Lesson 1 of 8 · 8 min read · last verified 2026-08-26
Why English gets the better answer
In this lesson you will:
- Explain why model quality varies by language
- Recognise the practical symptoms of a lower-resource language
Every lesson in this curriculum so far has assumed you were working in English. Most people on earth are not, and this module is about what changes when you are not.
It starts with an uncomfortable fact, stated plainly rather than apologised for.
The gap is real
Models generally answer better in English than in other languages. Not marginally, and not uniformly.
The cause is simple. These systems learn from text, and the text available to train on is enormously skewed. English is heavily over-represented relative to the number of people who speak it, because of who put the internet together and what got digitised first.
A model has therefore seen far more English — more phrasing, more subject matter, more argument, more of the world discussed through English. So it is better at English, and also better at everything else, in English.
It is not one gap, it is a gradient
Treating this as “English versus the rest” is wrong and unhelpful.
High-resource languages — Spanish, French, German, Mandarin, Portuguese, Russian, Japanese. Large digitised corpora. The gap is real but often modest, and narrowing.
Mid-resource — Hindi, Bengali, Turkish, Vietnamese, Indonesian, Arabic, Swahili. Hundreds of millions of speakers, far less digitised text than that would suggest. Usable, with more errors.
Low-resource — thousands of languages including many with tens of millions of speakers: Bhojpuri, Sylheti, Hausa, Quechua, most of Africa’s and India’s languages. Output ranges from stilted to unreliable to unusable.
Notice that speaker count and resource level are only loosely related. That mismatch is the injustice in this, and it is worth naming: the languages served worst are not the smallest ones.
What the gap looks like in practice
Recognising the symptoms matters more than knowing the theory, because they tell you when to switch strategies.
Stiff, translated-sounding phrasing. Grammatically fine, nobody talks like that.
More factual errors. Not only about language — about the world. Ask about local law, local history, local institutions in a lower-resource language and error rates rise noticeably.
Weaker instruction-following. Your format request, your word limit, your “do not include X” get dropped more often. Complex instructions degrade first.
Register collapse. Formality distinctions your language marks carefully come out flattened or wrong — a real problem in languages where getting this wrong is rude rather than merely odd.
Silent switching. It answers in English, or drifts into it midway.
It is improving, unevenly
Worth saying, because the picture is not static.
Model providers have invested in multilingual capability and each generation is better. There are also excellent region-specific models, several from Indian and African research groups, which outperform general models on their target languages.
But improvement follows commercial attention, which follows markets. Languages with buying power improve fastest. That is a prediction about the next few years, not a complaint — and it means the answer to “how good is this in my language?” changes, and is worth re-testing rather than assuming.
L2 is how you test it for yourself, which is the only measurement that applies to your actual work.
None of this is your fault
Stated because it needs stating.
If output in your language is worse, that is a property of what the system was trained on. It is not a signal that you should switch to English to be taken seriously, or that your language is less suited to the work.
The rest of this module is about getting good results anyway — which is frequently possible, and sometimes involves using English deliberately as a tool rather than as a concession.
Try it now (6 minutes)
Ask the same substantive question twice — once in English, once in your language. Something you can judge, ideally about your own field or region.
Compare not just the language quality but the facts. That difference is the one people miss, and it is the one that matters.
Check your understanding
Recap
Models answer better in English because far more English text existed to learn from, and the gap is a gradient rather than a binary — tracking digitised text rather than speaker numbers. Learn the symptoms: stiff phrasing, more factual errors, weaker instruction-following, flattened register and silent switching. It is improving unevenly, it is worth re-testing, and it is not your fault.
🗂 3 flashcards from this lesson join your daily review.