AI is on the cusp of a global leap—but only if we fix how we collect data.
A July 2025 survey of 50+ multilingual large language models (MLLMs) just exposed something uncomfortable:
Multilingual models aren’t truly multilingual.
They’re biased. They’re misaligned. And most critically, they’re overwhelmingly built on English-centric training data.
For businesses operating across borders—or training AI to serve diverse populations—this isn’t just a technical detail. It’s a blind spot. A costly one.
Let’s break it down.
The 3 Core Challenges to Multilingual AI
The study, conducted by Beijing Foreign Studies University, highlights three systemic issues:
Language Imbalance
→ 70–80% of the training data across leading models is English. The rest? A fragmented afterthought.
Over 100 major world languages—spoken by millions—remain dramatically underrepresented or completely excluded.
So yes, your model might technically support Amharic or Bengali. But without data depth, it’s like giving someone a dictionary and calling them fluent.
Cross-Lingual Misalignment
→ Languages don’t just differ in grammar. They differ in worldview. Yet most models carry over assumptions from high-resource languages—especially English—when parsing meaning.
So, even when a language is included, the model’s internal representation—how it “understands” meaning—doesn’t hold up across cultures or language families.
Translation may happen. Alignment doesn’t.
This creates semantic fractures. Meaning is lost. Context is bent. Answers sound right but miss the point entirely.
Embedded Bias
The training corpora aren’t neutral. They bring baggage.
They’re laced with the assumptions, stereotypes, and ideologies of their dominant sources—mostly Western, mostly English.
And here’s the kicker: bias scales. Social, political, and cultural biases are amplified.
Left unchecked, it shows up in sentiment analysis, in healthcare recommendations, in hiring tools, in systems people rely on.
Not just inaccurate. Dangerous.
Your Global AI May Not Be Global—Even If Your Model Says So
This is a digital language divide.
Recall Axios’s warning back in 2023 that native speakers of the world's more than 7,000 other languages will be excluded from fully accessing and benefiting from the capabilities of mainstream models.
When your product depends on multilingual AI—whether it’s powering content, chatbots, diagnostics, or anything in between- these flaws don’t sit in the background. They shape how users experience your brand. How they trust it. How they return to it.
- You launch a new language. But engagement tanks because the model doesn’t “get” the tone.
- You expand to Southeast Asia. But your AI misclassifies sentiment, turning helpful into hostile.
- You work in multilingual compliance. But the model glosses over nuance in local legal terms—and opens up risk.
Final Thoughts: Fix the Data, Not Just the Model
- AI can’t be multilingual if the data isn’t. Collect language data from diverse speakers, dialects, and contexts—not just high-resource defaults.
- Scalable doesn’t mean universal. Partner with in-region linguists and cultural consultants to capture nuance that scraped datasets miss.
- Focus on intent-rich, real-world inputs. You need authentic, scenario-based data—the kind users actually generate in critical domains like healthcare and government.
If your AI doesn't understand all users, it’s not global—it’s gated.
So no, this isn’t about “localization later.”
It’s about representation now.
Recalibrate your data strategy. Not for scale.
For equity. For trust. For global relevance.
Because multilingual AI isn’t built at inference.
It’s built at collection.