Ask almost any modern AI to write something in Indonesian and it will. The grammar will mostly hold, the words will be real. But anyone who actually speaks the language can usually tell within a sentence or two that something is off. Producing Indonesian and speaking Indonesian are not the same thing.
Indonesian is not one register
How you talk changes with the situation, and Indonesian changes a lot. You message a friend with “lo” and “gue”. You speak to an older colleague with “saya” and “Bapak” or “Ibu”. You write to a ministry in full tata bahasa baku, where the form is almost as important as the content.
A model that picked up Indonesian mostly as a side effect of learning English tends to flatten all of this into one register, and then use it everywhere. It writes a birthday note like a press release, or a casual reply that is stiff and oddly formal. The grammar passes. The judgment does not.
We rarely speak one language at a time
Real Indonesian is full of code-switching. A single message can carry Indonesian, English, and a bit of Javanese or Sundanese, and everyone reads it without blinking. “Tolong di-follow up ya, deadline-nya mepet.”
Models trained on cleaner, more separated text often stumble here. They over-correct the mix back into formal Indonesian, or they treat a normal borrowing like “di-follow up” as a mistake to fix. What reads as natural to a person reads as noise to a model that never saw enough of it.
Context carries half the meaning
A lot of Indonesian lives in context rather than in the words themselves. Honorifics like Bapak, Ibu, Mas, and Mbak carry respect and relationship that have no clean English equivalent. References to local places, offices, holidays, and customs are assumed, not spelled out.
An AI that does not hold that context tends to be polite in the wrong direction, or to fill the gap with something invented. It will address a stranger too casually, or describe a local institution that does not quite exist, and do it with total confidence.
Where models usually go wrong
Three failures keep showing up:
- Wrong register. Too formal for a chat, too casual for a document.
- A translated feel. Sentences that are grammatical but read like English wearing Indonesian clothes.
- Confident local mistakes. A made-up rule, a wrong office name, a misremembered fact, all delivered as if certain.
One common cause is simple. When Indonesian is only a small share of a model’s training and evaluation, it can learn the outline of the language without enough of its everyday use.
What it takes to get it right
This is not something you fix with a clever prompt. It takes Indonesian text, gathered broadly and cleaned carefully, so the model sees how the language is actually used across registers and regions. It takes evaluation by people who speak the language, so “sounds fluent” is never mistaken for “is correct”. And it helps to build close to where the language is spoken.
That is most of what we spend our time on. We measure our models on Indonesian work rather than English leaderboards, as we wrote in why we built our own benchmark, and we build with Indonesian usage, evaluation, and context in mind rather than treating the language as something added at the end. Speaking Indonesian well is not a feature you add late. For us it is the point.