Deep diving into small language models with Dev Aditya

Professor Phil Newton from the University of Swansea talks through his recent paper: Evidence of Learning Through Assessment. Protecting the Value of a Degree in the Age of AI. Phil highlights clear and present threats to higher education by the sector’s continued use of online, unsupervised and remote assessments. This, Phil argues, is compromising universities’ ability to properly certify learning, and has the potential to erode the value of a degree. In this conversation we talk through the six key recommendations put forward in the paper, and how they might realistically be implemented. We discuss where change can come from, the problem of SOUR assessments, the misaligned incentives in higher education, the loss of confidence from employers, the ethical concerns around placing the burden of trust on to students, how other countries have reacted, the need for training, resource barriers, the information that should be included on degree transcripts, and many other topics. Guest bio Phil Newton teaches Neuroscience across a variety of programmes in the School. He also teaches on evidence-based education as part of the Schools online MSc in Medical Education. He came to Swansea in 2009 from the University of California at San Francisco, having completed a Ph.D in Biochemistry and Molecular Biology at the University of Leeds (2001) and a B.Sc in Biology at Aberystwyth University. He also holds an M.Sc in Medical Education from Cardiff University (2012). LinkeIn https://www.linkedin.com/in/prof-phil-newton-21966b8a/ Web - https://www.swansea.ac.uk/staff/p.newton/ YouTube - https://www.youtube.com/user/newtonsneurosci Scholar - https://scholar.google.co.uk/citations?hl=en&user=77aJ0IMAAAAJ&view_op=list_works ORCID - https://orcid.org/0000-0002-5272-7979 Researchgate - https://www.researchgate.net/profile/Phil_Newton/ Further reading Newton, P. (2026). Evidence of Learning Through Assessment. Protecting the Value of a Degree in the Age of AI. Policy Exchange. [Online]. Available at: https://policyexchange.org.uk/publication/evidence-of-learning-through-assessment/. Newton, P. M. and Draper, M. J. (2025) ‘Widespread use of summative online unsupervised remote (SOUR) examinations in UK higher education: ethical and quality assurance implications’, Quality in Higher Education, 31(1), pp. 127–141. doi: https://doi.org/10.1080/13538322.2025.2521174

What does a reconfigured language learning ecology look like? My guest today, Professor Pascual Pérez-Paredes, argues that AI-driven environments can “expand opportunities for language learning while creating risks related to opacity and over-reliance”. We discuss the continued importance of studying real, attested language use through well-designed corpora, enabling evidence-based comparisons and supporting traceable, reproducible research. This, Pérez-Paredes claims, is more important than ever when unreliable, non-evidenced answers from chat bots are just a mouse click away. We also discuss corpora in teaching, how AI can make corpus tools more user friendly, costs and local models, and the need for critical AI literacy, open-box pedagogies, and transparency amid increasing synthetic language. Further reading Pérez-Paredes, P., Curry, N. and Aguado Jiménez, P. (2025) 'Integrating critical corpus and AI literacies in applied linguistics: a mixed-methods study', Computer Assisted Language Learning, pp. 1–27. doi: 10.1080/09588221.2025.2569351. Pérez-Paredes, P. (2026) 'Corpus linguistics and AI in the reconfiguration of language learning ecologies', Language Teaching, pp. 1–13. doi: 10.1017/S0261444825101109 Guest bio Pascual Pérez-Paredes is a Professor of Applied Linguistics at Universidad de Murcia, and former Lecturer in Research in Second Language Education at the University of Cambridge. His main research interests are the use of corpus linguistics methods in applied linguistics, corpora and digital resources in language education, learner language variation, and corpus-assisted discourse analysis. He was the Overall Coordinator of the MEd Research Methods Strand at the Faculty of Education, University of Cambridge (2016-2019). His most recent books are Corpus Linguistics for Language Learning Research, John Benjamins Research Methods in Applied Linguistics series with co-authors Geraldine Mark & Anne O’Keeffe; Data-driven Learning in and out of the Language Classroom, Cambridge University Press, co-authored with Prof Alex Boulton and //Corpus Linguistics for education. A guide for research// in the Routledge Corpus Linguistics series. He is currently PI of the ERASMUS+ project “Freedom of movement at play: EU citizens’ identity and transnational discourses” (2023-2025) funded by the European Commission and PI in the Spanish Research Agency -funded project Broadening the scope of Data-driven learning: a multi-site & multi-data approach. Twitter @perezparedes and Bluesky @perez-paredes.bsky.social Google Scholar Profile Research Gate Profile ORCID: https://orcid.org/0000-0002-2796-338X SCOPUS: https://www.scopus.com/authid/detail.uri?authorId=26038713600# WOS: https://www.webofscience.com/wos/author/record/G-6496-2011

Today we take a critical look at the promises and pitfalls of using GenAI in educational research. I'm joined by Jasper Roe, Assistant Professor in Digital Pedagogies at Durham University, and Stephen Gow, former Leverhulme Research Fellow at Edinburgh Napier University and lead author of the Student Experiences on Generative AI Project. We discuss Jasper's recent book, How to Use Generative AI in Educational Research, and Stephen's work on an AI assisted scoping review. We talk through the importance of foundational knowledge, how we talk to students about research, opportunities for classroom-based scholarship, efficiency trade-offs, concerns around AI detection, the need for 'humans in the loop', and other topics. Key reading Gow S, Illingworth S (2026), "Dynamic tensions: an AI-assisted critical scoping review of university students' qualitative experiences of GenAI". Artificial Intelligence in Education, Vol. 2 No. 1 pp. 67–89, doi: https://doi.org/10.1108/AIIE-06-2025-0151 Roe, J. (2025). How to Use Generative AI in Educational Research. Cambridge: CUP. DOI: https://doi.org/10.1017/9781009675338 Texts referred to in the conversation Jansson, D.G. & Smith, S.M. (1991) ‘Design fixation’, Design Studies, 12(1), pp. 3–11. doi: 10.1016/0142-694X(91)90003-F. Perkins M, Roe J (2024), "Academic publisher guidelines on AI usage: A ChatGPT supported thematic analysis". F1000Research, Vol. 12 No. 1398, doi: https://doi.org/10.12688/f1000research.142411.2 Pinochet, L.H.C., Miadaira Hamza, K. & Dwivedi, Y.K. (2026) ‘Editorial: Generative AI, human authorship and the transformation of scholarly communication’, Revista de Gestao, 33, pp. 31–40. doi: 10.1108/REGE-04-2026-0058. Your Undivided Attention (2025) ‘AI and Cancer: Why Superintelligence Won’t Get Us to a Cure’ [Podcast]. Available at: https://podcasts.apple.com/us/podcast/ai-and-cancer-why-superintelligence-wont-get-us-to-a-cure/id1460030305?i=1000764657602 (Accessed: 19 August 2026). Guest bios Dr Jasper Roe is a leading scholar in educational technology, artificial intelligence, and academic integrity. He has worked internationally, holding senior leadership positions at universities in Vietnam and Singapore, where he spent over a decade. His mixed-methods research is deeply interdisciplinary, drawing on applied linguistics, corpus and discourse analysis, sociology, and anthropology. Jasper is the author of How to Use Generative AI in Educational Research (Cambridge University Press) and has published over 30 peer-reviewed journal articles in prestigious outlets. He is best known as co-creator of the AI Assessment Scale (AIAS) and is a regularly invited keynote speaker at international conferences and for intergovernmental organisations, including for UNESCO. Some of his work in news media is available at the below links:Channel News Asia https://www.channelnewsasia.com/author/jasper-roe Times Higher Education Campus https://www.timeshighereducation.com/campus/authors/jasper-roe For academic publications please visit his Google Scholar Connect on Linkedin Dr Stephen Gow was the Leverhulme Research Fellow at the Department of Learning and Teaching Enhancement (DLTE) at Edinburgh Napier University. During this role he led the Student Experiences on Generative AI Project (StudentXGenAI), this project carried out the StudentXGenAI Survey with a response rate of over 7,000 students at UK institutions and interviews with students across the UK in addition to integrating GenAI into the research process. He is an expert on academic integrity, assessment and GenAI, and the Chair of the Northern Academic Integrity Forum. He is now associate staff with Department of Education, University of York and available for consultation and research projects related to GenAI in education. He can be contacted at [email protected] or via Linkedin: Stephen Gow | LinkedIn Corrections and clarifications At 44:05, Phil mentions a Linkedin post by Andrew Maynard which de

I’m joined by Dr Sophia Mavridi to talk about her doctoral research: Second language students’ experiences of AI-assisted writing in higher education: A phenomenographic perspective. Sophia identifies six broad categories that emerge when observing students’ use of AI assisted writing, and in this podcast we dive into these categories: achieving better grades, expressing ideas more clearly, managing time/workload, habituating academic writing norms (e.g., criticality/reflective writing), increasing interaction, and increasing sophistication. We go over inferences and implications of GenAI in student writing, how talk to students about developing a voice, and how we do this ourselves in our own writing, the tension between making your writing distinctive while adhering to academic norms, where we should set expectations for undergrads versus PhD students. Plus: AI slop on Linkedin. We also cover Sophia’s paper published last year in Technology in Language Teaching & Learning that critiques institutional responses to GenAI in education including prohibition, avoidance, teacher support, technological enthusiasm, and critical engagement. We discuss directions for language teacher education, including rethinking writing pedagogy, discernment, transparency, teacher educator development, and ethical dilemmas. Speaker Bio Dr Sophia Mavridi is a Senior Lecturer in Educational Technology and TESOL at De Montfort University (UK). She specialises in digital pedagogy, online learning, and AI in education. She has extensive experience in language teacher education, working with organisations such as the British Council, NILE, and Bell Foundation, and is committed to bridging research and practice. https://sophiamavridi.com/ Further reading Watch Sophia’s keynote form the TELSIG symposium earlier this year: Behind the scenes: How students use AI (and what we can learn from it). The survey Sophia alludes to at approximately 55 minutes in is the HEPI Student Generative AI Survey. Mavridi, S. (2025). Hype, Fear, and Everything in Between: A Critical Typology of Responses to AI and their Implications for Language Teacher Education. Technology in Language Teaching & Learning, 7(2), 103219 Available at: https://doi.org/10.29140/tltl.v7n2.103219 Timecodes 00:00 Podcast intro 01:41 Reflections on the PhD Viva 04:55 Teacher identity 08:19 Overview of the PhD research 09:26 Six AI writing categories 11:03 Integrity and voice 14:33 Participants and context 16:55 Self regulating AI use in writing 19:33 Students surprising sophistication 24:45 Struggle in L2 writing 32:12 Why writing still matters 38:19 Academic norms versus authorial voice 44:11 AI and the loss of voice 49:09 Institutional AI typology 51:51 Teacher education directions 57:05 AI detectors and in-class writing 01:07:30 Tech hype vs pedagogy 01:11:53 Reading AI research critically

All of us started our careers somewhere, and today I'm talking to two old friends and former colleagues about how our years teaching English as a foreign language shaped our view of education and the world. We talk about the resilience and adaptability that those early classroom days taught us, the overwrought lesson plans, learning humility from failure, and how these experiences inform our work now in higher education management, instructional design and ed-tech. We go on to discuss the direction of education, the resurgence of interest in humanities subjects in the face of AI, staying up to date with knowledge and skills in an ever-changing jobs market, and the general chaos of those unique times in our lives. Guest Bios Ben Naismith, PhD is a Staff Assessment Scientist at Duolingo where he works on research and development of the Duolingo English Test. Prior to joining Duolingo, Ben worked extensively in the field of English language teaching in numerous contexts as a teacher, teacher trainer, materials developer, assessment specialist, and researcher. His current work and research into the assessment of L2 English proficiency focuses on vocabulary assessment and automated scoring of speaking and writing. His work has been published in journals including Language Testing, Studies in Second Language Acquisition, Language Teaching Research, Language Learning, and Assessing Writing, amongst others. He is the lead editor of the forthcoming Routledge Handbook of Digital Language Assessment: Innovations and Insights from the Duolingo English Test. Andrea Flores is Senior Instructional Designer at MIT Sloan School of Management and Founding member of the Boston AI2030 Chapter. She’s an AI lead and human-centered design practitioner with a diverse background in education, public health, workforce development, and technology. She has worked across non-profits, higher education, and global health, and focuses on human impact and sustainable solutions.