Artificial intelligence is being increasingly woven into the fabric of Scottish schools and universities. For language teaching, tools like machine translation, automatic transcription and AI-based lesson planning may lighten workloads, enhance certain aspects of learning and improve accessibility. Yet many people are rightly concerned that if we rely on AI excessively, we risk losing fundamental skills — critical reasoning chief among them — and may even undermine the human essence of language itself. In a country facing declining uptake in modern languages, these are significant concerns. How do we incorporate AI in our schools and universities without eroding the language learning process?
What do we mean by ‘AI’?
In common parlance today, ‘AI’ has become synonymous with Large Language Models, or LLMs — systems like ChatGPT and Google Gemini. As AI ethics expert Margaret Mitchell argues, this conflation drives misunderstandings about what these technologies represent — and, crucially, what they can’t do. AI, broadly speaking, is any technology that attempts to accomplish tasks traditionally relying on human intelligence. A system that identifies possible tumours from an X-ray scan is a type of AI, as is the system underlying a self-driving car. So is an LLM. What they all have in common is pattern recognition. But what LLMs do differently — and this is both the remarkable and dangerous thing about them — is that they generate the very thing they’ve been trained on: fluent, even ‘artistic’ text, often in diverse languages.
We naturally, and very often unfortunately, use people’s fluency and verbal artistry as a proxy for their intelligence level. When an LLM produces polished, confident prose, it is easy to assume that some genuine understanding lies beneath it. It is not surprising, therefore, to hear people (and companies) discuss LLM-based chatbots as having ‘PhD-level’ intelligence (whatever that means) or being able to ‘reason’. Nevertheless, despite being very useful sometimes, LLMs function differently from human beings. At their core, they are just exceptionally good next-word prediction machines. And how well they play the next-word prediction game for a given language depends on how much data is available for it. While the performance of LLMs for larger languages like Spanish or French is now very impressive in many domains, this is not true for smaller languages, such as Scots or Scottish Gaelic.
Not all LLMs are created equal
In a recent study, a team of linguists and computer scientists evaluated 19 leading LLMs on their Gaelic language competence — spanning grammar, translation and cultural knowledge — and compared them against a baseline of 30 fluent Gaelic speakers. The results varied enormously. The best-performing model, Google’s Gemini 3 Pro Preview, scored 83.3% on an expert-authored grammar test, actually surpassing the fluent-speaker average of 78.1%. But many models performed only slightly better than chance. This pattern will be familiar to researchers working on other low-resource languages.
These findings have practical implications. If a company is trying to sell an educational product to Scottish schools that will save time for language teachers during lesson and assessment planning, one should start by checking which LLMs are being used under the hood and how well they perform for the languages involved. A model that excels for French may flounder with Gaelic. Oversight and evaluation are important here.
Cognitive offloading and the language classroom
Undoubtedly, more should be done to educate our teachers and students about how to use tools like LLMs and machine translation responsibly and productively, without handicapping their own language and reasoning abilities. Over-reliance on AI encourages what psychologists call cognitive offloading — the tendency to let machines do the thinking for us. If pupils routinely outsource grammar, vocabulary recall and composition to machines, they risk short-circuiting their own acquisition processes. The question, then, is how AI should feature in language education: what kinds of applications might enhance language learning, and what kinds might impede it?
What augmentation looks like in practice
We know that communicative competence in a language requires many different abilities, among them comprehension, production and inter-cultural awareness. Only the most sophisticated AI systems can emulate a human interlocutor. But simpler systems can provide powerful interventions, if used well.
For instance, even a relatively simple subtitling system can help students gain language comprehension skills. The ÈIST project, led by the University of Edinburgh, recently launched such a system for Gaelic. This type of tool offers scaffolding towards understanding fluent speech at speed. By toggling between subtitles in the target language and the learner’s native language, students can quickly begin to recognise new vocabulary and idioms.
Figure: Gaelic and English subtitles produced by the ÈIST Gaelic Subtitling system
Crucially, this kind of technology also invites active learning rather than passive consumption. More advanced students could work with recordings that the system struggles to recognise — such as with unusual dialects or in noisy acoustic situations — and correct the errors on screen. In doing so, the learner becomes the authority and the AI becomes the object of study: a nice inversion that embodies the principle of intelligence augmentation (IA) rather than replacement (AI).
Conclusion
When considering the place of AI in Scottish language education, our priority should be augmentation, not replacement. In a country facing declining uptake in modern languages, we should embrace the possibilities of language technology for making language learning more accessible, effective and enjoyable. The interactive and immediate nature of good language technology apps could even draw students towards subjects that have a reputation for being difficult. But we must be clear-eyed about the limitations — particularly for languages like Gaelic, where LLM performance is uneven and undertested. Rigorous, language-specific evaluation is not an afterthought; it is a precondition for responsible use.
Professor Will Lamb is Personal Chair in Gaelic Ethnology and Linguistics at the University of Edinburgh since 2022. His research interests span music, linguistics, traditional narrative and language technology. He is known, in particular, for his work on formulaic language, traditional music, Gaelic grammatical description and Natural Language Processing (NLP). Most of his recent work has been in Gaelic NLP, and he recently finished an MSc in Speech and Language Processing. Prior to joining the University of Edinburgh in 2010, he was a lecturer at the University of the Highlands and Islands (LCC Benbecula) and an Honorary Research Fellow of Glasgow University.
