Bridging the Low-Resource Language Gap: How LangQuest and Codex Enable AI Translation for Underserved Languages

by Frontier R&D Team

A recent report from the Center for News, Technology & Innovation (CNTI) reviewed over 55 studies on AI transcription and translation in journalism, revealing a critical divide: AI tools perform exceptionally well for English and other dominant languages, but struggle significantly with what researchers call "low-resource" languages.

The report highlights that English represents more than 50% of web content, which means mainstream language models are heavily biased toward languages with abundant digital resources. Even widely-spoken languages like Urdu are considered "low-resource" because they lack sufficient digitized text for training AI models.

The Problem

The CNTI working group found that AI translation tools introduce significant errors when working with low-resource languages. In one study from Tanzania, researchers discovered that 13% of translated sentences contained mistranslations or inaccuracies. For example, the English term "street food" was incorrectly translated into Kiswahili as "food of the road."

Beyond simple errors, AI tools struggle with:

  • Language ambiguity and cultural nuance — Missing context that human translators understand intuitively
  • Formality levels — Languages like Korean and Japanese have strict formality rules that AI often mishandles
  • Inherent biases — Training data reflects existing inequalities, perpetuating gaps in access

This creates a vicious cycle: languages with less digital content get worse AI support, which means fewer resources are created in those languages, which further widens the gap.

The Solution: LangQuest and Codex

At Frontier R&D, we're building tools specifically designed to break this cycle. Two of our projects address different aspects of the low-resource language challenge:

LangQuest: Documenting the Undocumented

LangQuest accelerates data collection for low-resource languages. The platform enables rapid, efficient gathering of language data that can be used to train AI translation models. By streamlining the collection workflow, LangQuest makes it possible to build translation tools for languages that were previously considered too resource-intensive to support.

LangQuest addresses the fundamental problem identified in the CNTI report: low-resource languages lack the digitized text needed to train AI models. By making data collection faster and more accessible—especially for field workers and language communities—LangQuest helps unlock these languages for AI, translation, and ultimately access to essential information.

Codex: Translation Without Prior Training

Codex Translation Editor takes a different approach: it enables AI-powered translation for languages that haven't been explicitly trained in the model. Codex leverages advanced AI techniques to work with languages it hasn't seen before, breaking down the barrier that requires massive training datasets.

This is particularly powerful when combined with LangQuest's data collection. As LangQuest documents a language, Codex can begin working with that language even before traditional training datasets are fully assembled. This creates a faster path from language documentation to translation capability.

Why This Matters

The CNTI report emphasizes that "hybrid translation"—where AI translations are reviewed by humans—is one of the most promising approaches for newsrooms. But hybrid translation still requires AI tools that can produce reasonable output in the first place. For low-resource languages, that foundation often doesn't exist.

LangQuest and Codex work together to build that foundation:

  • LangQuest documents languages and creates the data needed for AI training
  • Codex enables translation work even with limited training data
  • Together, they create a complete pipeline from language documentation to translation to publication

This integrated approach allows small teams to tackle translation projects that would have been impossible just a few years ago. For Bible translation, journalism, and access to essential information, this means communities can work in their own languages rather than being forced to use dominant languages.

Looking Forward

The low-resource language gap isn't just a technical challenge—it's a barrier to information access for millions of people. As the CNTI report shows, this affects everything from news coverage to public meeting transcripts to basic communication.

By combining rapid data collection (LangQuest) with AI translation that doesn't require massive training datasets (Codex), we're working to ensure that every language community has access to the translation tools they need. The goal isn't just better AI—it's breaking down the barriers that keep entire language communities from accessing essential information.

To learn more about how LangQuest and Codex are addressing the low-resource language challenge, visit langquest.org and codexeditor.app.

For a practitioner diagnostic—what “low-resource” means in AI, why a “supported” language still fails, and what to check before trusting a model—see Why AI Translation Fails for Low-Resource Languages.