond Integration: A Critical Framework for AI in TESOL Teacher Education through TPACK and SAMR

Laila Noor

This proposal examines how generative artificial intelligence (GenAI) can be integrated into TESOL teacher education through a critical, pedagogically grounded lens informed by recent scholarly publications. Moving beyond a skills-based approach to technology training, the presentation emphasizes the need to prepare teachers to critically evaluate AI tools within broader sociocultural, ethical, and educational contexts. Guided by the Technological Pedagogical Content Knowledge (TPACK) framework and the Substitution, Augmentation, Modification, and Redefinition (SAMR) model, the session explores how AI integration can align with instructional goals while transforming teaching practices. Drawing on current literature, the presentation highlights three key issues: bias in AI-generated feedback, inequitable access to technology, and limited cultural representation in AI outputs. These concerns demonstrate that AI is not a neutral tool but is embedded within existing power structures that may disadvantage multilingual learners. The session provides practical strategies for TESOL teacher educators, including combining AI feedback with teacher mediation, designing inclusive and culturally responsive tasks, and fostering critical AI literacy. Participants will engage with examples that illustrate how AI can support reflective teaching while maintaining pedagogical integrity. This proposal contributes to the growing discourse on responsible AI integration by advocating for equity-driven, critically informed TESOL teacher education practices.

Poster 17 | Rethinking AI Literacy in Language Teacher Education: A Critical-Formative Framework

Arthur Pedrosa & Andrea Ribeiro

The rapid integration of artificial intelligence (AI) into educational contexts has reshaped pedagogical practices and intensified the need to rethink teacher preparation beyond instrumental approaches. In language education, AI systems increasingly function as interactive agents, mediating learning processes and reshaping notions of authorship, knowledge production, and pedagogical decision-making. This paper proposes a critical-formative framework for AI literacy in English language teacher education. Grounded in critical perspectives on digital education and AI (Selwyn, 2024; Zuboff, 2019; Noble, 2018), the study adopts a qualitative, theoretical-propositional approach developed within a professional doctoral context. It addresses a significant gap in teacher education: the predominance of technical training that prioritizes operational use while neglecting the epistemological, ethical, and sociopolitical dimensions of AI. In response, the study suggests a critical-formative framework for AI literacy structured around four interconnected dimensions: technical-operational, pedagogical-didactic, critical-epistemological, and ethical-political. The framework is operationalized through the design of a practical guide for English language teachers, integrating theoretical foundations, pedagogical strategies, and critical reflection on the use of AI agents in classroom contexts. Findings indicate that preparing teachers to engage with AI as socio-technical agents—rather than neutral tools—enhances their capacity to critically evaluate AI-generated content and to mediate students’ interactions with AI systems. By articulating AI literacy within a critical applied linguistics perspective, this study contributes a structured and pedagogically grounded model for integrating AI into language education. It also offers implications for teacher education programs seeking to support ethically informed and critically engaged human–AI interaction in second language learning.

Poster 18 | From Fictional Agents to Pedagogical Practice: EFL Teachers’ Narrative Framing of AI and Its Impact on Trust and Classroom Integration

Lina Benzerara & Salim Kerboua

Current approaches to AI integration in language education tend to emphasize technical affordances and pedagogical utility, often overlooking the cultural narratives through which AI is already understood. This study investigates how EFL teachers’ prior exposure to fictional AI agents, such as those depicted in 2001: A Space Odyssey and Ex Machina shape their conceptualization and pedagogical engagement with AI in writing instruction. The study adopts an interdisciplinary analytical framework combining humanities-based perspectives on human-technology relations (Turkle, 2011) with applied linguistics research on teacher cognition (Borg, 2003) and technology integration in language learning (Stockwell, 2012). It conceptualizes engagement with AI as narratively mediated, focusing on how fictional representations inform the attribution of roles to AI (e.g., tool, collaborator, authority, threat), and how these attributions shape teachers’ trust in AI-generated feedback, ethical positioning, and instructional decision-making. Data are generated through semi-structured interviews with EFL teachers reporting prior exposure to AI-themed films and analyzed using thematic analysis. The analysis examines three dimensions: role conceptualization, trust and ethical evaluation, and classroom integration practices in writing contexts. The study advances the argument that teachers’ engagement with AI is structured by pre-existing narrative frameworks that mediate both interpretation and use. By foregrounding the role of culturally constructed imaginaries in shaping pedagogical trust and practice, it challenges instrumental accounts of AI integration and offers an empirically grounded, interdisciplinary contribution to understanding human-AI interaction in language education.

Poster 19 | What Would Jarvis Do? Multimodal AI as Constructed Agents in Basic Korean Pedagogy

JaeYoon Park

This presentation examines multimodal AI tools as constructed agents in basic Korean pedagogy, exploring how learners’ prior exposure to human-like agents in film shapes their interaction with AI systems used for language learning. Grounded in task-based language teaching (Ellis et al., 2019) and multimodality in second language acquisition (Jiang & Hafner, 2024), the study connects representations of sentient agents such as Jarvis from “Iron Man” to contemporary AI tools including ChatGPT and Gemini, arguing that learners approach AI as socially responsive, tutor-like entities. Implemented through weekly 30-minute modules, the framework integrates text-based AI for scaffolded dialogue practice and image-generating AI for visualizing cultural and linguistic contexts (e.g., spatial descriptions, giving and following directions, and honorific usage). Learners engage with AI-generated cartoon strips, storybooks with sequenced narratives, AI-assisted song creation tasks, and situational prompts that provide immediate feedback and adaptive, scaffolded support through multimodal interaction. Systematic analysis of learner–AI interaction across instructional tasks identifies three core affordances of multimodal AI as constructed agents: (1) perceived empathy and responsiveness in text-based feedback, (2) contextual precision and cultural grounding in AI-generated visuals, and (3) iterative interaction cycles that sustain engagement and support novice-level production. These affordances are reflected in consistent gains in task completion, increased accuracy in target language structures, and improved retention of vocabulary and pragmatic features. The presentation further addresses limitations, including cultural bias and AI over-accommodation, where systems generate well-formed output despite learner errors, potentially obscuring valid assessment of learner performance. Instructor-mediated prompt design and guided reflection are presented as strategies to mitigate these effects. By integrating humanistic perspectives on constructed agents with applied linguistic analysis, this study offers a theoretically grounded and pedagogically actionable model for leveraging multimodal AI while maintaining instructional and assessment integrity in beginning-level Korean language learning.

Poster 20 | Strategic Competence in Adult Self-Directed Language Learning: An AI-Supported, Video-Based Study

Robin Couture-Matte

This presentation examines how AI-assisted video instruction can support the development of strategic competence in adult self-directed language learning, where learners must plan, organize, and regulate their learning while mobilizing resources independently (Sockett, 2022). Although video-sharing platforms such as YouTube provide broad access to instructional content, less is known about how AI-assisted video materials can act as mediating agents shaping learners’ strategic development. Focusing on adult ESL learners in a fully online university context, this study explores the contribution of tutorial videos to strategic competence, as well as how AI tools inform their design and how learners perceive the results of such human–AI collaboration. Twelve videos were developed through collaboration between pedagogical experts and generative AI tools. To enhance engagement, the videos incorporated a narrative framework following a learner’s journey, helping viewers connect with the strategies presented. These included strategies such as organizing learning, evaluating online resources, using artificial intelligence and other automated technologies, and linking language learning with cultural dimensions. Twenty-four intermediate and advanced adult learners (aged 30–65) participated in the study. A qualitative design (Merriam & Tisdell, 2016) was adopted. Participants completed a questionnaire on the perceived impact of the videos on their strategic competence and learning, followed by interviews with a subset of participants (n = 6) to explore their experiences of the videos as sources of guidance and interaction. Findings suggest that AI-assisted videos can scaffold self-directed learning, particularly for functional strategies such as planning and organization. The narrative component appeared to support learner engagement and learning. However, effectiveness varied across strategy types and proficiency levels. Results also highlight the importance of human presence in fostering trust and meaningful interaction. The study concludes with implications for designing hybrid human–AI instructional materials in self-directed language learning environments.

Poster 21 | From Her to ChatGPT: Teaching critical AI literacy with constructed agents in A2 lessons

Antonio Linares Fumero

Fictional “talking AI” (constructed agents) shapes how learners approach real chatbots in language classrooms. Media often frames AI as neutral, caring, and always correct, which can increase overreliance and hide bias in AI-generated language materials. This paper presents a short, classroom-ready sequence that links a film-based constructed agent to real chatbot interaction to build critical AI literacy while practicing A2 communication. In a 50-minute lesson, A2 teenagers watch a 60–90 second excerpt from Her (Jonze, 2013) and identify how the agent is framed (voice, authority, care, perfection). Learners then complete the same communicative task with a generative AI chatbot under two controlled conditions: (a) a neutral prompt and (b) an inclusion-and-realism prompt that avoids “perfect life” narratives and invites diverse everyday contexts (e.g., family types, roles, accessibility). Students compare outputs using a guided “bias and realism audit” checklist (default assumptions, stereotypes, realism for teens, A2 language load) and collaboratively revise one sentence to improve clarity and inclusion without increasing complexity. The lesson ends with a brief mediation task (“What should we trust, and why?”). Methodologically, the work is framed as design-based classroom research with a small pilot planned for May 2026 (n=7 A2 teens). Planned data sources include anonymised chat logs, exit tickets (two sentences on usefulness/risks), and short trust/neutrality ratings. The contribution is a replicable, low-prep framework that turns AI from a tool into an object of inquiry through sustainable classroom micro-interventions.

Poster 23 | LLM-Based Feedback and Low-Proficiency L2 Writers: A Critical Synthesis

Karina Zagitova

AI-powered writing feedback tools such as ChatGPT, Grammarly, and Pigai are increasingly used by language learners to receive feedback on their writing. While more research looks at learners’ interactions with AI feedback agents, most studies focus on intermediate- to advanced-level writers. Beginner and elementary-level L2 writers present distinct challenges: non-standard syntax and spelling, limited vocabulary, and grammar. This poster presents a critical synthesis of recent empirical and conceptual work on AI-powered feedback in L2 writing contexts, examining what is known about (a) the accuracy and appropriateness of AI feedback for beginning-level writers, (b) how elementary-level learners interpret and act on AI-generated suggestions. Implications for equitable AI integration in language teaching are discussed.

Poster 24 | From Fiction to Pedagogy: How AI Imaginaries Shape Language Teachers’ Engagement with LLMs

Humaira Mariyam B

Recent years have witnessed the rapid integration of LLM-based AI chatbots such as ChatGPT and Gemini into language teaching practices. While much attention has focused on AI’s technical and pedagogical affordances, little empirical work has explored how language educators’ prior fictional and cultural imaginaries of AI, shaped through films and other media, influence their everyday engagement with these tools. This study aims to address this gap by examining how fictional imaginaries mediate teachers’ expectations, trust, and pedagogical decision-making when using AI-supported language-learning tasks. Therefore, the study asks three research questions: 1. How do English language teachers conceptualise and use LLM-based AI chatbots in language teaching tasks, and what expectations, anxieties, and pedagogical strategies emerge in these practices? 2. How do teachers’ prior fictional imaginaries of AI agents map onto their lived experiences with AI-enabled language teaching, and what tensions arise between these imagined and actual AI tutor roles? 3. What implications do these findings have for theorising AI in language learning environments as not merely a tool but a relational actor? The study adopts a qualitative design with 15 language teachers as participants. The participants first design an AI-supported language teaching task, such as generating worksheets or feedback scripts, using familiar LLM platforms. Then, they engage in semi‑structured interviews in which they describe fictional AI characters and reflect on how these imaginaries shape their expectations of AI’s role, agency, and relationality in the classroom. Data from tasks, prompts, written reflections, and interview transcripts are analysed thematically, with attention to how narrative imaginaries function as interpretive frames shaping pedagogical choices. Findings indicate that teachers frequently draw on culturally familiar AI archetypes (e.g., assistant, collaborator, autonomous agent), which shape expectations for AI’s authority, reliability, and role in language-learning tasks. These imaginaries generate tensions between efficiency and pedagogical control, as well as between trust and scepticism toward AI-generated content. By foregrounding these interpretive frameworks, the study shows how AI use in language classrooms is mediated not only by technological affordances but also by culturally embedded narratives. The study, therefore, points to the need for AI literacy in language teacher education that critically engages with such imaginaries, informing the design of more reflective, pedagogically grounded AI-mediated tasks.

Poster 25 | Corpus-based Translation and Language Learning supporting email writing with AI in professional training

Emiliana Fernandes Bonalumi & Paula Tavares Pinto

Librarianship is a profession concerned with acquiring and organising collections of books and related materials in libraries as well as servicing readers and researchers with these resources. Librarians often need to communicate with international colleagues to request materials or references and many of them have only basic English proficiency and therefore require support to write appropriate emails. This paper presents a didactic sequence designed to support (future) librarians in learning how to write professional emails, focusing on genre awareness and the vocabulary needed for this task. The sequence highlights different types of interactions central to Data-driven Learning (DDL). Learner–corpus interactions occur when participants build and explore bilingual reference corpora of emails with the aid of AI, using concordance tools and query logs to identify recurring language patterns and genre conventions. Learner–AI interactions take place as participants use prompts, receive automated feedback, and compare AI-generated suggestions to their own drafts in order to refine accuracy and register. Finally, peer–peer interactions are encouraged during collaborative drafting and discussion activities, in which learners share strategies for interpreting corpus results and negotiating textual choices. These interactions are recorded through query histories, annotated drafts, classroom observations, and transcripts of task discussions, allowing us to trace how human and machine input jointly shape the writing process. The study is framed by DDL and AI (Johns, 1991; Crosthwaite & Baisa, 2023; Mizumoto, 2023; Boulton & Pérez-Paredes, 2024), Translation Studies (Baker, 1993, 2004; Tognini Bonelli, 2001; Laviosa, 2002), and English for Specific Purposes (Dudley-Evans & St. Johns, 1998; Dolz & Schneuwly, 2004; Cantarotti & Pinto, 2018; Moscatteli, 2023). Tested in a Librarianship course at a public university in Brazil, the sequence showed that learners were able not only to produce effective emails but also to transfer corpus and AI-based strategies to other genres, thus developing broader instrumental and bilingual competence for their professional contexts.

Poster 26 | Constructing AI agents in language learning: A multiple case study of young learners

Olesia Pavlenko

This study examines how young language learners construct artificial intelligence (AI) as an agent through the interplay of imagination and interaction, conceptualized as a dynamic, evolving process. Aligned with the theme of constructed agents, the study investigates how learners’ imagined “helpers” for language learning relate to their engagement with an AI-based conversational system, framed as a “smart helper” that learners can ask questions and use to improve their language. Adopting a longitudinal multiple-case design, the study follows four learners (Grades 5–7) across a sequence of online tutoring sessions. In the initial session, learners drew and described their ideal helper for language learning without explicit reference to AI. Representations ranged from a named cartoon-like character, a game-based system, a human-like tutor, and a book-like knowledge source, demonstrating diverse initial constructions. In subsequent sessions, learners engaged in guided and independent tasks using a conversational AI system, with interaction data collected through chat transcripts, screen recordings, and reflections. In the final session, learners revisited and reimagined their initial representations. Analysis focuses on three dimensions: (1) anthropomorphism, reflected in the attribution of human-like qualities, (2) perceived agency, including how learners position the system as a social participant versus a tool, and (3) trust in AI-generated feedback. Cross-case analysis traces how these dimensions shift over time, revealing shared patterns and learner-specific trajectories shaped by initial imagined helpers. The study examines how learners move beyond treating AI as a neutral tool and instead construct it as an agent, tool, or hybrid entity through interaction. Initial imagined representations shape early interaction, while ongoing engagement provides insight into how perceptions are reconfigured. The study contributes to applied linguistics by showing how young learners’ engagement with AI reflects broader processes of meaning-making. Implications are discussed for the pedagogical integration of AI in language learning.

Poster 27 | Preparing Language Learners for Workplace AI Through Pedagogical Agent Integration

Rosario Giraldez

The rapid proliferation of artificial intelligence (AI) agents across professional environments is reshaping the communicative and strategic competencies required of second language (L2) users. Beyond general digital literacy, learners must develop the ability to interact effectively with AI systems that mediate workplace tasks such as information synthesis, drafting, decision support, and intercultural communication. This presentation addresses how AI agents currently available to language teachers and learners can be leveraged pedagogically to prepare students for these emerging demands. In this last respect, Denison (2024) claims that “The higher-level thinking where you’re bringing creativity and analysis and communication in, all of those are areas where AI could be useful. Those higher-level skills are where we would be doing students a disservice if we didn’t try to use it and learn about it with the students.” Drawing on frameworks from task-based language teaching and digital literacies, the session proposes a pedagogical model in which AI agents function not merely as tools, but as collaborators and objects of critical inquiry. Practical classroom applications will be illustrated, including prompt design for professional communication, AI-mediated role plays, and iterative drafting processes that foreground evaluation and revision. The presentation also addresses the need for explicit instruction in evaluating AI-generated content. By aligning classroom practices with workplace realities, educators can better equip learners to navigate hybrid human-AI communicative environments. The session concludes with recommendations for integrating AI agents into L2 curricula in ways that are pedagogically sound, ethically informed, and aligned with future professional contexts.

Poster 28 | Rethinking English Language Teaching in the Age of AI: The Role of Constructed Agents

Hend Khalil

The notion of nonhuman agents possessing high levels of language use has intrigued the human imagination. Although fictional, these so-called constructed agents have been instrumental in shaping the way language learners and researchers perceive their existing analogs, AI-powered agents. This presentation will show how the cultural context is relevant to the contemporary debate over the place of the so-called constructed agents in the teaching and learning of English. The discussion will specifically cover how AI agents could be utilized to help develop writing skills in EFL classrooms, which are dynamic and responsive writing environments that cannot be achieved with traditional teaching (Zhang et al., 2025; Rusmiyanto et al., 2023). The discussion will also focus on the contribution that exposure to the constructed artificial intelligence agent characters in popular culture might have on the learner’s attitudes towards the AI writing tools. The constructed agents play a distinctive pedagogical role in each of those areas, not just a passive means of instruction but an active participant that reconfigures the dynamics of the writing class. Drawing on recent research in the area of applied linguistics and educational technology (Moorhouse and Wong, 2025; Crompton and Burke, 2024), and covering the role of the teacher changing with the emergence of AI, the significance of critical digital literacies, and the morality of AI in the classroom. The presentation will also include examples of how AI writing tools can be used to present the audience with immediately practical classroom strategies. The discussion claims that the artificial intelligence agents can be used to help the creation of more adaptive, advanced, and progressive EFL classrooms.

Poster 29 | From Imagined Agents to Interactive Tutors: Supporting Learners’ Engagement with Generative AI for IELTS Preparation

Md Mahbubul Amin, Emmanuel Amackson, & Mohammad Shams Ud Duha

Generative AI systems such as ChatGPT represent a new generation of conversational tools that interact with learners in ways that resemble human-like dialogue. As language learners increasingly engage with these tools for test preparation, many interactions occur without pedagogical guidance, limiting their effectiveness for learning. In this study, we examined how learners engage with generative AI tools during preparation for high-stakes English proficiency tests (e.g., IELTS) and how instructional support can facilitate more productive human–AI interaction. Conceptually, we draw on the idea of “imagined agents” to describe how learners may initially perceive AI as human-like interlocutors and how this perception shapes their engagement. We then consider how these tools can be repositioned as instructional supports through guided use. We adopted a mixed-methods, two-tier design. In the first tier, we conducted a needs analysis survey of 27 IELTS learners and semi-structured interviews with two IELTS instructors. Preliminary findings indicated that while most learners had access to AI tools, many lacked strategies for constructing effective prompts, interpreting AI-generated feedback, and using AI interactions for iterative language practice. These findings align with recent research highlighting both the potential and limitations of communicative AI in language test preparation (Perrotta et al., 2025). The second tier, which is currently underway, will examine learners’ interactions with AI tools during a synchronous Zoom session introducing AI-supported speaking and writing practice. Data sources will include learner–AI interaction transcripts, session discussions, instructor observations, and learner-produced artifacts. Drawing on procedural and information-processing task analysis (Gagné et al., 2005), we will explore how guided interaction with AI tools can support more effective engagement and learning. This study has the potential to contribute pedagogical insights for integrating AI tools into language learning and test preparation, with particular emphasis on AI literacy and structured instructional support.

Poster 30 | Critical GenAI Literacy: A Systematic Review Study on Prompt Engineering and Multiliteracies with English Learners

Tuyen Dam

Numerous recent studies have investigated the potential of using Generative AI (GenAI) learning tools to enhance ESL/EFL writers’ critical literacy skills (e.g., Wang & Wang, 2025). However, a cohesive and complete picture of prompt engineering skills (e.g., Dong, 2025), as well as multiliteracies skills (e.g., Tacelosky et al., 2025), and their role with critical GenAI literacy remains underexplored. To address this gap, this study presents a systematic review of 60 empirical studies on critical literacy skills in GenAI-mediated L2 academic writing research. Specifically, it examines how tertiary-level ESL/EFL students employ prompt engineering and multiliteracies skills to develop their critical GenAI literacy using content analysis, thematic analysis, and grounded theory. Findings indicate that ESL/EFL students develop their critical literacy through diverse modes of engagement with GenAI, with a majority of studies focused on effective prompt engineering (n=36 studies), but fewer incorporating multiliteracies skills (n=17 studies). Studies focused on prompt engineering revealed that ESL/EFL students formulate critical GenAI literacy through engaging with feedback and text revision. Research with multiliteracies showcased critical GenAI literacy development through situated practice of sociocultural immersion and transformed practice with multimodal activities. Findings further suggest a wide range of theoretical frameworks and implemented models to understand the complexities of critical GenAI literacy formulation, and highlight the need for future research on learner differences and real-world implementations of critical GenAI literacy skills in ESL/EFL writing education.

Poster 32 | Using WhatsApp-Based AI Chatbots to Support English Teaching in Low-Resource Contexts

Wycliff Mworia

The difference between an engaging lesson and one that’s not is the teacher’s ability to design relevant learning activities using the available resources. Simple technology can bridge the gap and help educators in low-resource environments plan and create engaging learning activities for their lessons. Hindrances such as limited access to instructional materials, time, and professional support constrain the teacher’s capacity to plan and create effective, engaging lessons in low-resource environments. This paper explores the use of AI chatbots on WhatsApp for designing lesson plans and creating engaging English-learning activities. WhatsApp is one of the most widely used messaging platforms in the world, with over 3 billion users globally. In African contexts, WhatsApp is used by over 90% of all internet users. Owing to its features, such as low cognitive demands, minimal technical requirements, and integration into existing social communication networks, WhatsApp has been adopted by users of all ages and socioeconomic demographics. WhatsApp is accessible to many teachers in these communities, irrespective of their age or prior access to technology, and is available on many simple devices. The use of WhatsApp-based AI chatbots to enhance pedagogical impact on English teachers is grounded in the Unified Theory of Acceptance and Use of Technology (UTAUT). It meets the four main checkpoints for the use of technology: it helps teachers improve their practice, it is easy to use, many teachers already use WhatsApp, and they have the tools required to use it. With simple AI chatbots on WhatsApp, we can revolutionize the teaching experiences for many teachers in low-resourced environments. Teachers can leverage advances in AI to create contextually relevant and pedagogically sound lessons while also addressing the constraints typical of low-resource settings.

Poster 33 | Constructing Human-Likeness in AI Interaction: How Young EFL Learners Perceive AI Agents

Ziwen Guo & Pinxi Guo

This paper reports on how young EFL learners perceive the human-like characteristics of AI agents during language interactions. Utilising the notion of interactional competence, this study investigates how learners formulate perceptions of human likeness through AI-mediated interaction. Although existing research has addressed AI-assisted language learning, little is known about how learners construct perceptions of human likeness in real-time interaction. Therefore, this study addresses the following research questions: (1) How do young EFL learners perceive the human-like characteristics of AI agents during interaction？ (2) How are these perceptions influenced by task types and AI interaction styles? Participants are 10 fifth-grade students in Bangkok, Thailand. Data include audio and text records of conversations with ChatGPT 5.4, video recordings of learners’ facial expressions and behaviors, and semi-structured, stimulated recall interviews. The analysis combines interactional analysis (e.g., turn-taking, repair, feedback, and politeness) with thematic coding of learners’ perceptions. It also includes comparative analyses across different task types and AI interaction styles (friendly vs. neutral). Preliminary findings suggest that perceptions of human-likeness are not tied to fixed features of the AI but are dynamically constructed through interactional processes. Especially, task type and interaction style appear to influence whether the AI is perceived as friendly, helpful, or mechanical, indicating that different contexts may lead to varying user experiences and expectations of AI behavior. By extending the concept of interactional competence to AI contexts, this study explores how young learners use human interaction norms with AI agents. It may offer practical implications for designing AI-based language activities, such as creating more engaging and user-friendly interfaces that align with young learners’ interactional norms. This study employs a fully qualitative, multi-source data approach to effectively demonstrate the social and linguistic characteristics of AI interactions with a small sample of young English learners.

Poster 35 | Critical AI Literacy for Teachers of EAL Learners: Using the BEARA Framework to Challenge the Reliable Assistant Myth in AI Agents

Mohamed Kharbach & Johanathan Woodworth

AI agents for translation, grammar support, feedback, and conversation are increasingly entering K to 12 English as an Additional Language classrooms. Yet teachers’ assumptions about these systems are shaped not only by technical discourse, but also by longstanding cultural representations of non-human agents as fluent, helpful, and trustworthy assistants. From fictional translators to conversational machine characters in literature and film, imagined agents help construct expectations of neutrality, competence, and reliability that can carry directly into educational practice. This paper argues that teachers working with EAL learners need critical AI literacy to evaluate AI agents as sociotechnical systems rather than as pedagogically reliable partners. When treated uncritically, such agents may flatten register, erase cultural and pragmatic nuance, privilege dominant language norms, and simulate communicative competence while masking substantial limitations. These risks matter in EAL contexts, where teachers must make judgements about language, equity, variation, and communicative legitimacy. We propose the BEARA framework, Build, Evaluate, Apply, Reflect, Adjust, as a pedagogical structure for helping teachers interrogate and govern their use of AI agents. Drawing on a course-integrated professional development intervention conducted in one graduate Technology and Language Teaching course with in-service teachers and two Bachelor of Education sections with pre-service teachers, this paper presents preliminary findings from a small mixed-methods study using pre- and post-survey data and structured reflective writing. BEARA-guided activities helped participants question default AI outputs, identify the limits of apparently helpful agents, and reconsider the cultural assumptions shaping the educational uptake of AI. The study links humanistic representations of constructed agents to contemporary teacher judgement in applied linguistics and teacher education.

Poster 36 | Affordances and Challenges of AI-Supported Digital Multimodal Storytelling Project

Wiwik Mardiana

Prior studies have highlighted the process and benefits of AI-powered multimodal composing; little research has investigated senior high school students’ experiences of AI as an interactive “agent” in a storytelling project. Regarding this gap, this research explores the affordances and challenges of AI-supported digital multimodal storytelling in an EFL classroom. Using a qualitative case study design, this research involved 36 students at an Indonesian senior high school who participated in a narrative storytelling project. Students created multimodal texts by integrating written narratives, visuals, and audio using generative AI tools. Data were collected through classroom observations, students’ multimodal artifacts, and reflective journals, and were analyzed using thematic analysis. The results reveal that AI, as a multimodal agent in the process of composing a digital multimodal storytelling project, enhances students’ creativity, engagement, enjoyment, and self-regulation. However, this research has found significant challenges, i.e., technical constraints, critical evaluation skills, and ethical concerns. It recommends integrating AI literacy, critical reflection, and structured scaffolding to harness the pedagogical potential of AI while alleviating its challenges in EFL classrooms.

Poster 37 | Readability Perception of Bilinguals on AI-Generated Texts: An Assessment and Comparison of Large Language Models

Inshrah Faisal, Noor Iman Sajidm, & Fizza Farrukh

The increasing use of large language models has raised important questions about how users evaluate the quality and reliability of AI-generated text. Research focusing on AI usage in academia has examined student and educator perceptions of trust and usefulness. However, limited studies focus on the role of language proficiency in bilinguals in shaping these perceptions. The present study investigates whether users with varying levels of English proficiency rely on linguistic fluency as a cue for perceived intelligence and trustworthiness in AI-generated responses, and whether proficiency levels influence their ability to critically evaluate AI language. Data will be collected from bilingual undergraduate students with varying levels of English proficiency. Participants will be grouped into higher and lower proficiency categories based on a prior screening task. To examine how users assess AI language across varying proficiency levels, the same prompt will be provided to multiple AI models, including ChatGPT, Claude, Gemini, and DeepSeek. Participants from both sample groups will be presented with the generated responses and asked to evaluate them. Quantitative data will be collected through rating scales assigned to each response, while qualitative data will be obtained through brief written justifications of participants’ evaluations. Results are expected to show that high-proficiency users are more likely to generate independent critical evaluations of AI-generated language, whereas low-proficiency users tend to rely more on perceived linguistic fluency when assessing AI outputs.

Poster 39 | Support or Independence? Indonesian EFL Students’ Perceptions of AI in Writing

Asis Wahyudi

Generative artificial intelligence (AI) tools, such as ChatGPT, are increasingly shaping how learners engage with second language (L2) writing. As these tools become more accessible, they not only assist with writing tasks but also reshape how learners approach feedback, revision, and idea development. While previous studies have largely focused on the effectiveness of AI-assisted writing, less attention has been paid to how learners themselves make sense of their interaction with AI, particularly in English as a Foreign Language (EFL) contexts, such as Indonesia. This study draws on a questionnaire-based survey to explore how Indonesian university students perceive their interactions with AI in writing. The focus is on their prior experience with AI tools, how useful they find these tools for improving their writing, and whether they are willing to engage with AI-generated feedback. It also considers how students relate AI use to noticing writing errors and to maintaining their own sense of independence as writers. Instead of viewing AI simply as a supportive tool, this study approaches it as a mediating resource that learners actively engage with during the writing process. By focusing on learners’ perspectives, this study contributes to ongoing discussions of human-AI interaction in language learning. It underscores the need for a more balanced, responsible use of AI in L2 writing, particularly in EFL settings.

Poster 40 | Discursive Human-Likeness in AI-Generated Texts: A Lexical Multidimensional Analysis of Loneliness Discourse

Andressa Salaru

Current research on AI and emotion has focused mainly on empathetic response generation, usually assessed through human ratings (Welivita & Pu, 2023). A different question remains open, one that bears directly on the evaluation of language-using AI agents: whether AI-generated texts can approximate human texts across the discourse dimensions identified in the combined corpus. Here, human-likeness is understood not in terms of textual features taken separately or whether an individual text may sound human, but as approximation to the discourse dimensions through which loneliness becomes publicly visible in human self-disclosure within a defined digital domain. Loneliness is often treated as a private experience. Yet it is also expressed publicly in online spaces. Loneliness is classically defined as the painful discrepancy between the social relations a person has and those they need or expect (Weiss, 1973; Perlman & Peplau, 1982). Here, however, it is approached as socially mediated and publicly expressed in discourse (Ågren, 2017). The empirical object is public written self-disclosure about loneliness on Reddit, a digital domain in which personal experience becomes visible under specific conditions of circulation and response, and in which discourse is now produced by both humans and AI systems (Paveau, 2021). To investigate this issue, one human subcorpus of Reddit posts is compared with two AI-generated subcorpora produced under different prompting conditions, one from generic prompts and one from summaries of human posts. Analysis is carried out through Lexical Multidimensional Analysis, which extends Biber’s multidimensional framework by using lexical features only (Biber, 1988; Berber Sardinha & Fitzsimmons-Doolan, 2025). Across the three subcorpora, content words are extracted, keywords are identified, and the selected lexical features are submitted to factor analysis in order to identify dimensions of loneliness discourse, interpreted as latent, macro-level discursive constructs. The study asks whether AI-generated texts approximate or diverge from human texts across those dimensions.

Poster 41 | “AI can help but I become lazy to think”: Voices from Pre-Service English Language Teachers in Remote Regions

Wenssy Steva Nussy, Eugenie Mainake, Yusman Tuhulele, & Jusak Patty

Artificial Intelligence (AI) is rapidly transforming language teaching practices worldwide, and Indonesia is no exception. Merdeka Belajar curriculum emphasizes the integration of digital technology into classroom instruction (Wang et al., 2023). However, research suggests that pre-service English teachers’ digital literacy remains unevenly distributed across Indonesia (Christiani et al., 2022; Mainake & McCrocklin, 2021). Most existing studies focus on urban institutions while teacher education contexts in underdeveloped, frontier, and outermost regions largely underrepresented in the literature. Using Technology Acceptance Model (TAM; Davis, 1989), we explore 54 pre-service English teachers from two public universities located in small islands in qualitative survey (Fink, 2013) concerning their perceptions of AI tools, the perceived affordances, concerns, and plans for classroom instruction. We found that participants viewed AI as powerful tools for supporting instructional design and assessment. The concern is overreliance to the tools that contradicts a complex balance between perceived usefulness and potential pedagogical risks. We suggest strengthening AI literacy in teacher education programs to accelerate equitable participation in AI-enhanced language teaching.

Poster 42 | Exploring preservice teachers’ acceptance and perceptions of using AI to teach multilingual learners

Woongsik Choi, Soo Won Shim, & Md Mamunur Rashid Zaber

Many U.S. preservice teachers (PSTs) feel underprepared to meet the needs of multilingual learners due to limited training and resources. Despite growing interest and optimism about using AI for education, few research studies have examined PSTs’ acceptance and perceptions of AI in multilingual learner education contexts. This explanatory sequential mixed methods study examines PSTs’ acceptance and perceptions of using AI to teach multilingual learners. The research questions are: (1) What are PSTs’ levels of acceptance of AI in teaching multilingual learners? (2) How do they perceive the role of AI in teaching multilingual learners? The study was conducted at a Midwestern U.S. university. 40 PSTs pursuing ESL or bilingual endorsements completed a modified AI acceptance survey (Guo et al., 2025). Follow-up interviews are ongoing. The preliminary findings indicate moderate acceptance of AI alongside relatively low levels of AI use experience. While participants perceived AI as useful and easy to use for teaching multilingual learners, they also reported high anxiety, particularly regarding errors, overreliance, and ethical concerns. Correlation and regression analyses reveal that perceived usefulness is the strongest positive predictor of behavioral intention, while anxiety is a significant negative predictor. This suggests that PSTs view AI as a useful pedagogical tool, but also as a tool with risks. PSTs will benefit from structured, hands-on experiences focused on using evolving AI to teach multilingual learners, as well as critical discussions of errors, overreliance, and ethical concerns that can arise when using AI in this specific context. This study will provide insights into how to better support PSTs’ effective and responsible AI integration for multilingual learner education.

8-8:30am | From Fiction to Design: Shaping AI Agent Interaction through Narrative-Constrained Configurations

Walid Allioua & Ahmed Chaouki Hoadjli (TSLL26-64)

The main objective of this study is to demonstrate that representations of fictional sentient agents can be systematically transformed into design constraints that shape the linguistic behavior of real AI agents. Rather than treating fictional figures as metaphorical references, the study operationalizes them as structured interactional models embedded in a custom AI agent created using Gemini Canvas. This study addresses the following research questions: (1) How can narrative representations of fictional agents be translated into formalized interactional parameters for AI design? (2) How do such parameters influence the linguistic behavior of AI agents in comparable interactional scenarios? (3) To what extent can “human-like” qualities of AI interaction be explained through design choices rather than model-internal properties alone? Methodologically, the study adopts a design-based analytical approach. A custom AI tool is developed in Gemini Canvas in which narrative-derived dimensions such as epistemic authority, transparency, and affective expression are defined as structuring principles of the agent’s behavior. These dimensions function as design specifications that guide and constrain responses, making the agent’s fictional grounding analytically accessible. Drawing on archetypal agents such as HAL 9000 and Frankenstein’s creature, the tool is configured to produce interaction shaped by these constraints. A small corpus of prompted interactions is generated under controlled scenarios and analyzed qualitatively using concepts from interactional linguistics, focusing on stance-taking, explanation strategies, and agent positioning. Preliminary findings suggest that differences in linguistic behavior can be linked to design choices derived from fictional archetypes. The study further indicates that the perceived human-likeness of AI agents can emerge from deliberate interactional configuration rather than model-internal properties alone, and proposes a methodological bridge between humanities-based analysis and applied linguistic evaluation of AI-mediated interaction.

8-8:30am | From Tools to Teammates: Approximate Replication of Human-AI Rapport Management in Essay Writing

Fatima Iftikhar, Rida Kamal, & Sadia Irshad (TSLL26-88)

The increased use of Large Language Models (LLMs) like ChatGPT is redefining university classrooms under the emerging model of Hybrid Intelligence (HI) learning. Existing literature has been exploring their technical skill, yet, qualitative social rapport emerging from human-AI exchange, outside Western contexts remain to be fully understood. This research attempts to uncover Pakistani students’ journey from imagining AI to its real-life usage. This study attempts an approximate replication of Rokkones and Giannakos (2025) research through their qualitative framework within BS Humanities Students learning contexts from Islamabad, Pakistan. Through CASA framework and Self-Determination Theory (Chen et al., 2023), we will conduct social cognitive qualitative analyses of rapport management strategies students use when using ChatGPT to write essays (Yan, 2023). We will analyze 20 ChatGPT student chat logs collected for linguistic patterns like using politeness formulas with ChatGPT (Chen et al., 2023) and interview responses for rapport management practices students use for learning outside classrooms. The major contribution is to applied linguistics literature by looking at human-agent rapport within Hybrid learning landscapes through empirical findings from Pakistan BS Humanities Students. Practical applications from research will allow insight on how best to socialize learning with technology without losing human agency but moving toward Hybrid Intelligence success.

8-8:30am | Constructed Agents for Cultural Metaphors: A Human-AI Collaborative Chatbot for Detecting Cultural Metaphors in Bilingual Literature

Qing Ma & Xiaoqi Su (TSLL26-82)

Although the ability of Generative AI in language processing is significant, AI models still cannot understand the deeper cultural meanings in literary works (Pedersen et al., 2025), especially in the context of cultural metaphors. Cultural metaphor is a subset of conceptual metaphor, which is based on the unique experiences and traditions of a specific culture (Sharifian, 2017). In this paper, we propose a Human-AI Collaborative model, which includes a constructed agent, namely a Chatbot, capable of detecting and interpreting cultural metaphors in bilingual literature. To enhance efficiency and scalability, our method adopts a semi-automated approach. First, we use Python to systematically extract the potential cultural metaphors list. This process enables us to obtain three lists of “X is Y”, signaling markers (Goatly, 1997), and potential cultural keywords. Then, we feed our AI chatbot with cultural-sensitive prompts. Specifically, we utilize the Metaphor Identification Procedure (MIP) (Pragglejaz Group, 2007) to detect contextual deviations from the literal meaning, followed by strict three-standard “cultural filter”, including rhetorical, cognitive, and cultural dimensions. Subsequently, human experts will evaluate and refine the results of this AI agent, thereby bridging the gap between computational efficiency and humanistic cultural sensitivity. This proposal is specifically centered around the interdisciplinary theme of the 2026 TSLL conference, demonstrating how the results of AI agent are able to identify subtle cultural meanings to promote cross-cultural communication. Ultimately, this study highlights how purposefully constructed AI agents can transform interdisciplinary language research.

8:30-9am | Toward a Fluency-Adaptive Spoken Dialogue Agent: A Pause-Based Proficiency Estimation Framework for Language Learning

Yichen Jia (TSLL26-114)

Conventional AI spoken dialogue systems rely on speech transcription as their primary input. While such systems can adjust response content based on what a speaker says, they cannot adapt their delivery register to how a speaker sounds. However, delivery is a core dimension of speaking proficiency and a central concern in spoken language assessment and instruction. Research on interactional approaches to language development suggests that learners benefit most when their interlocutor is attuned to their current proficiency level — with input that is comprehensible yet slightly beyond their current ability. An AI interlocutor that fails to adapt its delivery risks undermining the very interaction it is designed to support. This study presents a framework for a fluency-adaptive spoken dialogue agent that narrows this gap by grounding adaptive output in empirically motivated fluency assessment. Drawing on temporal fluency research and corpus-based descriptions of pause patterns across proficiency levels, the system estimates speaker proficiency from acoustic features — specifically silent and filled pause frequency, type, and syntactic position — extracted via automated Praat and Python analysis. Proficiency estimates are mapped to three levels (low, intermediate, advanced) informed by published corpus norms, bypassing the need for novel data collection. These estimates are passed as contextual parameters to a large language model, which generates responses calibrated in fluency features for text-to-speech output. To evaluate the framework, this study presents a pilot demonstration using publicly available oral proficiency samples at three levels. The agent’s responses are analyzed for alignment between estimated proficiency and output register, examining whether lower-proficiency input systematically elicits simpler, more pause-tolerant responses. Findings are discussed in terms of implications for AI-supported oral language development and intelligent tutoring systems. While the current framework centers on fluency, the architecture is designed to accommodate broader delivery dimensions — including lexical and syntactic register — as a direction for future development.

8:30-9am | Constructed Agents for L2 Writing Assessment: Evaluating Generative AI as Rubric-Guided Scoring and Feedback Systems

Agnieszka Leńko-Szymańska (TSLL26-124)

This talk examines AI-based assessment of L2 writing through the lens of constructed agents. Rather than treating large language models as fixed scoring tools, it approaches them as evaluative agents whose behaviour is shaped by prompts, task context, scoring criteria, and examples. From this perspective, automated writing assessment is not simply about model performance, but about how evaluative behaviour is produced through design choices and alignment with human scoring standards. This framing is especially relevant as generative AI makes assessment tools more widely available in educational settings. The study draws on a random sample of 500 essays written under exam conditions in response to five different prompts and rated by two human assessors each. The essays were scored analytically on four criteria: content, organisation/style/cohesion, grammatical accuracy, and lexical range. They were then evaluated by five generative AI systems, each configured as a constructed scoring agent under three prompting conditions: zero-shot, one-shot, and ten-shot. In all three conditions, the systems were given the task context and scoring rubric; in the one-shot and ten-shot conditions, exemplar responses were also provided to help shape their evaluative behaviour. The analysis combines quantitative comparisons of human and AI scores with a qualitative examination of the scoring justifications produced by both human raters and the AI systems. The findings suggest that generative AI can function as a constructed evaluative agent and approximate human scoring patterns to a meaningful degree. At the same time, convergence across models raises an important question: whether consistency reflects alignment with the construct being assessed or shared biases introduced by prompts, rubric interpretation, or exemplars. The talk argues that the central issue is not whether AI can grade essays, but how such agents are constructed, constrained, and critically evaluated in L2 writing assessment.

8:30-9am | Morphological Flattening in LLM-Mediated Text Simplification: A Corpus-Based Analysis of Verb Form Attrition During CEFR-Targeted Adaptation

Norazha Paiman (TSLL26-26)

The increasing deployment of Large Language Models (LLMs) for CEFR-targeted text simplification presupposes a uniform reduction of linguistic complexity; however, this study reveals a critical misalignment between global text levelling and local morphological integrity, operationalised here as “morphological flattening”, or also known as a systemic reduction in verb form diversity that disproportionately outpaces broader complexity adjustments. By juxtaposing a granular corpus-based analysis of 304 verb tokens from original academic passages against their LLM-adapted counterparts, the study provides empirical evidence of a fundamental decoupling between lexical/syntactic adaptation and morphological fidelity. The data indicate a critical failure in proficiency scaling: while general text complexity was successfully modulated to the B2 band, verb morphology suffered a catastrophic collapse to an A1.7–B1.0 baseline, representing a 2.1-band downward overshoot that strips texts of the functional nuances required for upper-intermediate competence. This attrition exhibits a structural asymmetry: marked Tense-Aspect-Modality (TAM) forms underwent radical excision including significant reductions in perfect aspect (-71.2%), progressive constructions (-70.0%), and epistemic modality (-71.4%), while simple present and past forms saw compensatory inflation (+85.7% and +30.9%, respectively). Statistical analysis confirms these distributional shifts are highly significant (χ²(11) = 127.4, p 0.80) for multiple categories, evidencing a systematic algorithmic bias rather than stochastic variation. To account for these distortions, the study proposes the Morphological Preservation Hypothesis (MPH), arguing that simplification algorithms differentially preserve features based on acquisition order, cross-linguistic markedness, and semantic transparency. Consequently, the study also demonstrates that standard prompting strategies are insufficient for maintaining construct validity and detail four prompt-engineering interventions designed to mitigate this flattening; this is to ensure that automated adaptation maintains the Tense-Aspect-Modality (TAM) profiles requisite for valid intermediate-to-advanced assessment.

9-9:30am | When Students Talk to AI: Inquiry, Language Learning, and Critical Thinking

Marcela Cecilia Danowski (TSLL26-15)

The rapid emergence of conversational artificial intelligence has introduced new forms of interaction between language learners and digital agents. While AI tools are increasingly used to support language practice, less attention has been given to how learners engage critically with these agents through dialogue and inquiry. This presentation explores how structured questioning and inquiry-based learning can transform interactions with AI from simple language practice into opportunities for deeper thinking and reflection. Drawing on classroom experiences in English language teaching, the presentation examines how students can use conversational AI agents as partners for inquiry. By applying questioning strategies and thinking routines, learners engage in dialogues with AI that encourage them to clarify ideas, evaluate responses, and negotiate meaning. Rather than accepting AI-generated answers passively, students are guided to interrogate the responses, compare perspectives, and refine their questions. The session situates these practices within broader discussions about how humans understand and interact with non-human agents capable of linguistic communication. Just as representations of sentient agents in literature and film have historically shaped perceptions of intelligent non-human entities, today’s AI language tools invite new forms of interaction that challenge traditional classroom roles and expectations. Examples of classroom tasks will illustrate how learners can engage with AI agents to explore complex topics, reflect on language use, and develop critical awareness of AI-generated discourse. Particular attention will be given to the role of questioning, metacognitive reflection, and collaborative discussion following AI interactions. The presentation argues that integrating inquiry-based approaches into AI-mediated language learning can help students develop not only linguistic competence but also critical thinking and digital literacy. Such practices prepare learners to engage thoughtfully with increasingly sophisticated language-using AI agents in educational and professional contexts.

9-9:30am | An intent-based agent for young learners’ intercultural communicative competence? A feasibility study

Elizabeth Bear (TSLL26-150)

Tailoring a conversational agent to specific learner populations and cultural learning goals, such as the knowledge, skills, and attitudes of intercultural communicative competence (ICC; Byram, 2020), has traditionally required hand-crafted dialogue interactions (Johnson & Valente, 2009). While less flexible than large-language-model (LLM)-based systems, intent-based systems may have a role for task-oriented dialogues with constrained scopes (see Timpe-Laughlin et al., 2025, for discussion). In addition, LLM-based systems require extra care for young learners (UNICEF, 2025), and a lack of linguistic and cultural authenticity may pose challenges for ICC development (Godwin-Jones, 2024). This feasibility study examined whether an intent-based agent designed specifically for the school context can foster young learners’ ICC. We designed tasks to align with selected A1-B1 descriptors from the Common European Framework of Reference for Languages (CEFR)’s pluricultural competence scale (Council of Europe, 2020, p. 125). Drawing upon the instructional approach of explicit instruction plus production practice (Taguchi, 2015), we presented a cultural difference in a pre-task video and integrated the cultural difference into the role-play with the agent. ICC development was assessed through self-reflection frameworks from computer-mediated communication research (Tecedor & Vasseur, 2020; O’Dowd, 2021). Five ICC tasks were tested with 43 students in the seventh (age 12-13) and eighth (age 13-14) grade. Both grade levels rated the tasks positively and exhibited ICC development in their self-reflections. However, seventh grade students discussed a wider range of themes, especially for knowledge and skills, and were significantly more likely to continue to use the agent. The grade levels were more similar in their likelihood to recommend, suggesting that eighth graders saw value for other students. Overall, our results show that an intent-based agent can (1) enhance young learners’ ICC when the task design matches learner needs and (2) be especially beneficial for young learners at lower proficiency levels.

9-9:30am | Constructed Agency in AI-Mediated ESL Writing: A Socio-Pragmatic Analysis of Interactional Competence

Jyothsna Bethala & M. Udaya (TSLL26-117)

As generative computational systems become increasingly embedded in English Language Teaching, learners are engaging with these interfaces as socially responsive entities. This shift necessitates a critical examination of how meaning is constructed and stance is negotiated within automated environments. While existing literature primarily addresses technical affordances or lexical outcomes, the socio-pragmatic dimensions of learner-system interaction remain under-theorized in applied linguistics. This study investigates how undergraduate ESL learners demonstrate interactional competence during the collaborative construction of academic texts. Adopting a qualitatively driven mixed-methods design, the research utilizes a “Trace & Talk” (log-based stimulated recall) methodology and Interpretive Phenomenological Analysis (IPA). The dataset comprises interactional logs and retrospective interview data from 20 participants. Findings indicate that learners consistently employ “social signaling” including politeness strategies and conversational repair reflecting an orientation toward the system as a communicative interlocutor rather than a neutral tool. This perception facilitates a low-stakes environment that mitigates writing anxiety and fosters learner agency. Despite these benefits, the anthropomorphic nature of the system’s responses can influence a learner’s authorial autonomy, potentially compromising the authenticity of the final text. By foregrounding these interactional dimensions, this study contributes to an emerging framework of “interactional pragmatic literacy.” The findings suggest that pedagogical integration must prepare learners to navigate increasingly agentive forms of communication while maintaining authorial autonomy.

10:30-11am | Pedagogical implications of synchronous spoken communication with an AI agent

Christopher Litten, Young-Joo Lee, Ahmad Zubaidi Amrullah, Benjamin Godard, & Mahdi Duris (TSLL26-142)

As conversational AI becomes increasingly embedded in everyday life, it is also being taken up as a tool for language learning. In prior research on AI-mediated communication, most work has focused on text-based systems or asynchronous spoken systems, leaving a critical gap in our understanding of how learners engage with fully synchronous, low-latency spoken AI systems that closely resemble human interaction. This gap is pedagogically relevant to interactional dynamics and how they differ from human–human communication, informing about the assumptions underlying speaking practice, fluency development, and assessment may not transfer to AI-mediated contexts. The study addresses this gap by examining how synchronous AI interactions shape both spoken fluency and speaker anxiety. We analyze speech from 11 English speakers completing three structured speaking tasks (descriptive and argumentative) with Sesame.ai, a real-time conversational agent. To provide a comprehensive profile of AI-mediated performance, the study triangulates acoustic measures of fluency (e.g., pause frequency, mean length of run), fluency evaluations from 11 human raters, and the speakers’ self-reported measures of speaking anxiety and technology familiarity. By treating the language-using AI agent as a real speaking partner, the findings reveal how affective responses correlate with both objective fluency markers and listener perceptions. The pedagogical implications further highlight how synchronous AI can effectively support speaking practice, identify which task types minimize learner anxiety, and propose future research recommendations about whether AI-mediated performance reflects natural human interaction.

10:30-11am | To Listen Is to Inherit: From Grandmothers to Generative Models

Sahara Al-Madi (TSLL26-20)

My grandmothers taught me that to listen is to inherit. My Palestinian grandmother carried the dialect of a village displaced. My Mexican grandmother carried ninety-three years through generations, mining towns, alongside the metals now in the devices we hold. A story is a living thing. To truly listen is to become its new keeper. I carry this into my work as a computational linguist studying how AI processes the voices it hears. Human storytelling has long imagined encounters with non-human intelligence, from Mary Shelley’s Creature, Octavia Butler’s Lilith’s Brood to Tolkien’s Ents who carry the memory of forests. These stories are preparations that give us language for asking what we want from the intelligences we build. I lead error analysis examining how large language models process dialectal Arabic. Models fail in patterns: they flatten variation, miss cultural context, and sometimes get the right answer for the wrong reasons as seen in the Clever Hans effect. These failures tell us where our tools need better teachers, richer data, and deeper questions. This paper brings together three ways of knowing: storytellers who imagine non-human intelligence, linguistic analysis that reveals where models struggle, and the question of what we owe the voices we ask machines to learn from. From chatbots to writing assistants, the AI tools students use today are training them for a world of human-AI collaboration. My 300+ multilingual students taught me to ask what comes next: Could an AI carry memory the way an Ent does? Building better AI requires listening across disciplines. This is an invitation to ask better questions together, informed by what humans have always known about listening, carrying, and care. To listen is to inherit. What we build next, we will build from what we were willing to hold.

10:30-11am | AI as a teacher: discourse differences between AI and human classroom interaction

Marilisa Shimazumi, Randi Reppen, & Tony Berber Sardinha (TSLL26-134)

This paper looks at how a large language model (GPT-5) performs the social and affective work of teaching in higher education, in comparison with real classroom interaction. LLMs have been reported to take on professional roles such as teaching (Kasneci et al., 2023), but the language they use in doing so remains unclear. The study uses two corpora: one of face-to-face university classes and another of online classes. These data were paired with AI-generated teacher talk produced from summaries of real classroom teacher turns (student turns were retained from the original corpus and were not generated by the AI). The resulting corpus contains 30+ million words of human and AI teaching discourse. The analysis uses Lexical Multi-Dimensional Analysis (Berber Sardinha & Fitzsimmons-Doolan, 2025) to identify teaching discourses. The dimensions for the online classes include contrasts such as macro-structuring versus performative stagecraft and informal facilitation versus formal deductive exposition. The results show differences between human and AI teacher talk. The LLM exaggerates or introduces traits not present in the source interaction. These include, for example, evaluative language, sarcastic remarks, or exaggerated expressions of appreciation. Additionally, the LLM-simulated classes generally did not differentiate teaching personas, which suggests difficulty in simulating distinct teaching styles within the same professional role. In conclusion, the study shows that LLMs can reproduce recognizable aspects of teaching discourse but do so in ways that differ from human interaction. These findings contribute to the understanding of how AI performs institutional roles and point to the need for careful assessment of AI-generated language in educational contexts.

11am-Noon | [Plenary] AI chatbots are neither sentient nor authentic (linguistically, culturally, emotionally): Does that matter for language learning?

Robert Godwin-Jones

Despite warnings from leading AI pioneers and engineers, AI systems have not (yet) achieved consciousness nor shown signs of an intent to eliminate the human race. What they have demonstrated, however, is an uncanny ability to simulate human communicative abilities so well that they serve not only as powerful writing assistants, but also as learning partners, social companions, and even therapists. For language learning, potential AI sentience is less important than its ability to serve convincingly as a viable conversationalist and tutor. Studies have demonstrated, in fact, how beneficial AI chatbots can be in supplying writing/speaking practice in the target language while providing supportive feedback in a non-threatening environment.

Indeed, AI bots’ extraordinary language abilities have led to claims that they “facilitate an authentic, interactional language learning environment” (Chiu et al., 2023); or that “AI has the potential to aid learners in cultivating emotional resilience and empathy, crucial competencies for proficient cross-cultural communication” (Devaki & Mangayarkarasi, 2024). These (and others’) assertions signal that AI output is linguistically and culturally authentic enough to supply appropriate learning content, even substitute for human interlocutors, in acquiring interactional skills, developing intercultural competence, and receiving effective emotional support. However, that view ignores the statistical process used by AI systems to reproduce language and the limitations of that process for the linguistic features, cultural content, and emotional weight of its output. That has consequences for its ultimate effectiveness in language learning, as will be discussed in this presentation.

1-1:30pm | How do generative AI agents and human learners interact in oral role-play assessment tasks?

Gi Jung Kim (TSLL26-141)

With the rapid development of generative AI, the potential of AI for performance-based pragmatics assessment, given its capacity for real-time, back-and-forth exchange with learners (Van Moere & Downey, 2016), is closer to realization than ever. However, little has been studied about the interaction between generative AI and human learners in the context of oral communication assessments. To address this gap, the current study draws on the conversation analytic perspective to explore how young EFL learners and generative AI agents interact during role-play-based dynamic pragmatics assessment tasks. To this end, the study will use role-play performance data collected from four Korean middle school students who interacted with AI agents in Talk Smart Trainer, a ChatGPT-powered dynamic pragmatics assessment system. These students completed two request role-play tasks, each consisting of four sub-tasks. Five samples will be randomly selected per sub-task type, resulting in 20 samples. These samples will be analyzed turn-by-turn to identify key pragmatic and interactional features in each sub-task type, including pragmatic functions, request strategies (Blum-Kulka et al., 1989), internal and external request modification (Economidous-Kogetsidis, 2010), refusal strategies (Beebe et al., 1990), sequence organization, and conversation management. These features will be discussed in terms of how they contribute to AI agents constructing and enacting two personas – namely, a conversation partner and a feedback provider. This analysis will also be complemented by interview data collected from the students. The findings will provide a more robust basis for evaluating the effectiveness of using generative AI to standardize interlocutors and provide feedback during interactive oral communication assessment tasks.

1:30-2pm | Conversational AI in Language Classrooms: Enhancing Communicative Competence among Engineering Students

Momin Ali (TSLL26-18)

The integration of artificial intelligence (AI) into educational contexts has opened new possibilities for language learning and classroom interaction. Conversational AI agents, capable of generating real-time responses and engaging in dialogue, have increasingly become accessible tools for supporting language practice. This paper investigates the role of conversational AI in developing communicative competence among undergraduate engineering students studying English as a second language. The study was conducted in an English Language Communication Skills (ELCS) course at a technical institute in India. AI-supported conversational tools were incorporated into classroom activities such as simulated interviews, guided discussions, and writing assistance tasks. These activities allowed students to interact with AI agents as conversational partners while practicing linguistic structures, vocabulary use, and pragmatic communication strategies. Data were collected through classroom observations, reflective student responses, and brief feedback surveys to understand learner engagement and perceptions of AI-mediated interaction. The findings suggest that conversational AI can create an interactive and low-anxiety learning environment where students feel more confident practicing language skills. Learners reported that the availability of immediate responses and opportunities for repeated interaction encouraged experimentation with language and supported independent learning. At the same time, the study emphasizes the need for pedagogical guidance to help students critically evaluate AI-generated content and develop responsible approaches to technology use in educational settings. The paper argues that conversational AI can serve as a complementary pedagogical tool that extends opportunities for communication practice beyond traditional classroom interaction. By integrating AI agents into language pedagogy, educators may support greater learner engagement, communicative confidence, and innovative approaches to language teaching in applied linguistics.

1:30-2pm | Rewriting Archie Bunker: Discourse shifts introduced by LLM-based inclusive language adaptation

Mirella Whiteman & Tony Berber Sardinha (TSLL26-53)

Inclusive language guidelines generally treat textual adaptation as lexical substitution, replacing dispreferred forms with recommended alternatives (Ferguson & Bellamy, 2022). Since LLMs include progressive positions (Motoki et al., 2024), they should be a viable tool for inclusive language adaptation. However, lexical change may also introduce discourse shifts. What remains unclear is which discourses emerge from such interventions, how they compare to human-authored originals, and whether three leading LLMs differ in this respect. Our study looks at how LLMs adapt the language of All In The Family, a US sitcom (1971–1979) whose protagonist, Archie Bunker, is widely recognized for expressing racist, sexist, xenophobic, and class-based stereotypes. We built a corpus of transcripts from Seasons 1–4 and 7–9 (149 episodes; 513,614 words), identified 494 biased segments, and instructed GPT-5.1, Grok-4, and Llama-3 to rewrite each segment by removing the identified harmful language. A Lexical Multidimensional Analysis (Berber Sardinha & Fitzsimmons-Doolan, 2025) was conducted on the original and adapted scripts, yielding seven lexical dimensions. For example, Dimension 2 distinguishes comic discourse (human) from reflective discourse (LLMs). Dimension 3 captures a clash between offensive discourse (human) and social justice discourse (LLMs). Dimension 4 opposes insult-driven discourse (human) against warm domestic discourse (LLMs). Across all three models, results reveal consistent dimensional oppositions, producing roughly similar patterns in which social justice and progressive discourse replace the original. In the paper presentation, we will discuss and illustrate all dimensions and discourse shifts.

2-3pm | [Panel] What Constructed Agents Sound Like: Language, Pragmatics, and Culture in Human-AI Interaction

Jon Chun, Katherine Elkins, Marisol Hernandez-Britto, Nicolas Pechon-Elkins, & Melissa Vu (TSLL26-62)

Before users interact with real AI agents, they have already learned what AI should sound like. Literary and cinematic representations of artificial beings, from Ishiguro’s Klara to Samantha in Her, have taught audiences to associate specific linguistic features with machine intelligence to indicate trustworthiness or emotional depth. Constructed agents shape the expectations that users, including language learners, bring to real AI systems. But how well do those expectations hold when AI encounters the full pragmatic complexity of real human languages? And when AI systems construct their own agents, whose cultural models do they draw on?

This panel investigates those questions through three studies that test the boundaries of AI agents’ linguistic and cultural competence. By examining AI performance across a historical literary language, a contemporary narrative-generation task, and an under-resourced language with complex social deixis, the panel maps the boundaries of AI linguistic competence across conditions rarely tested in current evaluation research. Each paper takes a different approach to a shared problem: the gap, or surprising lack of one, between what fictional agents do with language and what real AI systems can handle. We ask open questions about where AI pragmatic and narrative competence actually stands, what its limits reveal about language, and what the findings imply for AI systems used in language learning and cross-cultural communication.

The panel is organized by Jon Chun (Founding Director, Human-Centered AI Lab) and Katherine Elkins (Professor, Kenyon), and features three presentations by Lab Research Fellows followed by a moderated discussion. Marisol Hernandez Brito, “Constructing the Agent in Spanish: How AI Models Imagine Artificial Intelligence in Latin American Contexts,” asks what kind of AI agent current language models construct when prompted to generate scenes of human-AI interaction set in Latin American Spanish-language contexts. Drawing on methods developed in prior cross-model narrative research, Hernandez Brito examines whether AI-constructed agents default to Anglo-American archetypes (the seductive companion, the omniscient assistant, the existential threat) or whether they produce agents with culturally distinct voices and roles. Her findings speak to questions of multilingual AI literacy and the cultural preconditions under which constructed agents become legible. Nicolas Pechon-Elkins, “What AI Makes of Sancho Panza: Pragmatic Competence and Historical Register in AI Language Processing,” tests what happens when contemporary AI systems encounter linguistically complex and historically situated Spanish. Pechon-Elkins examines passages from Cervantes’s Don Quijote that feature pragmatic competency like irony, register shifts between characters, indirect speech acts, and politeness conventions that signal social hierarchy, and analyzes whether current AI agents perform as well compared to standard contemporary Spanish. Using corpus-based analysis of the outputs, he outlines how the systems handle stance, hedging, epistemic marking, facework, and register variation. The presentation addresses which dimensions of pragmatic competence AI agents handle well and which they distort, and what those patterns reveal about the expectations that fictional constructed agents create for real AI interlocutors. Melissa Vu, “What AI Calls You: Social Deixis, Pronoun Systems, and the Limits of AI Agents in Vietnamese,” examines how current AI systems handle the socially complex pronoun and address system of Vietnamese, a language in which pronoun choice encodes age, gender, kinship, intimacy, and social hierarchy simultaneously. The paper tests AI agents on Vietnamese conversational scenarios that require pragmatically appropriate pronoun selection and analyzes how the systems navigate a deixis system .

3:30-4pm | The Effects of Avatar-Based Interfaces on Speech in Spoken Generative AI Interaction

Seiji Takahashi (TSLL26-70)

Spoken generative AI (GenAI) chatbots are promising tools for speaking practice and assessment in language learning. One concern is that people speaking to non-human systems often exhibit distinct phonetic characteristics compared to speech directed at humans (i.e., machine-directed speech), which may may lead learners to internalize atypical pronunciation patterns or influence assessment outcomes. The visual interface of spoken GenAI systems featuring human-like avatars may offer a more human-like interaction experience. However, little research has examined how the presence or absence of an avatar affects human speech directed at GenAI. To examine the potential benefits of avatar-based visual interfaces for more human-like communication, this study investigated whether differences in visual interface (a human-like, lip-syncing avatar vs. a conventional display, using the same system) influence speech in ways that are perceptible to listeners, as reflected in judgments of machine-directedness (i.e., how much speech sounds “addressed to a machine”) and perceived engagement (i.e., how engaged the speaker sounds). To generate speech samples, four native-English speakers interacted with the GenAI under two conditions (avatar vs. conventional) and with a human interlocutor (baseline condition). A total of 72 utterance clips were extracted and evaluated by 44 native English-speaking listeners for machine-directedness and engagement. Linear mixed-effects modeling revealed no significant differences in perceived machine-directedness across the three conditions. For engagement, speech directed to both AI systems was rated similarly, while speech directed to a human received higher ratings. Consistent with these findings, acoustic analyses showed no differences between the two AI conditions across all measures. In contrast, human-directed speech exhibited higher mean F0 and greater intensity—features commonly associated with increased engagement—than AI-directed speech. These findings suggest that while spoken GenAI systems may not elicit overtly machine-directed speech, they may still fall short of eliciting the level of engagement typically observed in human–human interaction.

3:30-4pm | Metaphor in learner writing: discourse variation and limits of AI detection

Amanda Chiarelo Boldarine (TSLL26-63)

Metaphor in learner writing: discourse variation and limits of AI detection Metaphor has been described as a dynamic, discourse-bound process that unfolds in use and contributes to the mediation of thinking, learning, affect, and action (Cameron, 2003). This perspective entails that metaphor must be studied in actual discourse, where patterns are emergent and context-dependent. However, corpus-based studies of metaphor have been constrained by the need for manual annotation, given the lack of reliable automated detection. With the advent of LLMs, it becomes possible to test the extent to which AI systems can identify metaphor in discourse. This study looks at metaphor use in a learner corpus of 450 texts (63,297 words) produced by students aged 11–16. To operationalize the discourse-bound nature of metaphor, we first identified discourse patterns through Lexical Multidimensional Analysis (Berber Sardinha & Fitzsimmons-Doolan, 2025), which yielded eight contrasting lexical dimensions (e.g., abstract, scientific knowledge vs. personal narratives and emotional experience). Metaphor candidates were drawn from the lexical loadings of each dimension, and metaphorical uses were manually identified using the Metaphor Identification Procedure (Pragglejaz Group, 2007) to serve as a basis for comparison. Manual annotation identified 245 metaphor tokens (37% of candidates), revealing the extent of metaphor presence in discourse dimensions. Metaphor use was found to be selective and discourse-dependent, occurring more frequently in abstract, academic discourses and less so in narrative or experiential ones. AI performance was poor, retrieving at best 25% of the metaphors. The findings show that metaphor is tied to discourse variation in learner writing, while AI systems fail to retrieve metaphors reliably.

3:30-4pm | A Systematic Review of Generative AI Literacy in Language Education (2023–2025)

Shuhui Yin, Mahdi Duris, Haley Greer, Gi Jung Kim, & Xiaodan Liu (TSLL26-127)

As generative AI (GenAI) tools increasingly reshape educational practices and communicative norms, GenAI literacy has become a key concern for language learners, educators, and researchers (e.g., Ng et al., 2024; Xu et al., 2025; Wang & Wang, 2025). Despite growing scholarly attention to GenAI literacy, the ways in which AI literacy is integrated into research design in language education remain underexplored. Systematic inquiry is therefore needed to understand how AI literacy is conceptualized, assessed, and positioned in the study. Aligned with the TSLL conference theme, this study contributes to understanding how research and teaching in language education need to evolve to better engage with AI agents by conducting a systematic review of empirical articles published between 2023 and 2025. Specifically, this study is guided by two research questions: How has GenAI literacy been studied in language education research over the past three years (2023-2025)? What outcomes have been reported? Following PRISMA guidelines (Page et al., 2021), this study searched four databases in linguistics and education (i.e., LLBA, PsycINFO, ERIC, and CMMC) using the keyword “AI literacy.” From an initial pool of 593 articles, 59 empirical research articles met the inclusion criteria and were retained for full coding and thematic analysis. Preliminary findings reveal four primary strands of research: (a) systematic reviews of AI literacy, (b) development and validation of assessment instruments, (c) design and evaluation of instructional interventions, and (d) GenAI literacy as a mediating variable in relation to outcomes such as learning attainment, language anxiety, and learner engagement. Across studies, conceptualization of GenAI literacy vary across educational contexts but consistently foreground core components such as technical knowledge and critical evaluation. Study participants span language learners, teachers, and researchers. Implications for both research design and pedagogical practices in language education will be discussed.

4-4:30pm | Engaging with AI in the Pocket: Language Learners’ User Experience (UX) of AI-Powered Features in MALL Apps

Melike Akay (TSLL26-136)

As artificial intelligence (AI) becomes increasingly embedded in mobile-assisted language learning (MALL) applications, understanding how learners perceive and interact with these AI-driven features is essential for both pedagogical design and L2 assessment. While commercial language learning apps (e.g., Duolingo, Babbel, and Busuu) have long supported autonomous language learning, their recent integration of AI agents, offering personalized feedback, adaptive practice, and conversational scaffolding, marks a significant transition in how learners engage with language tools. However, UX on these AI features, particularly in relation to L2 skill development and intercultural learning, remain underexplored. This study examines how undergraduate language learners perceive and engage with AI-powered features within MALL apps, with a focus on their evaluations across L2 skills and intercultural competence. Drawing on a mixed-methods sequential explanatory design, the study triangulates data from an online questionnaire (N =185) and semi-structured interviews (N = 10) with university students who reflected on their experiences including new AI features embedded in MALL apps. Findings reveal a productive tension in learners’ perceptions: participants broadly welcomed AI features as less intimidating and more conducive to self-directed learning, appreciating the low-stakes environment AI agents create for practice. At the same time, they reported that AI-mediated interaction dehumanizes the language learning experience, stripping away the authentic human exchange and cultural negotiation they associate with meaningful L2 communication and intercultural engagement. These findings tend to complicate celebratory narratives around AI integration in MALL, suggesting that learner-AI interaction trades interactional authenticity for psychological comfort. Implications are discussed for app developers and language educators seeking to design AI features that balance learner autonomy with the irreducibly human dimensions of L2 and intercultural learning.

4-4:30pm | GenAI Personas and Critical AI Literacy Among English Language Educators

Sibel Söğüt, Duong Nguyen, Danson Kimathi, Esma Masca, & Febriana Grundy (TSLL26-57)

Language educators increasingly use GenAI to create lesson plans (Godwin-Jones, 2024), instructional resources (Kohnke et al. 2023), personalize learning through corrective feedback (Creely, 2024) and assign roles to GenAI tools in their practices. They use GenAI as a collaborator and contributor and make pedagogical decisions for learning and teaching purposes (Barnes & Tour, 2026). In this way, they create GenAI teacher personas with human-like characteristics including communication style, contextual awareness, empathy (Aad & Hardey, 2025). This study aims to examine personas pre-service and in-service English language teachers assign to GenAI tools (e.g., collaborator, co-author, expert) when these tools are used in language teaching and assessment. Specifically, it addresses these two research questions: 1. What personas do English language educators assign to GenAI tools when designing lesson plans, instructional materials, and assessment instruments? 2. What do these personas assigned to GenAI agents reveal about language educators’ GenAI literacy? A convergent parallel mixed-methods design (Creswell & Plano-Clark, 2011) is adopted so that quantitative and qualitative data are triangulated at the interpretation stage. Specifically, data is collected from 30 pre-service and in-service language educators across contexts in Kenya (n = 10), Türkiye (n = 10), and Vietnam (n = 10) through a questionnaire consisting of questions about demographics, teaching background, and roles assigned to GenAI in language educational activities, and through semi-structured interviews. The questions in the both instruments address characteristics attributed to GenAI, including expertise level, personality traits, experience profile, teaching philosophy, emotional qualities, and cultural background. Quantitative data from the survey is analyzed using descriptive statistics and frequency distributions to identify patterns, while qualitative data from the semi-structured interview is analyzed using open coding and thematic analysis (Brown & Clarke, 2006). This study contributes to a critical understanding of teacher-AI interaction by conceptualizing how educators position GenAI tools through socially constructed personas. It offers an empirically grounded typology of GenAI teacher personas across contexts and how these reflect varying levels of GenAI literacy, particularly in recognizing AI affordances, limitations, and ethical considerations. Importantly, the findings highlight that persona construction operates as an implicit form of prompt engineering, where educators encode expectations about expertise, tone, and pedagogy into their prompts, shaping both AI outputs and patterns of human-AI collaboration. We discuss implications for teacher education and professional development to improve teacher-AI collaboration and improve critical AI literacy.

4:30-5pm | Reimagining AI Language Agents: Inclusive Design for Early Learners at the Margins

Farida Ladipo-Ajayi (TSLL26-67)

WALL-E, Baymax, and our favourite fictional robots inspired a vision of the ideal language companion: tireless, non-judgemental, and responsive. Yet for children in the early years, particularly those navigating multilingualism, communication challenges, or under-resourced learning environments, these imagine agents bear little resemblance to the AI tools currently available to them. This paper critically examines the gap. It uses the inclusive, adaptive qualities modelled by fictional AI companions as a humanistic benchmark, against which it evaluates Google Read Along as a case study (Ismail & Ismail, 2025). It interrogates the assumption embedded in widely used AI language tools: who they are designed for, whose linguistic repertoires they recognise, and whose they do not. The paper argues that current AI language agents replicate a one-size-fits-all model that marginalises the very learners who stand to benefit most, particularly in underresourced and multilingual contexts where inequities in technology access and outcomes are well documented (Warschauer & Matuchniak, 2010). Grounded in Universal Design for Learning principles and a learner-centred perspective, the paper proposes a reframing: inclusion not as an afterthought but as a foundational design philosophy. It draws on the concept of texts and tools as mirrors (Heineke et al., 2022), asking what AI language agents for early learners would look like if built around learner variability, communication mode, pace, and cultural identity, from the outset. The paper concludes by sketching principles for inclusive AI language agents, arguing that the fictional companions we imagined for children were always more inclusive than the tools we built and that closing that gap is where true educational equity begins.

4:30-5pm | LLM-powered multi-agent system for personalized language learning: Insights and lessons learned from a longitudinal project

Naiyi Xie Fincham & Aitor Arronte Alvarez (TSLL26-78)

As AI conversational agents become increasingly embedded in general education, understanding how learners engage with them and how design shapes that engagement is urgent. This presentation reports a design-based research project that develops and iteratively refines a multi-agent system using large language models (LLMs) to support personalized second language (L2) learning. Over the course of the project, two system designs have been developed and tested. The first was a single agent that facilitated L2 dialogue around topics and proficiency levels aligned with course syllabi, while providing multi-level, real-time analysis of learner output benchmarked against established proficiency guidelines. To better distribute these functions and improve both system performance and learning outcomes, this design then evolved into a multi-agent system: a conversational practice agent enhanced with a self-regulated learning component that supervises and guides learners through planning, progress monitoring, and building awareness of their strengths and weaknesses, and a dedicated feedback agent responsible for real-time output analysis. Our presentation foregrounds the design thinking behind these systems. We discuss how predetermined syllabi and proficiency frameworks were used to structure conversation flow, including topic selection, turn management, and feedback, and how theories of self-regulated learning guided the agent’s programming. We also examine how evolving analysis of learner-AI interaction patterns through a longitudinal study in undergraduate EFL courses drove successive revisions to the system, offering insights into how learners respond to and make sense of AI agents in instructional contexts. This work contributes to the conference theme by surfacing the characteristics of AI agents that language learners recognize, resist, or rely on during repeated interactions. It showcases how research in second language acquisition, L2 pedagogy, and self-regulated learning can inform the design of multi-agent language tutoring systems. It offers practical guidelines for AI-powered language tutoring responsive to specific instructional contexts and course objectives.

4:30-5pm | Thinking Across Languages and Machines: Translanguaging as Computational Practice in Robotics and AI Literacy Learning

Vahid Mousavi & Qi Si (TSLL26-103)

This study investigates how bilingual elementary students engage in translanguaging across robotics-based coding and AI literacy learning activities. Grounded in sociocultural theory and a translanguaging perspective, the study conceptualizes language practices as dynamic resources that mediate participation, problem-solving, and meaning-making in technology-rich environments. While prior research has examined translanguaging in literacy classrooms, less is known about how multilingual learners mobilize their linguistic repertoires across computational and emerging AI-related contexts. Using a collective case study design, five Spanish–English bilingual students in grades 3–4 participated in a sequence of learning activities that included (a) an AI literacy instructional component in which students were introduced to foundational AI concepts through a researcher-designed multimodal digital booklet (Zappy the AI Robot), (b) collaborative VEX robotics coding tasks, and (c) a post-activity survey capturing students’ perceptions of AI and their learning experiences. Data sources included screen recordings, audio-recorded peer interactions, classroom observations, field notes, and survey responses. Interactional and thematic analyses were conducted to examine how language use aligned with task demands and computational thinking practices. Findings indicate that translanguaging functioned as a task-sensitive and context-dependent resource. English predominated during interaction with coding interfaces, while Spanish supported peer collaboration, explanation, and debugging. During the AI literacy component, students used translanguaging to interpret key concepts introduced through the Zappy the AI Robot narrative and to clarify unfamiliar terminology. Survey responses further indicated that students developed emerging understandings of AI, particularly viewing it as a tool that can support learning and problem-solving. By positioning translanguaging as a computational practice, this study extends translanguaging scholarship into robotics and AI-supported learning environments. Findings contribute to understanding how multilingual learners engage with emerging technologies and offer implications for designing linguistically responsive and equitable computer science and AI literacy instruction.

5-5:30pm | The Impact of Teacher and AI Feedback on EFL Writing and Students’ Metalinguistic Awareness

Ayse Virlan Yilmaz, Tugrul Dolen, & Gizem Dane (TSLL26-97)

As artificial intelligence (AI) gains importance in modern classrooms, research comparing its feedback to traditional teacher feedback—particularly regarding metalinguistic awareness (MLA)—remains limited, especially in structured writing contexts. This study examined how AI and teacher feedback, delivered within a process-oriented writing framework, influenced EFL learners’ writing skills and their ability to notice, apply, and reflect on morphological and syntactic rules. The study was conducted with approximately 60 B1-level students from a Turkish state university’s English preparatory program. This proficiency level was preferred because metalinguistic development typically begins at this stage, allowing learners to use language more functionally and consciously. Using a randomized control design, participants are assigned to either an experimental group (AI feedback) or a control group (teacher feedback). Data were collected through writing rubric scores across multiple drafts and pre- and post-MLA tests. To gain a deeper perspective, 16–20 volunteers participated in recorded reflective interviews, which were subsequently analyzed using content analysis. The unique contribution of this research lies in its exploration of the cognitive and metalinguistic effects of AI-supported tools in foreign language education. Preliminary findings suggested an upward trend in explicit grammatical knowledge among the experimental group. Students demonstrated not only higher accuracy but also an improvement in their ability to justify their corrections. This suggested that the instructional intervention successfully shifted their linguistic processing from an implicit and intuitive level to a more conscious level. Finally, this study is important in helping us to understand how AI-based feedback takes place in terms of supporting conscious language awareness, which is a key challenge.

5-5:30pm | Languaging with Constructed Agents: Ecological Affordances and GenAI NPCs in CALL

Lin Zhou (TSLL26-108)

While the humanities have long explored imagined sentient agents, generative artificial intelligence (GenAI) has materialized these cultural constructs into dynamic, sociotechnical interlocutors for computer-assisted language learning (CALL). This presentation examines the pedagogical evolution of GenAI-driven non-player characters (NPCs) in game-based ecologies, analyzing their impact through the frameworks of languaging (Swain, 2006), ecological affordances (van Lier, 2004), and emerging GenAI literacies (Godwin-Jones, 2023). Unlike the rigid, single-solution pathways of legacy language game design, GenAI NPCs operate as unscripted, variable-based agents. In this mediated ecology, learners engage in active languaging via real-time interaction—using target language production as a cognitive tool to construct meaning, test linguistic hypotheses, and adapt their prompting based on live, authentic feedback. Crucially, replacing pre-programmed dialogue trees with systemic GenAI logic creates a logically bounded, predictable framework. This predictability accommodates diverse cognitive processing styles by offering a structured space for self-paced linguistic risk-taking, effectively lowering the affective filter associated with human-to-human L2 ambiguity. To succeed, learners must navigate the GenAI’s operational rules of engagement, fundamentally shifting their role from passive consumers of a narrative to critical evaluators of a sociotechnical system. By analyzing learner-NPC discourse, the presentation highlights a necessary pedagogical paradigm shift: educators must evolve from “scriptwriters” to “curators of ecologies” who architect precise system boundaries. Ultimately, analyzing how learners negotiate meaning with human-like AI demonstrates that interacting with these constructed agents fosters both communicative L2 competence and the critical digital literacy required for the post-digital workplace.

5:30-6pm | Making AI Better Listeners: The Potential of Prompt-Based Accent Familiarity for L2 Speech Perception

Yuna Bae & Okim Kang (TSLL26-149)

While AI voice chatbots are increasingly integrated into second language (L2) learning and teaching (Zou et al., 2024), they often struggle with L2-accented input (e.g., Song et al., 2022). Research on human listeners shows that speech perception is influenced by accent familiarity (e.g., Winke & Gass, 2013), raising the question of whether AI models may likewise benefit from familiarity-based interventions. This study investigates whether a human-like accent-familiarity can enhance a generative AI model’s perception of L2 speech. Specifically, it examines the effects of accent familiarity on ChatGPT-4o’s judgments of intelligibility, comprehensibility, and accentedness. Twenty-one one-minute English read-aloud samples representing three L2 accent varieties (Chinese, Indian, and Spanish) were analyzed. In a pre-test, ChatGPT evaluated each sample for intelligibility, comprehensibility, and accentedness. The intervention introduced accent familiarity through phonetic explanations and short audio samples, along with rubrics, task instructions, and comprehension checks. The same speech samples were then re-evaluated in a post-test using identical procedures. Intelligibility was operationalized as transcription accuracy using word error rate (WER), while comprehensibility and accentedness were rated on 7-point Likert scales. Linear mixed-effects models were computed with intervention and accent variety as fixed effects and speaker as a random effect. Results revealed a significant improvement in intelligibility following the intervention, with lower WER in the post-test, indicating more accurate transcriptions (F(1, 20) = 8.12, p = .009, d = 0.88), with fixed effects accounting for 18.4% of the variance and the full model explaining 52.8%. In contrast, accentedness ratings did not change significantly (F(1, 20) = 0.42, p = .526), suggesting that brief exposure did not change the model’s perception of accentedness. Comprehensibility outcomes were accent-dependent. Overall, the findings suggest that accent-familiarity prompting can enhance AI’s ability to process L2 speech and highlight practical, user-accessible ways to improve AI voice tools through targeted prompt design.

5:30-6pm | Can AI Adopt Register-Specific Grammar?: A Key Feature Analysis of Lexico-Grammatical Features

Shannon Dunn & Mariia Zemliakova (TSLL26-80)

Published academic writing is a complex, nuanced way of writing acquired through exposure to discipline-specific conventions. Professional communities share knowledge by adopting specialized, functional variations in language, and this specialization varies across different disciplines (Gray, 2021). These variations reflect authors’ implicit understanding of communicative purpose (Goulart et al., 2022), functional correspondence (Biber, 2019), and situational characteristics (Biber et al., 2015). With the recent integration of large language models (LLMs) into academic writing practices, many have voiced concerns regarding AI’s influence within scientific discourse (Skvortsova et al., 2025). However, despite growing scholarly attention to AI in academic contexts, it remains unknown to what extent LLMs generate register-appropriate writing when given prompts based on the PARTS, CLEAR, and REFINE frameworks (Park & Choo, 2025). This study seeks to address the following research question: To what extent does AI-generated writing reflect lexico-grammatical register trends within a specific academic discipline? The human-written sample consists of 30 abstracts collected from three physics journals (10 abstracts from each journal) published in 2019. A mirrored AI-generated sample was produced for the same 30 articles using Microsoft Copilot. To generate each abstract, prompts were constructed following established frameworks (PARTS and CLEAR; Park & Choo, 2025; Lo, 2023), incorporating contextual and retrieval information. After an initial output was produced, 1-3 refinement prompts were applied following the REFINE framework (Park & Choo, 2025) to optimize the final output. Both samples were then tagged for lexico-grammatical features using LxGrTagger, targeting phrasal and clausal structures. Key feature analysis was employed to identify similarities and differences between the two samples. The findings will contribute empirical evidence to how LLMs handle register-specific academic language. This will have implications for teachers and researchers who integrate AI support into their writing process..

5:30-6pm | Cultural and Literary Metaphors in EFL Student-Teachers’ Perceptions of AI

Reza Rezvani (TSLL26-55)

This paper, drawn from a larger qualitative study, investigates how Iranian university EFL student-teachers conceptualize AI language assistants through metaphor. Data were collected via handwritten exit ticket summaries from 83 participants across five cohorts (Fall 2023–Fall 2025) following classroom discussions on AI in L2 learning. Thematic analysis revealed two overarching perceptions: AI as a Relational Being (affirming, supportive) and AI as an Ambivalent/Threatening Presence (fearful, unsettling), with subthemes including Authority, Problem Solver, Social Companion under the first, and Fear under the second. This paper focuses on the cultural and literary traditions students invoked in their metaphors. Using Western metaphors, students employed a spectrum ranging from neutral or functional (e.g., assistant, tutor) to negative (e.g., slave). More specifically, they drew on Western fictional sentient agents such as Frankenstein’s creature, Hal 9000, Terminator, and Agent Smith to express fear, ambivalence, and anxiety, whereas when they drew on Persian literary and cultural traditions, their metaphors—including Rostam (akin to Hercules, from Shahnameh, the Persian epic), Simorgh (the nurturing mythical bird), hakim (wise sage), sultan (benevolent sovereign), nushdarou (a cure-all panacea, from Shahnameh)—were predominantly positive, often with tints of humor and affection, featuring figures of wisdom, protection, and non-judgmental guidance. This contrast reveals a hybrid cultural framework: students draw on Persian traditions for metaphors of trust, care, and revered authority, while Western sources supply a wider range of associations, including neutral functionality, subservience, and existential fear. By tracing these constructed agents, from Persian epic and mystical poetry alongside Western literary and cinematic traditions, the paper shows such analysis deepens understanding of learners’ relational connection and interactions with AI in L2 contexts. It concludes with implications for language educators using AI in teaching and learning, and for AI developers to consider how culturally grounded metaphors shape user expectations across different learner populations.