Ask a chatbot the same question in English and Hindi, and it may answer both times.
What users do not always see is that the second answer may cost more to produce, take longer and rely on a thinner base of training data. AI can appear to treat every language equally, but the systems behind it often do not.
Only about 19% of the world’s population speaks English, according to estimates based on total speakers. Yet many AI systems were built, trained and optimized around English first.
For the rest of the world, the technology may still work, but it can work at a disadvantage.
Neha Heera, a senior engineering manager with 18 years of experience building large-scale mobile and AI systems, has worked on the technical challenges that appear when software is expanded into languages and scripts it was not originally designed to support.
Her essay, The Latin Default, argues that the language gap in AI is not only a translation problem. It is also a systems problem that starts before a model generates an answer.
The hidden cost of language
Most conversations about AI bias focus on the output, such as a stereotype in a caption or a skewed recommendation.
Language bias can be harder to see.
A Hindi speaker may receive a coherent answer, so the product appears to be working. But underneath that exchange, the system may need more tokens to process the same meaning, which can increase cost, slow response time and reduce how much information fits into the model’s working context.
That hidden penalty can affect users every time they ask a question, write a document, search for information or use an AI tool inside another app.
“When something works for the people who built it, nobody files a bug,” Heera said. “The failures that matter most are the ones your own team never hits, because you are not the user paying three times as much to ask the same question.”
Why tokenization matters
Before a language model processes text, a tokenizer breaks that text into smaller pieces called tokens.
Those tokens affect cost, speed and memory. If the same sentence requires more tokens in one language than another, the user experience can change even when the meaning is the same.
Many tokenizers learned their shortcuts from mostly English text. That can make English more efficient to process, while other scripts may be broken into more pieces.
Research has found that some non-Latin languages can require several times as many tokens as English to express the same meaning.
The model is not intentionally charging one group more than another. It is responding to the way the system was built.
“A tokenizer trained mostly on English has an accent, and that accent decides who gets efficient service and who does not,” Heera said. “You can hide it behind a clean interface, but the meter is still running faster for half the planet.”
How the gap affects businesses
The token gap becomes more than a technical detail when it turns into cost.
In a breakdown of why AI costs more in Hindi, Arabic, and Thai than in English, Heera examined how language inefficiency can affect businesses that serve large non-English user bases.
Some workloads may require three to four times as many tokens in certain languages compared with English for similar use. That can affect the cost of running a product, the speed of responses and how much information a model can handle at one time.
For a startup serving users in India, the Middle East or Southeast Asia, that can mean paying more to offer the same kind of AI feature that an English-first company provides.
A company can absorb the added cost, limit features or avoid full support for certain languages. None of those choices is neutral for users.
“Two founders with the same idea are not competing on a level field if one of them pays a language tax the other never sees,” Heera said. “We like to say technology is a great equalizer. This particular piece of it is doing the opposite, very efficiently.”
Why English became the default
The imbalance did not come from one decision.
Large language models learned from the internet, and the internet is not evenly distributed across languages. One estimate says about 54% of all web content is written in English, far above the share of people who speak it.
Systems trained on that data can inherit the imbalance. English becomes the center of the model’s experience, while other languages are treated as edge cases.
Correcting that requires deliberate engineering. It also requires testing AI systems in the languages users actually speak, not only the language most common inside the company building the product.
Heera has worked on that kind of challenge before. Years before large language models became widely used, she helped bring support for five Indian languages to Kindle devices for the India launch. That work included support for Hindi, Tamil, Marathi, Gujarati and Malayalam on devices with tight storage and performance limits.
“On a Kindle, you cannot wave your hands about supporting a script,” Heera said. “Either the letters join correctly or they do not, on a device with kilobytes to spare. That kind of constraint teaches you respect.”
What could improve
The language gap in AI is not permanent.
Tokenizers can be rebuilt to treat scripts more evenly. Models can be trained on deeper non-English data. Products can be measured against benchmarks in more than one language before they ship.
Some teams already build tokenizers tuned for specific languages, which can reduce token counts and improve efficiency for those scripts.
Heera is also a peer reviewer for the 2026 ACM Conference on Conversational User Interfaces, where researchers study how people interact with machines through conversation. That work keeps her close to one of the places where language gaps show up first.
Her message to engineers is straightforward: measure cost and quality in the languages your users actually speak. Once the gap is visible, the fixes become easier to prioritize.
“The good news is that this is an engineering problem, which means it has engineering answers,” Heera said. “Nobody decided to make AI unfair to most of the world. It happened by default, one reasonable-looking shortcut at a time. Defaults can be changed, and that is the whole reason to keep pointing at this one until it is.”