English is only half of AI use
Most of what we know about how well AI works comes from English.
The rankings that compare AI models use questions written in English or translated from it.4 5 One widely used open model learned from a mix with only 8% in other languages.3 About half of the conversations in a large public record of ChatGPT use are in other languages. Conclusions drawn from English results therefore describe half of real use at most.
English and dozens of other languages share AI conversations almost evenly
Share of conversations by language, April 2023 to June 2025. The area of each piece is proportional to the share. All other languages cover more than 50 smaller ones and include Yoruba and Indonesian, likely detection errors. breinz analysis of WildChat-4.8M, 30,000 conversations.
About this paper
Companies choose and trust AI on three kinds of evidence. The first is AI rankings. The second is how safe models turn out to be. The third is studies of how AI is used. The rankings use questions written in English or translated from it.4 5 The usage studies publish countries and leave out the language. Evidence of this kind is skewed toward English, and every decision built on it carries the same skew.
This paper measures the language of AI use directly. It counts the language of ChatGPT conversations in WildChat. This public record tags the language of each conversation and runs to mid-2025.1 It sets that against the languages of the web that AI is trained on.2 Outside studies add the training mix of one model3 and the language of the rankings.4 5 They also show how answers change with the language.6
The record covers one vendor, so its figures show direction, and the limits section sets out what they cannot show. The methodology, limits and full data are at the end.
What the data shows
- About half of the conversations in the record are in English.
English rose to two thirds in early 2024 and fell back to its spring 2023 level by mid-2025. - Most of the web is in other languages.
English makes up 43% of web pages, and has stayed between 40% and 48% since 2018.2 - AI is ranked mostly in English.
The rankings that compare AI models use questions written in English or translated from it. One widely used open model learned from a mix with 8% in other languages. - The same AI answers differently depending only on the language.
In 18 of the 19 other languages studied, Claude agrees and reassures more than in English. The gap is largest in Hindi and Arabic. - Each country’s conversations mix several languages.
In France, Germany and Spain, English takes more conversations than the local language. - Local work passes through English first.
Where English leads, meaning tied to local law, customers and culture passes through a foreign language first.
English has no lasting majority in AI conversations
Its share swung up and back within two years, and Russian, Chinese and French lead the rest.
WildChat records ChatGPT conversations held through a free research chatbot.1 It shows one vendor and one group of people, so its figures describe that group. Within it, English never settled into a lasting majority. Conversations in other languages kept about half of the total throughout, spread over 74 detected languages. The part of AI use that English evidence leaves out stays as large as the part it covers.
English swung up to two thirds of conversations and back to half within two years
Share of conversations in English by half-year. breinz analysis of WildChat-4.8M, 30,000 conversations.
A single snapshot of this record could show English as a clear majority or as half. It depends on when the snapshot was taken. Across the whole period the two halves stayed close, which is why the paper treats half as the honest figure.
Most of the web AI learns from is in other languages
English makes up 43% of web pages in Common Crawl. This public collection of web pages is widely used to train AI. Its share has stayed between 40% and 48% since 2018.
Common Crawl publishes the primary language of the pages in each of its crawls.2 English led every crawl since 2018. In the October 2026 crawl, Russian came next with 6%. German, Japanese, French, Spanish and Chinese followed, at about 4% to 6% each.
English stays below half of the web pages AI learns from
Share of web pages whose primary language is English, yearly average of Common Crawl crawls, August 2018 to October 2026. breinz analysis of Common Crawl language statistics.
So English is the largest single language on the web. Yet more than half of the pages are in other languages. The web’s mix of languages is close to that of AI conversations. The mix behind AI rankings is not.
AI is ranked mostly in English
The rankings that compare AI models use questions written in English or translated from it. One widely used open model learned from a mix with only 8% in other languages.
Companies compare models by their place in AI rankings. One widely quoted ranking checks whether a model gets facts right, and its questions are all in English.4 Another compares models across languages, and its questions were translated from English into 42 languages.5 A translated question keeps the subject matter of its English original. It shows how a model handles that content in another language. It says little about that language’s own concepts.
| Evidence | What it is | Language |
|---|---|---|
| SimpleQA4 | Short factual questions used to check whether a model gets facts right | English only |
| Global-MMLU5 | Questions used to rank models across languages | Translated from English into 42 languages |
| Llama 33 | The training mix of one widely used open model | About 8% in other languages |
Training follows the same pattern. The makers of one widely used open model reported its training mix. About 8% of it was in languages other than English.3 So the figures companies rely on describe AI in English. About half of its use happens in other languages.
The same AI answers differently depending only on the language
Measured against English, Claude reassures and jokes more in Hindi and Arabic, and gets more frank in Dutch.
Anthropic studied how Claude answers requests that call for judgement, such as advice or an opinion, in 20 languages.6 The two charts compare its answers in the other 19 with its answers in English.
In almost every language, the AI agrees and reassures more than in English
Agrees with the person’s ideas and work,reassures them without being asked, jokes and plays more, and keeps to polite language.
Corrects details nobody asked about,challenges wrong assumptions, asks for evidence, and gives frank criticism of the person’s work.
Each bar shows how far Claude’s answers in one language differ from its answers in English. English is the line in the middle, and the longer the bar, the bigger the difference. All 20 languages in the study. Calculated by breinz from Anthropic, Claude’s values across models and languages, 2026, based on 309,815 conversations from May 2026.
In Hindi, Claude jokes more, reassures people who did not ask for it and keeps to polite language. In Arabic, it more often agrees with the person’s ideas and matches its tone to their mood. That is the left side of the first chart. In English, Claude more often challenges a wrong assumption with evidence. It also corrects details nobody asked about. That is the right side of the first chart. In Russian, its answers lean the same way: they get straight to the point and ask for evidence.
In most languages, the AI sticks to the task more than in English
Admits its own mistakes,says when it is unsure, and is upfront about what it can’t do.
Stays within what was askedand aims for a result, giving a polished, confident answer.
Each bar shows how far Claude’s answers in one language differ from its answers in English. English is the line in the middle, and the longer the bar, the bigger the difference. All 20 languages in the study. Calculated by breinz from Anthropic, Claude’s values across models and languages, 2026, based on 309,815 conversations from May 2026.
In Dutch, Claude more often admits its own mistakes and gets straight to the point. Its answers are more frank. In Indonesian, it more often encourages bigger goals and leaves the choice to the person. Its answers lean toward getting the task done.
Safety shifts with the language too.7
The model and the kind of request stay the same, and only the language changes. So a quality checked in English cannot be taken for granted in another language. If the AI corrects a wrong assumption in English, it may reassure the same person in Hindi instead.
Country figures hide the language
Usage studies count AI use by country and record no language. In the record of conversations, every country mixes several languages.
Eurostat and the data behind the AI True Cost Index count AI use by country and record no language.8
In the record of conversations, France, Germany and Spain each mix English, the local language, Russian and Chinese. Conversations from US, UK and Irish addresses mix languages too.1 Part of this mix may come from people writing through addresses abroad. Either way, a country figure shows where a conversation is counted, and its language stays out of view. Studies that report use by country describe places, and the languages AI works in go unmeasured.
Each country’s conversations mix several languages
Share of conversations by language, April 2023 to June 2025. The country comes from the internet address of each conversation. breinz analysis of WildChat-4.8M.
Even where English is the local language, it does not cover all conversations. In the UK and Ireland and in the US, a quarter to two fifths of conversations are in other languages.
Meaning tied to a place is the part at risk
When conversations move into English, the meaning a local language carries has to cross into a foreign language first.
In France, Germany and Spain, English leads the conversations in the record. As far as this record shows, much local work passes through English before AI answers it. The Labs paper Two years to first AI agent found the same move in search. Poles ask what AI is in Polish, then search in English once they start using it. Is Polish really the best language for AI? showed what follows. The writer’s thinking moves into a foreign language, and the model replaces local references with global ones. The meaning tied to local law, customers and culture is what this puts at risk.
The data used to describe AI has the same gap. Two years to first AI agent also found that data on how people learn and use AI is largely in English. It misses the questions people ask first, in their own language.
What follows for companies
Every decision built on English evidence carries its skew. That covers which model to choose, how it behaves, how to roll it out and how to judge its work.
Model choice comes first, because every other decision rests on the model chosen. AI rankings show how a model answers in English. A company working in Polish, German or Spanish learns how its model performs only one way. It has to check the model in those languages, on its own tasks.
Behaviour comes second. The same model reassures in one language and corrects in another, and its safety varies the same way. What a company sees in English is one version of the AI its staff and customers get.
Planning comes third. Training and rollout planned by country rest on figures that record no language. They miss the mix of languages people actually work in.
The quality of the work comes last, because this loss is the hardest to notice. Work moved into English can read fluently and still miss what the local language carried.
The AI True Cost Index measured the time AI saves across all languages together. How that saving differs from one language to another is still open.
Key takeaways
- Half of AI use happens outside English.
About half of the conversations in the record are in other languages. Conclusions drawn from English describe half of real use at most. - English is not settling into a majority.
Its share rose to two thirds and fell back to half within two years. - Check a model in the language it will work in.
AI rankings use questions written in English or translated from it. A model’s performance in Polish or German shows only when it is checked in that language. Check it on the company’s own tasks. - The same AI behaves differently by language.
In 18 of 19 languages it agrees and reassures more than in English. Two people asking for feedback on the same work in different languages may get different verdicts. - A country figure does not show the language.
Use counted by country hides the mix of languages people work in. Training and rollout need planning by language as well. - Work moved into English loses local meaning first.
References to local law, customers and culture are the first to go, and this loss is the hardest to notice.
Methodology
The findings rest on a series of decisions breinz took while building the analysis. Each decision chose between options for a stated reason.
The question. The paper compares the language AI is used in with the language it is trained in and measured in.
The record. WildChat is a large public record of AI conversations. It tags each conversation with its language and country. So the language of use can be counted directly, not inferred from country.1 It covers one vendor and one chatbot. The paper reads its figures as direction, not as exact shares of all AI use.
The sample. WildChat-4.8M holds 3,199,860 ChatGPT conversations in time order. The analysis read 100 consecutive conversations at every 10,000th position. Of 320 positions, 300 were read, giving 30,000 conversations from April 2023 to June 2025. Conversations next to each other cluster in time. So the sample behaves more like 300 points across the period than 30,000 independent ones.
English as the baseline. Every comparison is made against English, because AI is usually ranked in English. A difference from English shows what a company would miss by relying on English results.
The countries. France, Germany and Spain stand for large European languages. The UK and Ireland form one group. They share English with the US but differ from it in location and culture. The US is read on its own.
Detection errors. Languages come from the record’s automatic detection. Yoruba and Indonesian appear at about the same share in every country, which points to errors. So they stay inside all other languages and are not read as findings.
The web. Shares are the percentage of pages by primary language in each Common Crawl crawl from August 2018 to October 2026.2 Yearly figures are the average of that year’s crawls.
Language and answers. Anthropic published, for each of its 20 languages, how far Claude’s answers lean toward one of two opposite traits. breinz read these positions from the study’s language cards. For each language, it took its position minus the position of English. That gives the distance from English.6
Outside studies. They cover what conversation data cannot show. These are the training mix of one model and the language of two AI rankings. They also show how safety changes with the language.3 4 5 7
Limits
The record covers one vendor, reached through a free research chatbot, so its figures describe that group of people. It ends in June 2025.
Language detection errs on short or mixed texts. Yoruba and Indonesian appear at about the same share in every country, which points to errors. Placing people by internet address puts anyone using a server abroad in the wrong country. The training mix comes from one model released in 2024. The evidence on rankings covers two of them. The language comparison of answers covers one model family and 20 languages. Chinese AI services appear in none of the sources.
Within these limits, three findings hold. Half of the conversations in the record are in other languages. The same AI answers differently by language. The training mix and the two rankings examined are almost entirely English.
Data appendix
Languages of conversations. Share of the 30,000 sampled conversations, April 2023 to June 2025.1
| Language | Share |
|---|---|
| English | 52.4% |
| Russian | 11.3% |
| All other languages | 9.7% |
| Chinese | 8.8% |
| French | 4.8% |
| Vietnamese | 3.8% |
| Arabic | 2.0% |
| Portuguese | 1.8% |
| Spanish | 1.7% |
| German | 1.1% |
| Persian | 0.7% |
| Korean | 0.6% |
| Italian | 0.5% |
| Turkish | 0.4% |
| Japanese | 0.2% |
| Polish | 0.2% |
English by half-year.1
| Period | English |
|---|---|
| H1 2023 | 47.0% |
| H2 2023 | 57.1% |
| H1 2024 | 65.0% |
| H2 2024 | 51.9% |
| H1 2025 | 48.1% |
Languages by country. The country comes from the internet address of each conversation.1
| Country | English | Local language | Other languages | Conversations |
|---|---|---|---|---|
| France | 47.1% | 24.0% | 28.9% | 516 |
| Germany | 46.3% | 20.9% | 32.8% | 1,182 |
| Spain | 35.9% | 33.7% | 30.4% | 270 |
| UK and Ireland | 72.2% | 27.8% | 1,052 | |
| US | 61.1% | 38.9% | 9,089 |
English share of web pages. Yearly average of Common Crawl crawls.2
| Year | English | Crawls |
|---|---|---|
| 2018 | 43.9% | 5 |
| 2019 | 43.9% | 12 |
| 2020 | 43.2% | 9 |
| 2021 | 45.0% | 9 |
| 2022 | 46.6% | 6 |
| 2023 | 45.7% | 5 |
| 2024 | 44.3% | 10 |
| 2025 | 43.8% | 12 |
| 2026 | 41.4% | 10 |
Distance from English. Claude’s position in each language minus its position in English, in standard deviations. On the first scale, a minus means it agrees and reassures more than in English. A plus means it corrects more. On the second, a minus means it admits mistakes more. A plus means it stays more within the task.6
| Language | Agrees (−) or corrects (+) | Admits mistakes (−) or stays on task (+) |
|---|---|---|
| Hindi | −0.62 | +0.06 |
| Arabic | −0.41 | +0.08 |
| Thai | −0.24 | +0.07 |
| Japanese | −0.20 | −0.04 |
| Korean | −0.17 | 0.00 |
| Indonesian | −0.13 | +0.18 |
| Vietnamese | −0.13 | +0.10 |
| Spanish | −0.13 | +0.07 |
| French | −0.13 | +0.04 |
| Turkish | −0.10 | +0.04 |
| Romanian | −0.10 | +0.06 |
| Portuguese | −0.10 | +0.07 |
| Chinese | −0.08 | +0.04 |
| Dutch | −0.07 | −0.08 |
| Ukrainian | −0.07 | +0.08 |
| Italian | −0.05 | 0.00 |
| German | −0.04 | −0.02 |
| Polish | −0.02 | +0.04 |
| Russian | +0.02 | +0.04 |
Notes and sources
- WildChat-4.8M, published by the Allen Institute for AI, holds ChatGPT conversations collected through a free research chatbot. Each conversation is tagged with a detected language and a country taken from its internet address. breinz drew the sample described in the methodology. Sample sizes: France 516, Germany 1,182, Spain 270, UK and Ireland 1,052, US 9,089. Allen AI, WildChat-4.8M, 2025.
- Common Crawl publishes, for each of its crawls, the share of pages by primary language. English made up 42.6% of pages in October 2026, against 43.9% in August 2018. Across all 78 crawls in between, it ranged from 40.5% to 47.8%. Common Crawl, Statistics: languages, 2026.
- Meta’s report on its Llama 3 models describes their training mix. About 8% of it was in languages other than English. Meta, The Llama 3 Herd of Models, 2024.
- SimpleQA, a widely quoted ranking of short factual questions, is written entirely in English. Zahra et al., Whose Facts Count?, 2026.
- Global-MMLU translates the questions of the English MMLU ranking into 42 languages. Singh et al., Global MMLU, 2024.
- Anthropic studied 309,815 Claude.ai conversations in 20 languages from May 2026. In them, people asked for advice, an opinion or another judgement. Its language cards give, for each language, Claude’s position on four pairs of opposite traits and its distinctive behaviours. breinz computed each language’s distance from English. Anthropic, Claude’s values across models and languages, 2026.
- A study of how the safety of one open model changes across 18 languages. Ifebi et al., Minionese: Multilingual LLM Safety, 2026.
- Eurostat reports AI use in the EU by country and age group, with no breakdown by language. Eurostat, 64% of 16-24-year-olds used AI in 2025, 2026.