Originally posted at LinkedIn.  

By Holly Zanville, Founder/Director, Learn & Work Ecosystem Library 

Artificial intelligence is often described as global and multilingual. But what if the system underneath isn’t?

Today’s AI can respond in dozens of languages. Yet much of it is built on English—trained on English data, shaped by English patterns, and often processing meaning through English before producing an answer.

That gap between what we see and how the system actually works is what’s increasingly being called the “English Machine”—and it has important implications for how knowledge is interpreted and understood.

When people say AI “thinks,” they are using shorthand for how AI systems process language and generate responses—not human cognition. But the shorthand is useful, because it points to something real: AI doesn’t just process language, it inherits the structure of the languages it is trained on.

So, what’s the issue? If AI can translate between languages, isn’t this just a technical challenge that will improve over time?

The issue is more than translation—it’s about how meaning is formed.

Think about how human authors work. A novel or a poem is written in one language, shaped by its vocabulary, rhythm, and cultural context. When that work is translated to other languages though, there is always a concern: can the translation fully capture the meaning, tone, and nuance of the original? This has long been especially true for poetry, where word choice, sound, and structure carry deep significance.

A similar dynamic is now emerging with AI—but at scale.

When AI systems rely heavily on English, questions are often interpreted through English-language assumptions. Concepts may be reshaped to fit English structures. Answers may reflect English-speaking contexts, even when that’s not the user’s reality. In some cases, AI doesn’t just translate words, it reshapes meaning to fit English patterns.

To understand why this matters, it helps to step back and think about what a language actually is. A language is not just a set of words—it is a system for organizing meaning. It shapes how relationships are described, how time and space are understood, and what distinctions are important enough to name.

When AI systems are trained primarily on one language, they don’t just learn vocabulary—they learn patterns for how meaning is structured in that language. Over time, those patterns influence how questions are interpreted and how answers are generated. In that sense, AI systems are not language-neutral; they are shaped by the dominant language of their training data.

Languages don’t just use different words—they organize knowledge in different ways. For example, some languages use very specific “kinship” terms to describe family relationships, distinguishing between relatives that English would group together under broad terms like “aunt” or “uncle.”

Others organize everyday communication around physical space, using directions like “north” or “uphill” instead of “left” and “right.”

Many languages also reflect cultural expectations about how knowledge is shared. In some communities, certain knowledge is held by elders, shared only in specific settings, or passed down orally rather than written. These patterns shape how meaning is expressed—and they don’t always translate neatly into English.

The roots of the English Machine reflect how the digital world developed. A large share of online content is in English, and many early AI systems were built in English-speaking environments. Research, data, and investment have been concentrated in a relatively small number of languages.

The result: AI tends to perform best in the languages it has seen the most. This then creates a feedback loop. More English data leads to better-performing tools. Better tools generate more English content. That content further strengthens performance in English, while other languages fall further behind.

So, what is lost in this process? AI can often describe linguistic and cultural differences but it may struggle to reason within them. Subtle meanings can be flattened. Cultural context can be reduced or misinterpreted. In some cases, AI may generate incomplete or incorrect information when data in a language is limited. For Indigenous and underrepresented languages, this is not just a technical gap—it raises questions about preservation, representation, and voice.

This is where the implications become practical across the learn-and-work ecosystem. If AI is becoming part of how people learn, search for opportunities, build resumes and profiles, and navigate careers, then language shapes who benefits. Access to AI-powered tools may be uneven across languages. Skills-based hiring systems may misinterpret or overlook experience.  Credentials and qualifications may lose meaning in translation. Local and community-based knowledge may become less visible. The bottom line: if the system processes the world primarily through one language, it may not fully recognize knowledge expressed in others.

The good news is this isn’t going unnoticed. Multilingual AI models are expanding beyond English. Community-led initiatives are building tools grounded in local language and knowledge systems. Collaborative networks are developing resources for underrepresented languages. National efforts are creating language data repositories. Practical steps include expanding digital content in more languages, providing local context to guide AI outputs, building community-controlled data systems, and developing language-specific or “sovereign” AI models (systems controlled and governed within a country). These approaches point in the same direction: AI needs to reflect more than one way of knowing.

The English Machine isn’t just a language issue—it’s a signal. As AI becomes part of the infrastructure for learning and work, we have to ask whose knowledge is being represented, whose language is shaping the system, and who might be left out without even realizing it?

If we don’t pay attention, we risk building systems that look global but operate through a narrow lens.

We often say AI will expand access to information. It can but only if the system recognizes the full range of human knowledge, not just the parts that fit easily into English. That’s the real challenge behind the English Machine—and worth paying attention to now, before these systems become even more embedded in how people learn, work, and move through opportunity.

____________

See Topic Brief: Language & AI: Understanding the English Machine | Learn & Work Ecosystem Library

See Definition: The English Machine | Learn & Work Ecosystem Library