The multilingual gap is a governance gap

In February 2026, 13 frontier artificial intelligence companies signed the New Delhi Frontier AI Impact Commitments at the India AI Impact Summit, pledging to strengthen multilingual and contextual evaluations and to collaborate with local ecosystems to develop benchmarks for low-resource languages. 

It was the first time a major multilateral AI summit put multilingual evaluation at the center of a formal industry commitment. The question is straightforward: why did it take this long?

The honest answer is that most AI governance frameworks still assume the systems they regulate communicate in English. This is rarely stated outright, but it runs through nearly every layer of current governance infrastructure, how risk assessments are written, how compliance documentation is structured and how safety benchmarks are designed. 

Why English-only testing misses real risk

A systematic review of nearly 300 large language model safety publications from 2020-24 found the language gap in safety research is not only significant but widening, with even high-resource non-English languages receiving minimal attention. When organizations deploy the same foundation model across dozens of languages but evaluate it only in English, they are assessing a fraction of their actual risk surface. That is not a niche, technical problem. It is a compliance problem.

Think about where this gap shows up in practice. Most red-teaming exercises, toxicity classifiers and content moderation tools are built on English-language datasets. Models that pass safety benchmarks in English may produce harmful, inaccurate or culturally inappropriate outputs in other languages — not because they are designed to behave differently, but because the guardrails were never tested there.