Live

Technology Was Supposed to Connect Us. Instead, It Can Divide.

Machine translation (MT) tools like Google Translate, DeepL, Bing Translate, or even ChatGPT are everywhere. They give us instant access to content in dozens of languages and are built on powerful AI models — mostly neural networks that interpret context, predict meaning, and generate translations.

But these systems don’t just translate words. They also carry and reproduce the biases hidden in language — and that has consequences.

At the heart of many translation errors lies something deeper: discrimination embedded in training data. Neural machine translation (NMT) systems learn from real-world texts — texts that often reflect decades of gender, cultural, or class-based prejudice. These systems don’t just learn meaning — they memorize associations and repeat them, even when they’re outdated or harmful.

The result? A tool designed to make communication more inclusive can end up reinforcing inequality.

When AI Always Makes the Doctor a Man

The most obvious bias in machine translation is gender bias. Some languages, like Turkish, are gender-neutral. But MT systems often assign gender anyway — based on stereotypes.

Take the sentence “O bir doktor ve o bir hemşire”. A literal translation would be “They are a doctor and they are a nurse.” Yet many tools translate it as “He is a doctor and she is a nurse.” There’s no grammatical reason for this — just a reflection of sexist patterns in the data the AI was trained on.

This misrepresentation does real harm. It reinforces the idea that doctors are men and nurses are women, perpetuating what researchers call representational harm. Meanwhile, allocational harm refers to the poorer quality of services received by marginalized groups — in this case, women being served inaccurate or biased translations.

When Cultures Disappear in Translation

Gender isn’t the only issue. MT also struggles with cultural and identity-specific expressions.

Consider the Japanese sentence “お花見に行きましょう” — literally translated as “Let’s go see flowers.” But this misses the cultural meaning of hanami — a cherished Japanese tradition of gathering under cherry blossoms to celebrate spring, life, and renewal. The translation is technically accurate, but emotionally and culturally empty.

The same issue occurs with identity terms. In English or French, partner is often used to refer to a significant other without specifying gender — especially important in LGBTQIA+ contexts. Yet when translated into Polish, the word might be automatically masculinized, even if context suggests otherwise.

This is more than a linguistic error. It’s a form of identity erasure, where the speaker’s experience is distorted to fit heteronormative or binary assumptions. It’s systemic discrimination — disguised as “neutral” translation.

When Languages Don’t Count

Another issue? Minority languages get left behind. MT tools are most accurate for widely used language pairs like English–Spanish, often reaching over 90% accuracy. But for less common languages — Armenian, Corsican, or Romani — accuracy can drop to 55% or less.

The result: mistranslations, broken sentences, or content that’s just unreadable.

The reason? These systems rely on large, well-labeled text datasets. For many minority or regional languages, such data simply doesn’t exist. That leads to digital exclusion — entire communities cut off from information access.

Worse yet, the dominance of a few languages reinforces the idea that some cultures are “normal” or “standard,” while others are lesser, irrelevant, or strange.

When technology does real harm

These aren’t just technical glitches — they have real consequences.

In professional settings, mistranslations in legal, medical, or official documents can lead to serious misunderstandings. For nonbinary individuals or speakers of regional languages, MT can render their identities invisible.

On a social level, biased translations fuel stereotypes and deepen social divides. They promote cultural homogenization — a world where linguistic and identity diversity is seen as a problem, not a strength.

In international relations, the risks are even greater. One infamous case? The Japanese word mokusatsu was mistranslated during World War II as “not worthy of comment,” leading the U.S. to believe Japan was rejecting peace negotiations — a misstep that helped justify the bombing of Hiroshima.

Can we fix the machine?

So, is fair machine translation even possible?

Some solutions are already in the works. Tools like Gemini and DeepL now offer multiple gender options for certain languages. Others are expanding their training data to include minority languages and more inclusive vocabulary.

But the key is human oversight. Linguists, cultural experts, and identity advocates need to be involved in how these systems are designed and trained. We can’t rely on technology alone.

More importantly, we need a cultural and educational shift. Users must understand the limits of the tools they use every day. Critical thinking — comparing outputs, checking context, reporting errors — can help improve these tools and hold their creators accountable.

As long as training data reflects social inequalities, and as long as developers ignore diversity, NMT systems will continue to reproduce exclusion — even if they were meant to connect us.

Written by

Shape the conversation

Do you have anything to add to this story? Any ideas for interviews or angles we should explore? Let us know if you’d like to write a follow-up, a counterpoint, or share a similar story.

Write a counterargument article Write a related article

Same serie articles

Language and Identity: Why Minority Communities Need Their Language to Be Seen