Translation, and the back-translation check

Why some languages translate well and others do not

The reliability rule from the first block applies to languages the way it applies to facts: the model is good at what it has read a great deal of. It has read enormous amounts of English, Spanish, French, Portuguese, German, Chinese, Japanese, Arabic, Hindi and Russian. It has read a fair amount of Indonesian, Turkish, Vietnamese, Bengali, Urdu, Kiswahili, Tamil and Marathi. It has read some Yoruba, Hausa, Igbo, Amharic, Somali, Nepali, Sinhala, Khmer. It has read very little of the thousands of languages with a few million speakers or fewer, and almost nothing of most of them.

Translation quality falls off along that gradient. For the first group it is close to a competent human translator for ordinary text. For the second, it is usable with checking. For the third, it produces fluent sentences that may carry the wrong term, the wrong register, or a meaning that has drifted. For the fourth, it produces something that looks like the language and may not be.

The danger is that the output looks the same all the way down. A fluent Amharic sentence from the model looks as confident as a fluent Spanish one. You cannot tell from the surface which gradient you are on. You can tell from the check.

The back-translation check

Translate the worksheet into the target language. Open a fresh chat (a fresh one, so the model has no memory of the original) and paste the translation with: “Translate this into English.” Compare the result with your original.

Where the back-translation matches in meaning, the forward translation was probably sound. Where it has drifted (a “force” that became a “power”, a “solution” that became a “mixture”, a question that changed its meaning) the forward translation is wrong at that point, and it is usually wrong in exactly the way a fluent speaker would spot and you cannot.

This is not proof. A translation can go wrong and come back looking right, if the same error is made in both directions. But it catches most of the failures, it costs one extra minute, and it tells you which gradient you are on: a back-translation that matches closely is a large-language result; one that drifts everywhere means the model does not really have this language, and a speaker must check before students see it.

Terms are where it breaks

Ordinary sentences translate better than subject vocabulary. The model may translate “photosynthesis” into a coined word, a transliteration of the English, or the term used in a neighbouring country’s textbooks rather than yours. Students who have learned the term one way are confused by another.

So give it the terms. Paste a short glossary from your textbook (the ten subject words in this chapter, in both languages) and say “use exactly these terms”. It will. Without the glossary it guesses, and the guess is often the term from whichever textbook corpus it read most, which is not yours.

Register and variety

A language is not one thing. Hindi from the model tends toward a formal, Sanskrit-heavy register that reads like a government notice to a Class 6 student. Arabic comes out as Modern Standard when students speak a dialect. Portuguese may be Brazilian when the class is in Mozambique. Kiswahili may be the Tanzanian standard when the class is in Kenya.

Paste a paragraph from your own textbook and say “match this register and variety”. The model shifts. Without the example it defaults to the variety it has read most, which is often not the one in the room.

Exam papers and anything that carries weight

Do not translate an exam paper, a report to a parent, or a safeguarding document with a model alone. The back-translation check is a check for worksheets and practice material. For anything where a wrong word has a cost, a speaker of the language reads the translation before it is used, and the model’s role is to produce the draft that saves that speaker an hour. That is a large saving, and it is the correct division of labour.

Translating the other way

Students write in their home language; you read it in yours. The model does this well for large languages and it gives you access to what a student can express in the language they think in, which is often far more than they can express in English. For marking, the same rule as everywhere: feedback, yes; a grade on the translation, no, because the translation may have made the argument better or worse than the original.

The honest limit

For the languages at the bottom of the gradient, no amount of prompting makes the model a translator, because it has not read enough to be one. The last lesson in this block is about what to do then. For everything above that, the back-translation check and the glossary get you a usable draft, and “usable draft, checked by a speaker” is a great deal more than most multilingual classrooms had before.