Why AI Works Worse in Languages Other Than English

Every launch demo is in English. So is almost every benchmark table, most of the documentation, and the overwhelming majority of the text these models learned from. If you speak English, that invisible fact works in your favour so consistently that you never notice it. If you don’t — and most of the planet doesn’t — you are using a quietly different, and quietly worse, product: one that costs you more per word, answers you less accurately, and protects you less well. Same logo, same price, same confident tone. Not the same tool.

This is not a conspiracy, and it is not the model being rude about your language. It is a structural consequence of how large language models are built, and it shows up in three separate places: in what you pay, in what you get, and in how safe you are. Worth taking each in turn, because the mechanism is more interesting — and more fixable in some places than others — than “AI is just better at English.”

The tax you pay before you type: tokenization

Start with money, because it is the least intuitive and the easiest to prove. Language models do not read words. They read tokens — the small chunks a piece of software called a tokenizer slices your text into before the model ever sees it. A token might be a whole word, a fragment of one, or a single character. And crucially, you are billed by the token, and the model’s memory — its context window — is measured in tokens too.

Here is the catch. Tokenizers are trained, and they are trained mostly on English. So English gets sliced efficiently: common words become single tokens, and a sentence turns into a tidy, short sequence. Feed the same sentence in another language through the same tokenizer and it shatters into far more pieces. One cost audit of tokenization put numbers on what speakers of those languages have long suspected: the same content routinely costs several times more tokens in other languages — on the order of three times as many for Arabic, and considerably more for many languages written in non-Latin scripts, where the tokenizer can fall back to encoding almost character by character.

Three times the tokens means three things at once, none of them good. It means roughly three times the API bill for the identical request. It means you hit the context limit three times sooner, so the document you can paste in, or the conversation the model can remember, is a fraction of what an English speaker enjoys. And, more subtly, it means the model’s internal representation of your text is more fragmented, which tends to make its answers slightly worse before it has even started reasoning. This is the same token economics we wrote about in the hidden cost of AI tokens — except that here the meter runs faster for you purely because of the language you were born into.

The quality gap: it learned mostly English

The second layer is competence, and it comes straight from the training data. A model’s abilities are downstream of what it read, and what it read was the web — which is heavily, disproportionately English, with a long tail of other languages that thins out fast. The industry has a slightly euphemistic term for the languages on that tail: “low-resource,” meaning not few speakers but little digitised text. Plenty of languages spoken by tens or hundreds of millions of people are, by this measure, low-resource, because their speakers’ lives are not conducted on the English-language internet.

A model that has seen a trillion words of English and a comparatively thin slice of, say, Yoruba or Burmese will be correspondingly more fluent, more accurate and more knowledgeable in the former. In practice, users of lower-resource languages report the tell-tale symptoms: more hallucination, idiom that is subtly or comically off, reasoning that degrades faster on hard problems, and a habit of appearing to “think” in English and translate back, losing nuance on the way. Practitioner analyses of the multilingual performance gap find the same pattern across tasks: quality falls as you move away from the high-resource languages the model saw most of.

A concrete example makes the gap tangible. Ask a model for help with a legal form, a school report or a medical leaflet in a high-resource language and it will usually produce something idiomatic and broadly reliable. Ask for the same in a lower-resource language and you are more likely to get stilted phrasing, invented terminology, or a confident answer that quietly imports facts from the English-speaking world — the wrong tax office, the wrong emergency number, the wrong cultural assumption baked in as if it were universal. The failure is rarely a blank refusal; it is a fluent answer that is subtly, unverifiably wrong, which is the hardest kind to catch precisely because it arrives with the same confident polish as the good English one. A user who can’t easily cross-check is exactly the user least equipped to notice.

What makes this gap so persistent is that the people building the models often cannot see it. Benchmarks — the scoreboards labs use to decide whether a model is good — are overwhelmingly English-first. A model can post a triumphant benchmark sweep while quietly underperforming for most of the world’s speakers, because the test never asked. Efforts like the LINGOLY reasoning benchmark, which probes low-resource and even extinct languages, exist precisely to surface the drop that the standard leaderboards paper over. It is another reason to treat a headline score with suspicion, the same way we argued that benchmarks mean less than you think: a number that averages over an English-shaped test set is not measuring the experience of a Swahili or Tagalog speaker at all.

There is a cruel second-order effect here, too. The tools built to detect AI writing inherit the same English bias in reverse: they disproportionately flag text written by non-native English speakers as machine-generated, because fluent-but-non-idiomatic writing pattern-matches to their idea of “AI.” We covered that failure in our piece on whether AI detectors actually work. The through-line is the same: an English-centre of gravity, and everyone else absorbing the error.

The dangerous gap: safety is English-shaped too

The third layer is the one that ought to worry the labs most, because it turns an inequity into a vulnerability. The guardrails that make a model refuse dangerous requests are not laws of physics; they are trained behaviours, taught with examples. And those examples, like everything else, are concentrated in English. So the safety net is densest exactly where the model is already strongest, and thinnest where it is weakest.

Researchers at Brown University demonstrated this with uncomfortable clarity in a study bluntly titled Low-Resource Languages Jailbreak GPT-4. They took requests the model would refuse in English — the classic “how do I build a weapon” category — and simply translated them into low-resource languages such as Zulu or Scots Gaelic. The refusal largely evaporated. Combining several such languages, they raised the rate of harmful responses from under 1% in English to roughly 79%. In one widely reported example, researchers got GPT-4 to explain how to make explosives in Scots Gaelic, a language with a few tens of thousands of speakers.

Read that finding from the other side and it is not really about attackers at all. It means a legitimate user working in a low-resource language is relying on a safety system that is, for them, far more porous — more likely to produce unsafe or simply wrong output, with fewer of the checks an English speaker gets for free. The guardrail is real, but it is unevenly hung. To be fair, this specific study is from the GPT-4 era, and the labs have since invested in multilingual safety training and red-teaming; newer models are harder to fool this way. But the underlying asymmetry — safety effort following the biggest, most English-speaking markets — is not something a single patch removes.

To be fair: it is getting better, and it is hard

None of this is evidence of ill intent, and it would be reverse-hype to imply the labs are indifferent. Building a model that is genuinely excellent across thousands of languages is a hard, expensive problem, and there is real work going into it: more multilingual training corpora, tokenizers designed to treat scripts more evenly, dedicated non-English evaluations, and safety datasets in more languages. Each model generation does narrow the gap, and for the highest-resource non-English languages — the Spanishes and Mandarins of the world — the gap is already small.

The honest position is that this is a structural bias rather than a switch someone forgot to flip. It flows from an English-dominated internet, from benchmarks written in English by English speakers, and from the plain commercial logic of spending your safety and quality budget where your loudest, best-paying users are. Those forces are not malicious, but they are durable, and they all point the same way. Which is exactly why it is worth naming: bias that comes from structure rather than intent is the kind that persists quietly, because no one has to decide to keep it. As we argued in algorithmic bias is not a glitch, the defaults are the politics.

It is worth being precise about who carries that cost, because “it’s improving” can quietly become an excuse. The languages furthest behind belong disproportionately to smaller communities and to the global south — to people least able to absorb a token surcharge, shop around for a better model, or fall back on an English version they never learned. When a disadvantage is structural, the people on the wrong side of it are not a rounding error to be tidied up in a later release; they are most of the world’s speakers, handed the same branded product as everyone else and quietly given less of it. Saying so is not doom-mongering. It is the difference between a bug the makers are racing to fix and a cost they have decided somebody else can carry for a while.

What to do if you use AI in another language

Knowing the mechanism turns a vague frustration into a set of concrete habits. If you work with these tools outside English, a few adjustments genuinely help:

• Expect the meter to run faster, and budget for it. If you are on a paid API, the same task can cost several times more in your language purely because of tokenization. It is not you doing something wrong; price accordingly, and prefer shorter prompts where you can.
• Verify more, not less. The lower a language’s resource level, the higher the hallucination risk. Treat confident answers in low-resource languages as first drafts to check, especially for names, dates, quotes and anything legal or medical.
• For hard reasoning, consider working in English and translating. It is an unfair workaround, but a real one: many models reason more reliably in English, so posing a difficult problem in English and translating the result can beat asking directly in a low-resource language. Keep the original for nuance.
• Don’t assume the guardrails travel. A model that behaves responsibly in English may be markedly less reliable — and less safe — in another language. That cuts both ways: it may refuse things it shouldn’t, or produce things it shouldn’t. Either way, don’t lend it more trust than its English self has earned.
• Push for multilingual evidence. When a vendor quotes a benchmark, ask what it looks like in the languages you actually use. The answer, or the absence of one, tells you how much they have thought about you.

The promise of these tools is universal: type in any language and get a capable assistant. The reality, for now, is that the assistant is fluent, cheap and well-behaved in one language and progressively less so in every other — and it rarely mentions the difference. That gap is narrowing, and the engineering behind closing it is genuinely hard. But until it closes, the majority of the world using AI in its own tongue deserves to know it is paying more for less, and to use the tool with its eyes open rather than on the trust the English demo was designed to earn.