Why AI Sounds Smarter in English
By Chatday Editorial Team ·
Ask an AI something in Spanish, Polish or Turkish and the answer arrives fast and fluent. Ask the exact same thing in English and something subtle happens. The reply is a little sharper. Better organised. More sure of itself.
You are not imagining it, and it is not your writing. The gap is real, it has been measured, and it comes down to something simple: these models did most of their reading in English.
The internet these models read is mostly English
Language models learn by reading enormous amounts of text scraped from the web. And the web they read is heavily unbalanced.
One review of the big web archives used for this training looked at a late-2023 snapshot of Common Crawl, a public crawl of billions of pages. English made up about 44% of it. The second language on the list, German, was about 5%. Everything else split what was left.
Now compare that with the actual world. Language databases put English speakers at roughly 1.5 billion, counting everyone who learned it as a second language, which is something like one person in six. So a language used by a sixth of humanity supplies close to half the reading material.
The model does not decide to favour English. It just gets far more practice in it. Every idiom, every awkward phrasing, every weird edge case shows up thousands of times more often in English than in Catalan or Vietnamese, and practice is what matters.
Your AI is probably thinking in English before it replies
Here is the part that surprises people. When you write in your language, the model does not necessarily reason in it.
A 2025 study probed what actually happens inside multilingual models while they answer. The researchers found that these systems make their key decisions in a representation space closest to English, no matter what language went in or comes out. In their words, the model produces something English-shaped for the meaningful words first, then converts it into the target language.
Think of a brilliant colleague who grew up reading in English and now works in your office. They understand you perfectly. But somewhere in the middle, the thinking is happening in their first language, and the version you get back has passed through a translation step.
Most of the time you never notice. You notice when the joke lands flat, when an idiom comes out translated word for word, or when the tone is technically correct and socially wrong.
Why your language can cost more and run out of room faster
There is a second, more mechanical reason, and it has nothing to do with intelligence.
Before a model reads your text, it chops it into small pieces called tokens. Roughly, a token is a chunk of a word. That chopping is tuned mostly on English, so English gets efficient pieces while other languages get split into many more of them.
Research on this found the same sentence can take up to 15 times more tokens in some languages than in English. Languages that do not use the Latin alphabet tend to fare worst.
Three things follow from that, and you can feel all of them:
- You hit limits sooner. The AI’s working memory is counted in tokens, so the same document uses more of it in your language.
- It can cost more. Where a service charges by token, the same task costs more in a language that splits into more pieces.
- It can feel slower. More pieces means more for the model to process.
None of this is a punishment. It is a side effect of a design tuned on English text, and it quietly shapes your experience.
So how big is the gap, really?
Big enough to matter, and very uneven. A multilingual benchmark called MMLU-ProX put the same set of nearly 12,000 exam-style questions into 29 languages, which makes for a clean comparison: same questions, different language, same model.
The pattern was consistent. Models did well in the widely published languages and fell off in the ones with less material online, with gaps reaching about 24 percentage points between the strongest and the weakest. On a test where a good score is in the 70s, losing 24 points is the difference between a confident expert and a shaky student.
| Where your language sits | Roughly who this is | What to expect |
|---|---|---|
| Very well represented | English | Sharpest reasoning, best idiom, most reliable formatting |
| Well represented | Spanish, German, French, Portuguese, Chinese, Japanese | Close to English for most everyday tasks, occasional stiff phrasing |
| Moderately represented | Many European and Asian languages with smaller online footprints | Good answers, more translated-sounding tone, more slips on local detail |
| Thinly represented | Smaller languages and most African languages | Noticeably weaker reasoning, more confident mistakes, check everything |
Two honest caveats. Your everyday questions are much easier than an exam, so the gap you feel is usually smaller than the gap on paper. And these numbers move every few months, which brings us to the good news.
The gap is closing, on purpose
Labs figured out that training on more languages is not charity, it is an upgrade.
A 2026 paper on this tested it directly and found that adding languages during training improves results across the board. Low-resource languages gain a lot, high-resource ones hold steady rather than degrade, and adding even a single non-English language improved the model’s English too. The authors called English-only training largely suboptimal.
That is why each generation of models feels a bit more natural in your language than the last. It is also why the practical advice below is worth redoing every few months. The answer to “which AI is best in my language” keeps changing.
Four ways to get better answers in your language
1. Name the language and the register
Do not just write in your language and hope. Say it. Something like: answer in Portuguese, in a natural everyday register, the way a person would actually speak.
That last part matters most. Left alone, models drift towards a formal, translated-sounding voice, because that is what most of the non-English text on the internet looks like.
2. Show it how you write
Paste two or three sentences you wrote yourself and ask it to match that voice. This works far better than describing the tone you want, and it is the single fastest fix for the artificial feeling in AI text in any language.
3. For hard questions, separate thinking from answering
If a question is genuinely difficult, try asking it in English and requesting the answer in your language. Given what we know about the English-shaped middle step, this sometimes helps.
Sometimes. Not always. It costs you one extra message to find out, and for anything where the reasoning is the hard part it is worth the test.
4. Change the model, not the prompt
This is the one most people never try, and it is usually the winner. Different labs train on different mixes of language, so their strengths in your language differ more than their marketing suggests. The same prompt can land beautifully in one model and stiffly in another, which is a specific case of the broader reason AI models give you different answers.
Gemini 3.1 Pro and Claude Sonnet 5 are good starting points, but do not take our word for it. Run your own prompt through a few and see which one sounds like a person from where you live.
The three-prompt test for any new model
When you want to know whether a model is any good in your language, benchmarks will not tell you. Thirty seconds of your own testing will. Here is what we look at, and why each one catches something different:
- An idiom or a joke. Ask it to explain a saying only locals use. This catches the models that are translating rather than understanding.
- A local detail. Ask something practical and specific to where you live, then verify the answer. This catches confident invention, which is the most expensive failure mode.
- A short piece of writing in your voice. A message to a colleague, in your normal register. This catches stiffness, wrong formality, and that unmistakable translated feeling.
Run those three in two or three models. You will have a favourite within a couple of minutes, and it may not be the best known one.
Where none of this helps
To be clear about the limits: prompting cannot add knowledge the model never had.
If your language has a thin presence online, the gap is in the training, and no clever instruction closes it. Local rules, regional legal details and small-town specifics are where models invent most confidently, in every language, and where you should verify before acting.
For anything with real stakes, a contract, a medical letter, an official form, treat AI output as a first draft and get a human who speaks the language to read it. And when the job is straight translation rather than conversation, a dedicated translation tool built for it beats a chat window. We walked through that workflow in how to translate anything with AI.
The short version
AI is not smarter in English. It is just better read in English, and better practised at it, and the machinery underneath was tuned for it. That gap is shrinking with every generation, faster than most people realise, because the labs discovered that teaching a model more languages makes it better at all of them.
Until it closes, the fix is not to give up your language. It is to be specific about what you want, show the model your voice, and stop assuming the first AI you tried is the one that speaks your language best.