Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages

The Invisible Wall in the Machine

Imagine the internet as a giant, bustling library where every book, video, and conversation ever recorded is stored. This is the "training data" that Artificial Intelligence (AI) uses to learn how to speak, write, and solve problems. For a long time, this library was almost entirely filled with books written in English. Because the AI only had English books to read, it became an expert at English but struggled to understand anything else. This is the core problem of "low-resource languages": if a language isn't well-represented in the digital world, the AI built to help people using that language is born with a handicap.

To understand why this matters, we need to look at two simple ideas. First, Cognitive Load is like a backpack your brain carries. When you learn something new, your backpack gets heavier. If you have to learn a difficult math concept and translate it from a language you don't know well at the same time, your backpack gets so heavy it bursts, and you can't learn the math. Second, Tokenization is how computers chop up words into tiny pieces to understand them. Think of it like a puzzle: if the puzzle pieces for one language are cut perfectly, the picture is clear. If the pieces for another language are cut into tiny, messy shards, the computer has to work much harder to put the picture back together.

Why should anyone care? Because AI is being sold as a magical tutor that can teach anyone, anywhere, how to code or solve problems. But if that tutor was trained only on English books, speaks only English, and only works when you have a super-fast internet connection, it fails the very people who need it most: students in rural areas speaking languages like Bengali. This paper investigates why this "magic" tutor is actually broken for them, not because the technology is bad, but because the foundation it was built on was never designed for them.

---

The Silent Failure: Why AI Ignores Bengali Speakers

This paper, titled "Structural Silence," tells the story of how AI infrastructure accidentally builds a wall around speakers of underrepresented languages, using Bengali as a case study. Bengali is a massive language, spoken by about 285 million people—roughly 4% of the entire human population. It has a history of over 1,400 years and is an official language in two countries. You might think a language that big would be a superstar in the world of AI. But the authors found that instead of being a star, Bengali is being "silenced" by four specific structural failures that stack up like a house of cards.

1. The Empty Bookshelf (The Web Presence Gap)

Imagine trying to teach a robot to speak Bengali by giving it a library to read from. The problem is that the library is almost empty. Even though Bengali speakers make up 4% of the world, the language accounts for fewer than 0.5% of all the content on the internet. In contrast, English, which has a similar number of native speakers, makes up about 49.5% of the web.

The AI models are trained by "crawling" the web to grab text. Since there is so little Bengali text online, the AI gets very little practice. It's like trying to learn to play the violin by listening to one song, while your English-speaking friend gets to listen to a million. The authors point out that this isn't just about who has internet access; it's about who has been writing and posting content online for decades. The result? The AI starts its training with a massive disadvantage before it even learns a single word.

2. The Data Starvation (The Training Token Deficit)

Because there is so little Bengali on the web, the AI gets starved of "tokens." Tokens are the tiny chunks of text the computer uses to learn. The paper highlights a shocking gap: in major training datasets, there are 67 English tokens for every 1 Bengali token.

Think of it like a diet. If English gets to eat a feast of 67 plates of food, Bengali gets a single crumb. The authors note that even when Bengali is included in "multilingual" models (models that try to learn many languages), it still performs much worse than English. The data isn't just scarce; it's so scarce that the AI simply hasn't seen enough examples to learn the rules of the language properly.

3. The Broken Puzzle Pieces (The Tokenization Penalty)

Here is where things get tricky. Even if you gave the AI the same amount of Bengali text as English, it would still struggle. Why? Because of how the computer chops up the words.

English uses the Latin alphabet, which is relatively straightforward for computers to slice up. Bengali uses an "alphasyllabary" script, where letters combine with little marks (diacritics) and join together in complex ways. Standard computer tools, designed for English, chop Bengali words into tiny, messy fragments. The authors call this a "tokenization penalty."

Imagine you are trying to assemble a puzzle. For English, the pieces are big, clear shapes. For Bengali, the same picture is cut into thousands of tiny, jagged shards. The computer has to work much harder to figure out how the shards fit together. This means the AI has to "eat" even more data just to understand the same amount of meaning. The paper suggests that to get the same performance as English, Bengali models need substantially more data just to overcome this messy chopping problem.

4. The Cloud Trap (Connectivity Exclusion)

The final failure isn't about the AI's brain; it's about the AI's body. Most AI tools today live in the "cloud." This means you have to send your question over the internet to a giant server, which thinks about it and sends the answer back.

The paper points out a harsh reality: in rural Bangladesh, where many Bengali speakers live, individual internet penetration is only 36.5%, compared to 71.4% in cities. Furthermore, only 9.2% of households own a computer. If an AI tutor requires a constant, fast internet connection to work, it is functionally invisible to the rural student. It's like designing a library that only opens when you have a private jet to get there. The authors argue that this isn't just an inconvenience; it's a design choice that assumes everyone is already connected, effectively locking out the people who need the help the most.

The Double Burden: A Heavy Backpack

When you combine these four failures, the result is a "double burden" for the learner. Imagine a student in a rural village trying to learn computer programming.

1. The Language Barrier: The AI tutor explains the code in English (because it was trained mostly on English). The student has to translate the English explanation in their head while trying to understand the complex programming concept.
2. The Cognitive Overload: Their brain is trying to do two hard jobs at once: translate the language and learn the math. Their "backpack" (working memory) gets too heavy, and they stop understanding.

The paper cites research showing that when students learn technical concepts in a second language without support, they learn less and retain less. For a Bengali speaker, an AI tutor that speaks English isn't just less convenient; it's often impossible to use effectively.

What the Paper Says We Should Do

The authors are careful to say that this isn't just a "fix the code" problem. They argue that we can't just tweak the AI a little bit to make it work. The whole foundation is built on assumptions that exclude non-English speakers.

• Stop treating data as a side project: The paper suggests that building datasets for languages like Bengali shouldn't be seen as "preliminary work." It should be recognized as a major scientific contribution, just as important as inventing a new AI model.
• Go "Offline-First": Instead of assuming everyone has the internet, we should design AI tools that work on a phone or computer without needing a connection. This isn't a "lesser" version of AI; it's the only version that works for the majority of the world.
• Listen to Linguists: The paper calls for linguists to step in and explain why standard computer tools fail at languages like Bengali. They have the vocabulary to spot these structural flaws that engineers might miss.

The Bottom Line

This paper doesn't claim to have built a perfect AI for Bengali speakers. Instead, it acts like a detective, pointing out the four structural reasons why the current system is failing them. It suggests that the gap in performance isn't because Bengali is a "hard" language or because the students aren't smart. It's because the AI infrastructure was built with English in mind, from the books it read to the way it cuts up words, and finally, to the assumption that everyone has a fast internet connection. Until we fix these structural silences, the promise of AI as a universal teacher will remain a promise kept only for a few.