AI outperforms humans not only at coding and playing chess, but also at language learning. What “strategies” does AI apply to master languages so successfully? Are we humans capable of performing the same strategies? And is it any good asking these questions?
What will AI become in the next 20 years?
I have just finished listening to one of the most chilling podcasts episode in recent history with Geoffrey Hinton about our future with AI. Hinton, one of the founding fathers of AI, wants to warn us about AI becoming so knowledgeable and intelligent in the years to come that it just might decide to eliminate its inferior creator.
To be honest, I had not worried about that until now. I have a growing sense that the usage of AI could make us lazy and less creative, that it will replace much ‘low-end’ intellectual workers (already happening), however, so far, I’ve been mostly amazed by the rapid development of AI. It has improved so much since it was first released that it would be difficult to put the toy back in the box. From quickly formulating an email, doing research to practicing languages. It seems that AI can do more every day and it is “just too good for too many things”.
AI knows my mother tongue better than I do
Coincidentally, I had another interesting experience today. It turns out that you can also chat in Frisian with ChatGPT. That is a language with approximately 400,000 speakers, has a small written tradition and is most certainly not a world language. It happens to be my first language, but I had not yet thought of testing it in conversation with the popular AI-assistant. Very much to my surprise, it turned out that ChatGPT also mastered this minority language and can even teach it to you.
What can we learn from AI?
This made me wonder how AI models are trained in various languages so effectively and what we humans can learn from that. Do they use strategies that work well for us humans too?
By the way, I see the irony of first creating human-like computer-based intelligence that imitates the way our brains work and, once it has surpassed human cognitive capabilities, it’s us looking for ways to resemble AI.
Key ‘learning strategies’ in AI language models
The answer I got from ChatGPT was (of course) very helpful and not in the least bit derogatory towards us humans. It stated that although the human brain and AI function very differently, there are key learning strategies in AI language models that human learners can meaningfully adopt – especially when it comes to acquiring new languages efficiently and intuitively. Let’s see if ChatGPT is just being kind or we can actually follow the same path.
1. Massive exposure + repetition (input-based learning)
In AI
Language models like GPT are trained on millions of text examples. So for “massive exposure” massive data-sets are needed. Theoretically, the model only needs to see the training material once and store it somewhere. It doesn’t forget and it can be trained 24/7. Imagine reading all the world’s great literature on Gutenberg.org or entire volumes of the Encyclopædia Britannica – and be able to remember every fact and tiny detail for the rest of time.
What’s more, AI models learn language purely through exposure to usage patterns, not abstract rules. Unlike the first AI researcher expected, building artificial intelligence from logic and abstract reasoning never really worked out.
For humans
This whole procedure is somewhat similar to comprehensible input theory, where consuming lots of content – like what I do in my vocabulary notes – that you’re mostly able to understand is key. The brain picks up patterns naturally from rich, frequent exposure on your language level.
The main difference: we humans are obviously more limited in our choice of input. We can’t just upload all classical Chinese literature to our brain nor can we absorb Chinese Wikipedia and start seeing the relevant patterns. Not to mention learning being too slow and frustrating if we would attempt to do so.
Human takeaway
If there’s anything that we can learn from all this then probably this:
- Read and listen a lot – daily, in meaningful contexts.
- Use narrow reading/listening: immerse yourself in one topic at a time for repetition.
- Don’t stress about full understanding (easier said for AI than for humans); regular exposure helps to build fluency over time.
2. Learning through prediction
In AI
AI learns by predicting the next word in a sentence – this forces it to internalize grammar, word order, and meaning. It reminds me of listening to a boring teacher for so long that you know his next words before he actually says them.
For humans
Neurolinguistics suggests our brains also work predictively when processing language. This is most obvious in conversations, we’re able to complete unfinished sentences or fill in information gaps, but also in texts. When you read “once upon a …”, your brain is likely going to add “time” for example, because it knows the pattern.
Human takeaway
What follows from this is actually surprisingly useful. You can:
- Pause while reading or listening and try to guess what comes next.
- Use fill-in-the-blank exercises (cloze tests).
- Practice active listening by anticipating sentence completions. I also find repeating sentences from hearing an activity that forces you to reproduce certain patterns. The faster you’re able to do it, the better you’ve mastered the patterns.
3. Learning in context, not in isolation
In AI
AI doesn’t learn abstract grammar rules – it learns usage patterns from rich context. It understands context extremely well, far better than Google search used to do – if that is a fair comparison.
For humans
We all know that grammar makes more sense when learned through real examples. It doesn’t make much sense to internalize abstract grammar patterns like a mathematical formula. Although it would be very nice both for learners and teachers, unfortunately, memorizing rules alone rarely leads to fluent usage. Your brain needs practice (in authentic contexts) to reproduce patterns without conscious effort.
Human takeaway
So what do we learn from this? That we should:
- Learn through chunks and patterns (e.g. “Would you mind…”, “I’m planning to…”). I think Chunking Chinese does a good job at this.
- Focus on how phrases are actually used, not just what they “mean.”
- Surround grammar with context – stories, conversations, etc. Make it real and authentic.
4. Transfer learning: build on what you know
In AI
AI is often pretrained on general language data, then fine-tuned on specific tasks. Without this fine-tuning, AI wouldn’t be able to produce meaningful output. ChatGPT 3.0, trained on massive amounts of data, was still struggling to produce a shopping list for example. You had to remind the model what a typical shopping list looks like, give hints and clues for it be able to produce the desired results. So without task-specific training it wasn’t much good. I think this also explains the acceleration of progress, when OpenAI opened ChatGPT to millions of users. It then started using that knowledge on such a scale that it started creating all kinds of new connections.
For humans
You can transfer knowledge from one language to another — especially between related languages. Chinese is not the perfect candidate for this – at least for many learners. But the richer your frame of reference, the easier to spot similarities. When determining the severeness of an insult in Chinese, you can link it to a similar word in English for example, or maybe you’ll find that it can be used in similar contexts and settings.
Human takeaway
So again, when trying to imitate AI learning strategies, you could:
- Leverage similarities between languages you know.
- Build on shared grammar, vocabulary, or sentence structure.
- Use contrastive analysis (to some extent) to accelerate learning. This can be useful for pronunciation. To give an example for people with a German language background: the consonant in ‘ich’ [ç] is quite similar to the Chinese consonant in ‘xi’ [ɕ]. Not identical, but close. Knowing this as a German speaker can be helpful.
5. Focus on frequency: learn the core first
In AI
Models learn high-frequency words and patterns first – because they appear the most in training data. In other words, AI becomes fluent in “high-frequency areas” first. And although by now, I something get the impression ChatGPT knows almost everything, it still might not be able to answer questions about a hardly known medical treatment for some obscure disease entirely correctly. It might use extremely low-frequency medical terms in the wrong way, precisely because they’re not commonly used.
For humans
It’s not so different for us humans. In fact, most learners of Chinese tend to focus on the 1000 most frequently used words or HSK vocabulary.
Human takeaway
So – quite obviously – we shouldn’t waste time on rare or academic words and instead:
- Focus on the top 1,000–2,000 most common words.
- Learn high-frequency expressions and sentence structures.
- Use frequency lists or tools like the CEFR vocabulary framework.
What AI does that humans can emulate (even partially)
| Trait | In AI | Takeaway for humans |
|---|---|---|
| Massive repetition | Processes millions of sentences | → Learn in small but consistent daily chunks |
| Zero ego or fear | Makes endless mistakes | → See mistakes as data, not failure |
| Unforgiving memory | Retains all frequent input | → Use spaced repetition tools (e.g., Anki) |
Summary — what can humans learn from AI language learning?
| AI strategy | Human learning tactic |
|---|---|
| Massive exposure | Read/listen constantly in the target language |
| Predictive processing | Practice guessing what comes next |
| Context-based learning | Learn through example-rich input, not rules alone |
| Transfer from other knowledge | Use related languages to your advantage |
| Frequency-first focus | Master high-frequency vocabulary and patterns |
Conclusion
So much for this little excursion into the world of AI. Of course, one might ask how meaningful it really is to try to learn from machines. When the steam train was invented, people didn’t start wondering whether they could learn to run faster by closely examining the locomotive’s mechanism either.
It’s also rather intimidating to compare your own brainpower to that of thousands of processors humming day and night. Just as the cabinetmaker of old couldn’t keep up with the pace of a furniture factory, the human brain is simply outperformed by AI.
Nevertheless, I believe we can retain the principles discussed here:
- Lots of comprehensible input
- Learn by predicting patterns
- Embedded learning, focus on ready-made chunks with context
- Build on what you already know, activate your frame of reference
- Focus on high-frequency vocabulary first
Thanks for reading! Any thoughts? Feel free let me know in the comments!
Affiliate links
Disclosure: These are affiliate links. They help me to support this blog, meaning, at no additional cost to you, I will earn a small commission if you click through and make a purchase.
