Language learning has never been more accessible. Most of us know somebody whose life has been taken over by a certain green bird, and as a scientist who now specialises in education technology, I welcome anything that can bring the classroom into the palm of your hand.
However, many language-learning apps are at their best at the beginning, teaching someone their first words, introducing basic grammar and gradually guiding them through a carefully ordered course. The difficulty is that this carefully ordered approach becomes less effective as a learner progresses.
It was the much more complex and individual stage that follows that caught my attention.
This “day 1,000” stage is a tricky point at which someone can already hold a conversation, read an article or follow much of a television programme in English, but still does not feel fluent. The problem is no longer a simple lack of knowledge. It is uneven knowledge, built up through years of lessons, work, travel, relationships and media.
Those remaining gaps can have real consequences. At work, they can make someone hesitate before speaking, misunderstand a colleague’s tone or struggle to express their expertise as confidently as they could in their first language. In healthcare, a missed nuance can make it harder for a patient to explain how they feel or for a professional to communicate reassurance and empathy. Socially, slang, humour and in-jokes can leave somebody who speaks English well feeling as though they are still standing outside the conversation.
This is why I decided to build an AI tutor that creates an evolving picture of what each learner understands, identifies the gaps that are unique to them and uses that knowledge to decide what they should learn next.
The shared learning path eventually ends
For a beginner, a predefined course makes sense. Most people starting a new language need to learn broadly the same foundations. They need common words, simple sentence structures and essential grammar. Some will progress more quickly than others, but the general direction can remain similar.
That changes once people begin using the language in their everyday lives.
Languages are different from many academic subjects because we do not only study them. We use them to work, watch films, listen to music, read the news, build relationships and make ourselves understood. Those experiences gradually shape what each person knows.
At this point, the intermediate learner is not simply standing halfway along a universal path. They have developed an individual collection of strengths and gaps. The further they progress, the more individual that collection becomes.
Knowing a word is not a yes-or-no question
We often talk about vocabulary as though someone either knows a word or does not. In reality, knowledge is far more complicated.
A learner may understand “bank” when it refers to a financial institution but not recognise it as the side of a river. They might understand a word when reading it but be unable to recall it during a conversation. They may know its meaning but not its pronunciation, spelling or the circumstances in which a native speaker would naturally use it.
The same applies to phrasal verbs, idioms, grammar and pronunciation. A broad description such as “intermediate English” is useful, but it cannot tell us which of these individual elements someone understands or what they should practise during their next ten minutes of learning.
A good human tutor begins to develop this picture naturally. Through conversation and observation, they notice hesitation, repeated mistakes and gaps in understanding. They then adapt what happens next.
This is where I believe AI has the greatest potential in education. Its most valuable role is not simply to generate more content. It is to help build a continually updated and highly detailed picture of what an individual knows.
Personalisation requires more than a chatbot
An AI system is not necessarily personalised simply because one person is having an individual conversation with it.
Generative AI can create an almost unlimited number of explanations, questions and exercises. That is useful, but unlimited content is not the same as education. If a system does not understand what a learner already knows, it can still produce the wrong activity at the wrong level.
A genuinely adaptive tutor requires several things. It needs a detailed map of the language, covering individual meanings, expressions, grammar rules and other elements of knowledge. It needs to interpret what the learner’s previous answers and interactions reveal. It must then decide which activity is most likely to help that person progress.
That decision should be updated continually. Every answer, mistake and moment of uncertainty provides another small piece of information. The system can use that information to refine its understanding and choose what should happen next.
Teaching is a little like playing a strategy game: there is an end goal, but reaching it depends on making the right decision at each stage. AI can help choose what a learner should do next, but it needs to understand which approaches actually work for different people.
Having millions of people follow the same course, or counting clicks, streaks and completed exercises, does not prove that anyone has learnt or remembered anything.
Everyday life could become part of the curriculum
Once somebody already uses English, personalisation creates another opportunity. Learning no longer needs to remain inside a separate course. It can become connected to the language that person encounters naturally.
Imagine that someone wants to watch a particular film in English. An adaptive system could examine the language used in the film, compare it with its existing picture of the learner and identify the words, meanings and idioms they are unlikely to understand.
The learner could explore those expressions beforehand. When they then hear them in the film, the story provides context and reinforcement. They are not memorising a random list of vocabulary. They are learning something because they have an immediate reason to understand it.
The same principle could apply to television programmes, songs, books, online articles or professional language. With the learner’s permission, their interests and everyday experiences could help determine what they learn next.
This relevance is powerful. People are more likely to remember an expression when it unlocks the meaning of a song they love, helps them understand a scene or allows them to communicate more confidently at work.
Our first attempt taught us an important lesson
Building strong technology does not guarantee a strong learning experience. My team and I discovered this when our first version failed.
We had developed promising algorithms, but the product was built primarily around short, screen-based vocabulary activities. It was almost a “TikTok for language learning”. We had underestimated how differently people learn at different moments of the day.
Someone might have five minutes on a train, be able only to listen while cooking, or have time for a focused lesson. Each situation needs a different approach.
That failure changed how I thought about the problem. Educational AI must fit around human behaviour rather than expecting human behaviour to fit around the technology.
AI cannot replace everything a human tutor provides
A good tutor does more than decide what somebody should learn next. They encourage them, build their confidence and notice when they are frustrated or losing interest. Knowing that another person cares about your progress can also help you keep going. I do not believe AI can replace that relationship, and it should not pretend to.
What AI can do is give more people access to learning that responds to their own needs. It can spot what somebody understands, where they are struggling and what would help them next, whether it is used between lessons or by someone who cannot afford regular private tuition. This is particularly valuable for non-beginners, because no two learners have the same gaps.
The aim is not an endless chatbot conversation, but useful support that offers the right challenge, in the right way, at the right time. Not on day one, but on day 1,000.
