AI models understand Haitian Creole culture noticeably worse than they understand French - even though both are spoken in the same country.
A new study offers the first systematic evaluation of cultural awareness in large language models for Haitian Creole, a language spoken by millions but barely represented in digital training data. Researchers built a benchmark of culturally specific prompts, written by native speakers, and tested models on a text-infilling task across four dimensions: specificity, bias, diversity, and variation. The results showed a clear gap between Haitian Creole and French performance, with Haitian Creole outputs more inconsistent across topics and more prone to French-language interference bleeding into the responses. When models generated stories, Haitian characters were repeatedly cast in narratives of hardship and resilience.
This matters because it complicates the usual story about AI's language gaps. The industry tends to frame low-resource languages as a data-volume problem: feed the model more text, close the gap. This study suggests something subtler is going on. Haitian Creole shares deep linguistic roots with French, yet models still can't reliably separate the two cultures - and when they try to render Haitian identity in fiction, they default to a narrow, if well-meaning, stereotype.
It's a useful reminder that a model fluent in a language isn't the same as a model that understands the people who speak it. Resilience-through-hardship is the kind of characterization that reads as respectful on the surface, which makes it easier to miss as a bias - and harder to fix than a simple vocabulary gap.