Abstract: In this talk, I present a corpus-based account of research on the language produced by Large Language Models (LLMs). I report on a set of studies comparing human and AI-generated texts across different contexts, using Multi-Dimensional (MD) Analysis (Berber Sardinha, in press-b; Berber Sardinha & Veirano Pinto, 2014, 2019; Biber, 1988; Veirano Pinto et al., in press) to identify the patterns of language use associated with each. I argue that Corpus Linguistics (Berber Sardinha, 2004, in press-a; Berber Sardinha & Barbara, 2009; Biber et al., 1998) is well-suited to investigate LLM output, as it enables the analysis of large volumes of textual data while attending to key linguistic attributes such as register, text, discourse, and situation of use (Berber Sardinha, 2024; Goulart et al., 2024; Mizumoto et al., 2024; Raffloer & Green, 2025).

The first studies look at the grammar of AI-generated texts. LLMs acquire grammar by processing examples from large collections of text, which is a remarkable achievement in itself. However, in doing so, they also absorb the typical characteristics of the texts included in their training data, most of which come from the web. These sources comprise an unbalanced range of registers, and the register properties of the data are rarely identified. As a result, LLMs tend to be register-unaware. When prompted to produce language in a particular register, they will deliver an output that looks like a good exemplar, yet they often fail to reproduce the grammatical patterns that define that register in actual human usage.

The remaining studies focus on the lexical patterns found in AI-generated texts, using Lexical Multi-Dimensional Analysis (Berber Sardinha & Fitzsimmons-Doolan, 2025). More precisely, the attention shifts to discourse analysis. Just as with grammar, LLMs acquire discursive patterns from the texts included in their training data. These patterns are later reinforced or suppressed during the fine-tuning phase. We will look at instances of AI-generated discourse from various sources, in addition to attempting to steer LLM reasoning through prompting.

Academic Bio: Dr. Berber Sardinha is a Professor with the Applied Linguistics and Language Studies Graduate Program at the Pontifical Catholic University of Sao Paulo (PUC-SP), Brazil.

Tony holds a BA in English and an MA in Applied Linguistics from the Pontifical Catholic University of Sao Paulo, and a PhD in English from the University of Liverpool, in the United Kingdom, under Prof. Michael Hoey’s supervision. He was a visiting researcher at Northern Arizona University, Flagstaff, USA, working under Prof. Douglas Biber’s supervision, in the early 2000s. He has used Corpus Linguistics approaches to look at various topics in linguistic description and application, such as Register Variation, Metaphor Analysis, Applied Linguistics, Language Teaching, Multi-Dimensional Analysis, Forensic Linguistics, and more recently, Artificial Intelligence.

Tony currently heads several funded projects, which rely on corpus-based approaches to investigate contemporary issues, such as the Infodemic, as well as to extend Multi-Dimensional Analysis to model both verbal and visual language, through corpora of texts and images. He is the Editor-in-Chief of the Linguistics Journal DELTA, Consulting Editor for Register Studies, and serves on the board of major journals such as the International Journal of Corpus Linguistics, Corpora, Applied Corpus Linguistics, English for Specific Purposes, Applied Linguistics, Metaphor in the Social World, in addition to being on the board of the book series Studies in Corpus Linguistics, and being the Corpus Linguistics area editor for the Encyclopedia of Applied Linguistics (2nd ed.).

Read more about Dr. Berber Sardinha here

Abstract: Generative AI and large language models (LLMs) have opened new opportunities to practice conversation meaningfully in a foreign language with a bot. I will revisit 40 years of research on dialogue-based computer-assisted language learning (CALL), before and after ChatGPT, to identify both how much it changed and what still holds up. I will explore the effectiveness of these “artificial” conversations to determine what works, for whom, for what and to what extent.

In particular, I will share preliminary findings from a new meta-analysis of chatbots for language learning, looking at both “technological eras”. In this systematic review, we observe that the field is moving beyond comparing the technology to comparing pedagogical and instructional conditions in their effects on language learning. I will call for more attention to task design and characteristics, and to scaffolding in conversation, particularly productive support for less proficient L2 speakers.

Academic Bio: Dr. Serge Bibauw is an Assistant Professor of French language teaching at the University of Louvain (UCLouvain), in Louvain-la-Neuve, Belgium. He is affiliated with the Institute for the Analysis of Change in Contemporary and Historical Societies (IACCHOS), co-directing the Cripedis research center, and co-affiliated with Girsef. He collaborates intensely with researchers at CENTAL, CECL and TeAMM research groups. He is also an associated professor at ITEC, an imec research group at KU Leuven. He was previously a professor and a French and English teacher trainer at Universidad Central del Ecuador.

He studies conversational AI for language learning, or dialogue-based CALL, more precisely task-oriented chatbots (or dialogue systems) for language learning, at the intersection of task-based language teaching (TBLT), computer-assisted language learning and natural language processing (NLP). His research focuses on the instructional design and the evaluation of the effectiveness of conversational AI for L2 proficiency development, particularly L2 vocabulary size, spoken and written L2 fluency, and longitudinal assessment of L2 proficiency in various instructed SLA contexts.

Read more about Dr. Serge Bibauw here

Abstract: Traditionally, human assessors such as teachers and peers have been the primary providers of feedback in language classrooms (Zou et al., 2024). With the development of digital technologies, however, feedback provision has expanded to include automated systems capable of delivering corrective feedback (CF) on learner production. More recently, generative artificial intelligence (GAI) technologies have opened new possibilities for providing personalized and interactive feedback (Yan, 2024). However, despite growing interest in GAI feedback for language learning, important questions remain about how it can be effectively integrated into the language learning environment, particularly regarding its pedagogical effectiveness, learner perceptions, and implications for instructional design.

In this talk, I will explore the potential and challenges of GAI as a source of CF by drawing on my recent classroom-based studies that illustrate possible applications of GAI feedback in instructional contexts. The first project examined GAI as an asynchronous feedback provider in student writing, while the second project investigated GAI as a conversational partner that provided CF during text-chat interactions. I will share findings related to how learners engage with GAI feedback, the extent to which it may support language development, and how learners perceive its usefulness and reliability. I will conclude my talk by highlighting possibilities and challenges of GAI feedback in language classrooms and outlining future directions for research and implementation in classroom settings.

Academic Bio: Dr. Sanghee Kang is an Assistant Professor of Second Language Acquisition, English as a Second Language and Korean Studies in the Department of Languages, Cultures & Applied Linguistics at Carnegie Mellon University (CMU), in Pittsburgh, USA. She earned her Ph.D. in Applied Linguistics from Georgia State University, where her dissertation focused on the role of chatbot-based interaction and learner characteristics in the alignment-driven learning of second language grammar and pragmatics.

Her research interests include technology-oriented second language acquisition, with a particular focus on the role of generative AI, task-based language teaching, and digital multimodal composition. Her current work examines generative AI in language learning, focusing on AI-generated feedback, collaborative writing, and chatbot-based interaction.

Her research has appeared in Modern Language Journal, System, Foreign Language Annals, Computers and Composition, Language Teaching Research, Language Awareness, and International Journal of Applied Linguistics, among others.

Read more about Dr. Sanghee Kang here

Abstract: Since the public release of ChatGPT in November 2022, Generative AI (GenAI) has become a buzzword in various professions. The technology has been widely deemed as a handy tool for reducing labour-intensive work done by humans. The promise of GenAI for language education has been recognised by applied linguists, too. Many view the technology as a “game changer” for second language learning and assessment but concerns around responsible and ethical use of AI remain (Chapelle, 2024, 2025; Xi, 2023). In this talk, we will discuss the promise and challenges of using GenAI in three areas of language education.

The second area is using GenAI in spoken dialogue systems (SDSs) for speaking assessment and practice. GenAI has fuelled a desire to move beyond monologic tasks toward dynamic scenario-based oral interactions and context-sensitive conversations tailored to user profiles. However, challenges remain in ensuring construct validity and test fairness in SDS-mediated tasks and developing relevant ethical and validation frameworks.

Academic Bio: Dr. Yasin Karatay is a Senior Research Manager at Cambridge University Press & Assessment, where he leads research projects on conversational AI and spoken dialogue systems in L2 speaking assessment and learning contexts. His research interests include computer-based language testing, spoken dialogue systems, automated writing evaluation, and test validation.

He holds a PhD in Applied Linguistics and Technology from Iowa State University. Before joining Cambridge, he was a lecturer at Düzce University (Türkiye) and a teaching and research assistant at Iowa State University (USA). Following the completion of his PhD, he worked as a Postdoctoral Research Associate at Iowa State University, contributing to the revision of a Global Online Course project sponsored by the U.S. Department of State. His current work focuses on leveraging cutting-edge technologies to enhance L2 speaking assessment and learning materials. He also teaches at the MSt in English Language Assessment programme at the University of Cambridge.

Read more about Dr. Yasin Karatay here

Abstract: Since the public release of ChatGPT in November 2022, Generative AI (GenAI) has become a buzzword in various professions. The technology has been widely deemed as a handy tool for reducing labour-intensive work done by humans. The promise of GenAI for language education has been recognised by applied linguists, too. Many view the technology as a “game changer” for second language learning and assessment but concerns around responsible and ethical use of AI remain (Chapelle, 2024, 2025; Xi, 2023). In this talk, we will discuss the promise and challenges of using GenAI in three areas of language education.

The third area is using GenAI in integrated learning and assessment (ILA). A GenAI agent can provide scaffolds for learning, personalised content and practice, and learning-oriented assessment and feedback. It holds promise for enhancing learner engagement and motivation, fostering learner autonomy, and supporting multilingual and culturally inclusive learning. However, concerns remain that AI may generate inaccurate or outdated information, lack cultural nuance, and potentially hinder critical thinking and creativity. We will conclude our talk by identifying research priorities in each area.

Academic Bio: Dr. Hye-won Lee is a Senior Research Manager at Cambridge University Press & Assessment, where she leads research projects focused on developing next-generation assessments. She also provides research-based evidence to validate the quality of both current and future assessments. Her current research interests include promoting language assessment literacy for the student constituency and defining language ability in data-driven diagnostic assessment.

Hye-won holds a PhD in Applied Linguistics and Technology from Iowa State University (USA), where she specialised in computer-assisted language assessment, argument-based validation, and quantitative research methods. Prior to joining Cambridge, she participated in long-term research projects as an intern at the Center for Applied Linguistics and Educational Testing Service in the USA. She also taught and advised in-service English teachers in TESOL Master’s programmes in South Korea. Hye-won has delivered numerous presentations at regional, national, and international conferences and has published in peer-reviewed journals (e.g., Applied Linguistics), edited volumes, encyclopedias, and research reports.

Read more about Dr. Hye-won Lee here

Abstract: Since the public release of ChatGPT in November 2022, Generative AI (GenAI) has become a buzzword in various professions. The technology has been widely deemed as a handy tool for reducing labour-intensive work done by humans. The promise of GenAI for language education has been recognised by applied linguists, too. Many view the technology as a “game changer” for second language learning and assessment but concerns around responsible and ethical use of AI remain (Chapelle, 2024, 2025; Xi, 2023). In this talk, we will discuss the promise and challenges of using GenAI in three areas of language education.

The first area is automated scoring of constructed responses. Developing automated scoring systems typically requires a sufficient amount of training data and expertise in machine learning (ML). To address the problem of data shortage, GenAI has been used to generate synthetic L2 data. Additionally, there is an increasing body of research on fine-tuning large language models (LLMs) for automated scoring. Compared to traditional ML techniques, this new approach is more accessible to nonexperts, has a lower demand on training data, and may increase the explainability of automated scores to some extent. However, there are outstanding validity and reliability issues related to LLMs’ lack of knowledge of learner language and the potential bias in LLMs.

Academic Bio: Dr Jing Xu is Head of Propositions Research-English at Cambridge University Press & Assessment, where he oversees innovative research on technology-mediated English language assessment and learning. He received his PhD in Applied Linguistics & Technology from Iowa State University.

His research interests are centred on the application of cutting-edge AI technologies to English language education and the related validity issues. In collaboration with his colleagues at Cambridge, his current work focuses on automarking of constructed responses in high-stakes L2 English assessment and AI-mediated English-speaking tasks. He is Co-Chair of the Automated Language Assessment SIG at the International Language Testing Association (ILTA) and a Subject Matter Lead for the Cambridge Institute for Automated Language Teaching and Assessment (ALTA). He also teaches at the MSt in English Language Assessment programme at the University of Cambridge.

Read more about Dr. Jing Xu here