Explaining in the language students actually speak

The language in the room is not in the textbook

A maths class in Delhi thinks in Hinglish. A science class in Nairobi thinks in Sheng and Kiswahili and English, shifting mid-sentence. A class in Lagos thinks in Pidgin and Yoruba and English. A class in Manila thinks in Taglish. The textbook is in formal English, or formal Hindi, or formal Kiswahili: a register that exists mostly in print and that students decode rather than read.

Good teachers already explain in the mixed language, out loud, because it works. What was not possible was writing in it: a worksheet, a worked example, a revision sheet in the register students actually think in. The model can do this, and it is one of its more surprising strengths, because the mixed languages are all over the internet (in forums, comments, messages, subtitles) and it has read them.

The mechanism, and why “in Hindi” fails

Ask for “an explanation in Hindi” and the model produces the Hindi it has read most in educational text, which is the formal, Sanskritised register of textbooks and government material. It is correct Hindi. It is not what a 12-year-old in Delhi speaks, and the check question’s students find it harder than the English, because they have never heard anyone say it.

The fix is the same as for register anywhere: show it. Paste three or four sentences of the way your students actually talk (record a student explaining something and transcribe it, or write it yourself the way you would say it in class) and say “explain in this register, this mix of languages, these words for these things”. The output shifts immediately.

Explain why 1/3 is bigger than 1/4 the way I would say it to my class. Here is how I talk to them: “Dekho, jab hum ek roti ko 3 pieces mein kaat-te hain, toh har piece bada hota hai. Agar 4 mein kaato toh chhota. Samjhe?” Keep that mix. Short sentences. Use roti and pieces, not abstract nouns.

The result reads like a teacher talking, which is the point.

Bilingual glossaries

The other thing the model does well here is the glossary: subject terms in English, the local-language equivalent, and a one-line explanation in the mixed register. Ten terms per chapter. Students keep it in the front of their book and stop being stopped by vocabulary in the middle of a problem.

Ask for it as a plain list (the formats lesson) and give it your textbook’s terms, because the earlier lesson explained that it will otherwise pick a term from a different textbook tradition. For languages where there is no established equivalent, the glossary should say so, and use the English term with an explanation, rather than inventing a word.

Reading and writing in the mix

Some teachers worry that written Hinglish or Sheng in a worksheet teaches bad language. The evidence on bilingual education is fairly clear that explanation in the strongest language, or the mixed one, improves understanding of the content and does not harm acquisition of the formal register, which is taught as its own thing. A worked example in the mix is a bridge, not a replacement. The formal version follows, and lands better because the idea is already there.

Say this to parents if they ask. Some will, particularly the ones who see English as the route out, and they are right that English matters and wrong that a bridge to it costs anything.

Which mixes work

Well: Hinglish, Taglish, Spanglish, Franco-Arabic, Sheng, Nigerian Pidgin, Singlish, Urdu-English. Enough of each is written online for the model to have the pattern.

Less well: mixes involving a language low on the gradient from the previous lesson. Sheng works because Kiswahili and English are both large; a mix with a small language produces a mix in which the small-language half is unreliable. The back-translation check applies to the small-language sentences.

Speech, not just text

The mix is spoken more than written, and the model’s voice modes can now speak it, unevenly. For a class with one phone at the front, a short explanation played aloud in the students’ own register is a real thing, and a later lesson in this block is about setting up the free tools for it. Check the pronunciation before class; the voice may have learned the language from text and say it strangely.

The honest limit

The mixed language belongs to the students and it changes by year and by neighbourhood. The model’s version of Sheng is a few years old and from somewhere else. Students will notice a word that nobody says any more, and it will read as an adult trying too hard. That is fine if the teacher acknowledges it (“the machine’s Sheng is a bit old, fix it for me”) and it becomes a small language lesson in itself. What does not work is presenting the model’s mix as the class’s own. It is a draft of their language, and they are the editors.