Part two of a diptych: Language, learning and play with AI.

While my professional engagement with LLMs began from the outset, a use that is serious, strategic and grounded in work, I had a parallel, personal track that began in a more playful way, as outlined in the previous article. Over time, I had a shift. The whimsy gave way to experimentation. The absurdity turned to structure. My personal engagement with LLMs began to reflect the same values that shaped my life as a polyglot, a lifelong learner and an educator, not necessarily of others, but of myself.

Filling the Gaps No One Else Filled

I spent decades learning languages, not just the major ones, but also the minor ones, the sometimes forgotten but culturally rich languages like Romansch, Frisian, Sardinian, Occitan, Letzebuergesch, Plattdeutsch and Galego. I also have studied in depth Creole languages like Papiamentu, Tok Pisin, Crioulo and French-based Caribbean and Indian Ocean creole languages. Studying them is quite an undertaking in itself: there are hardly any materials and whatever books there are do not serve my needs as an advanced learner in each of them. So, I built the courses for myself.

Using several LLMs, I created full courses:

·       Dialogues contextualised with detailed historical and cultural references

·       Grammar explanations tailored to my own level and curiosity

·       Vocabulary lists and thematic sets

·       Exercises, quizzes and tests.

Then I took it further: integrated text-to-speech (TTS) generated native-sounding audio, gamified interactions, flashcard generated directly from content. All this with no coding knowledge, just intention, knowledge of the languages in question and openness to co-create. It has been an incredible, absorbing experience.

From the Sterile to the Real

A lot of mainstream language learning materials feel like they were created in a vacuum. A well-to-do vacuum at that, with dialogues involving fine dining and art museums. This is hardly the typical experience of a lot of people who go to live in a new country. Real life is budgeting, shared housing in dodgy neighbourhoods, social friction, accents you do not understand, administrative headaches. And landlords who won’t deal with the rat problem or the black mould.

I started building courses that reflected this reality, such as:

·       Broke protagonists navigating low-income life in a rough part of Buenos Aires

·       A French learner surviving in outer Paris, getting to grips with colloquial Parisian French and Arabic

·       Dialogues about social challenges, such as unemployment, bureaucracy, racism, exclusion, gentrification, and so on

·       A Syrian migrant’s experience of settling into a northern English city.

This is not intended for fantasy tourism. It is immersion through friction. Real language, real contexts, for real learners.

Some of it was hit-and-miss, particularly with earlier versions of LLMs. For example, ChatGPT 3.5 and 4 would include terrible mistakes, such as ‘actuellement’ in French used to mean ‘actually’. Of course, this is an error a learner would not even make at school. Or Arabic words that would curiously appear in a left-to-right reading direction (how is that even possible?) or hallucinations would occur, as in nonsense words. Lesser-known languages would ‘veer’ into the closest majority language: Romansch would suddenly become full-blown Italian; Plattdeutsch would veer into Standard German, Occitan would veer into Catalan, and so on. However, none of this occurred with the o1 pro model and subsequent ones.

Slot Machine Learning: Spinning the Dial of Context

Then came what I call ‘slot machine learning’, i.e. further granularity, localisation and customisation on demand:

·       Target language to learn: French, Basque, Japanese, Tamazight, Quechua, etc.

·       Learning from the perspective of a speaker of: Portuguese, Mongolian, Swahili, Tongan, etc.

·       Cultural frame: LGBTQ+, Muslim, working class, undocumented, business traveller, etc.

·       Historical setting (just because you can): 1970s Portugal, Cold War Berlin, 1968 Paris, etc.

You can click spin and build the course. The possibilities are endless: A Japanese course for native Portuguese speakers. A Norwegian course for Mongolians. A Spanish course set in Cuzco that also teaches you some Quechua. A French course set in Quimper that also teaches you some Breton. A guide to moving to Warsaw as a Muslim woman. Settling into Bucharest as a Gambian student. A survival guide for Czech-speaking LGBTQ+ migrants in Singapore.

These are courses that perhaps nobody dares to imagine. They are not sanitised or scalable or easily monetised. These were just experiments for an audience of one: me. However, if they were to be used beyond only me, they would need to have expert steer and input from people who come from the appropriate backgrounds. It cannot be just text generated by an LLM under the auspices of Mark Gibson. Nevertheless, this is an illustration of how wonderfully tailored these courses can be.

The Miracle of Co-Creation

I do not know how to write code. In fact, I don’t think I would even recognise it if I saw it. I am not a software developer. But I can co-create things with LLMs that I would have thought impossible. A tool like ChatGPT scaffolds the idea, gives it form and lets it breathe. I do the reshaping, refining and personalisation. The magic is in the co-creation. How deeply personalised and customised these courses can be blows my mind.

In the last article, I wrote about the limits of language mastery. For a polyglot, you have to come to terms with languages that come in and out of dormancy in your head. The talent of the polyglot is to be able to bring them back with ease. With LLMs, I can keep the dormant languages closer. I can summon context. I can exercise the muscle. I can test comprehension. I can audit levels of atrophy and signs of vitality. It can allow the dormant languages to come back to me faster, when I need them to.

Where the Personal Meets the Professional

I began this article by saying that my personal and professional uses of LLMs are separate and run in parallel, at least I intended them to be. But somewhere along the way, they met and they converged. The creativity I explored in my own learning now feeds ideas into my professional work and vice versa.

This is the beauty of tools like LLMs. It is not so much in the answers they generate, but in the better, more targeted questions that they invite. They are tools, but they are also mirrors. They draw attention to what you want to build, what you have been missing and what you did not know that you needed. And maybe… what nobody else thought to make.

I mentioned in a previous article that at least one fifth of humanity are using a tool like ChatGPT. Just think of all of the millions upon millions of new ideas that this co-creation is likely to generate. It is exciting to think about.

Thank you for reading,

Mark Gibson

Leeds, United Kingdom, Easter 2025