Lost in Translation: How AI Fails to Respect Cultural Context

Large language models are touted as the future of personalized education, yet they stumble over a fundamental hurdle: cultural context. As these AI models adapt English math word problems into languages like Bengali, Hindi, and Italian, glaring inconsistencies emerge. Let's apply some rigor here. What's the point of teaching if the cultural context is lost in translation?

The Experiment

In a recent experiment, Claude Opus 4, GPT-4.1, and Gemini 2.5 Pro were tasked with adapting 60 English math problems into seven languages, including under-studied ones like Sindhi and Sicilian. The aim? To check if these adaptations maintain cultural diversity. What they're not telling you: cultural diversity took a backseat as models agreed on transformation type just 62.5% of the time, with specific substitutions aligning only 33.5% of the time.

Cultural Missteps

While it seems like a small blunder to use Bangladeshi taka for Indian Bengali students, it reveals a systematic issue: cross-cultural contamination. These AI models prioritize surface markers such as names and foods over deeper cultural structures. Color me skeptical, but an egg hunt adapted as an Eid activity doesn't scream cultural respect, does it?

Why It Matters

Let’s consider the broader implications. If a student's educational content fails to respect their cultural context, what message are we sending? the models preserve structural consistencies like grade levels, but that's cold comfort when the cultural facade collapses. The surface plausibility that makes these adaptations appear correct only obscures their deeper failures.

Education is about more than numbers and words. It's about context and culture, aspects these models fail to grasp. So, the question arises: should AI really be in the classroom if it can't understand the classroom?