Haitian Creole and AI: Why models get it wrong

Ask an AI assistant to translate a sentence into Haitian Creole. The result is often understandable, but a native speaker will notice immediately: the sentence "sounds" French, a word is misspelled, or an expression simply does not exist.

Creole is spoken by more than ten million people and has been an official language of Haiti since 1987. Yet, when faced with AI, it remains a second-class language. Here is why.

Few texts online

An AI model learns from texts published on the internet. However, Creole is primarily an oral language. Schools, administration, and much of the print press still operate mainly in French.

As a result, there are far fewer texts in Creole than in French or English. In natural language processing, this is known as a low-resource language. It is not a question of the number of speakers, but of available data.

An under-applied orthography

Creole has had an official spelling system since the late 1970s. But many Haitians learned to write in French, and online writing mixes both systems. The model therefore sees several forms for the same word:

• mwen, m, mouin

• kounye a, konnya, kounya

To us, this is obvious. To a model, these are different words, each with even fewer examples.

The French trap

Most Creole vocabulary comes from French, and that is precisely what traps the models. They have seen billions of sentences in French and only a few fragments in Creole. When they hesitate, they pull toward French, even though the grammar is very different:

• The determiner is placed after the noun: liv la, machin nan, pitit mwen an.

• The verb is not conjugated: small particles indicate tense, as in mwen te manje, m ap manje, m pral manje.

• li does not mark gender: it means "he" or "she," and the model has to guess when translating.

Confusion between Creoles

Haitian Creole is not the only French-based Creole. There are also those of Guadeloupe, Martinique, Réunion, or Mauritius. Models sometimes mix them up and produce a sentence that is correct in none of them.

A hidden cost

A model does not read words, but tokens—that is, pieces of words learned mostly from dominant languages. A Creole word is often broken down into more pieces than an English word. The same sentence therefore costs more to process, and each piece carries less meaning.

What we can do

Large corporations will not solve the problem on their own. The quality of models also depends on what we produce.

1. Write in Creole online, using the official spelling.

2. Create open datasets, for example Creole–French texts verified by native speakers.

3. Report errors to the AI tools we use.

4. Train Haitian developers in language processing.

An AI only speaks languages well if a lot has been written in them. If Creole is poorly served today, it is due to a lack of data, and data can be built.