Short answer. A model performs reliably only in the languages it genuinely learned, and training mostly on English produces four measurable penalties everywhere else: lower accuracy, higher running costs from token inflation, weaker safety guardrails, and cultural errors no benchmark catches. The safety finding is the one that changed how teams think about this — translating unsafe prompts into low-resource languages such as Zulu produced harmful responses from GPT-4 roughly 80% of the time in Brown University research. A guardrail that fails in any language is a guardrail anyone can route around with a free translation tool.
Multilingual data used to be a localisation task, handled after launch. It has become a core training decision, for reasons that have less to do with fairness than with cost, security and the exhaustion of English text.
This piece sets out what multilingual data actually changes inside a model, why running a product in another language costs more, why alignment does not travel, and what the evidence says about whether multilingual training helps English performance too.
Why does a model that scores brilliantly in English still fail abroad?
Because benchmark scores measure the language the model was trained in, not the language your customers use. Capability does not automatically cross a language boundary.
Consider a support assistant that handles refund disputes flawlessly in testing, then ships to customers writing in Indonesian. It answers a question about an instalment plan as though it were a loan default. It responds to polite formal phrasing with a bluntness that reads as rude. Occasionally it replies in English for no reason. Nothing in the launch checklist predicted any of it, because the checklist was written in English and passed in English.
Research on multilingual reasoning finds the same pattern repeatedly: models handle a task well when it is posed in English and degrade when the identical task is expressed in a lower-resource language. The knowledge is often present. What is missing is the ability to reach it reliably through a different language, because the pathways between concepts and words were built almost entirely on English examples.
What does multilingual data actually change inside a model?
Three things, at three different depths.
Tokenizer efficiency. Text is broken into tokens by a tokenizer, which learns to compress whatever it saw most during training. Feed one a diet of English and it learns compact representations of English and treats everything else as unfamiliar fragments. Multilingual data changes what the tokenizer considers ordinary — and that decision is fixed for the life of the model.
The conversion layer. Studies of how models handle facts across languages suggest knowledge is stored in a largely language-agnostic internal space, and that errors often appear at the final step, when that internal representation is converted back into a specific language. Multilingual training data strengthens exactly that conversion. It is the difference between a model that knows something and a model that can say it correctly in Vietnamese.
Generalisation beyond the training list. Because related languages share structure, exposure to many languages helps a model make better guesses in ones it has seen only briefly. Breadth matters even for languages that never appear in the plan.
Why is running an AI product in another language more expensive?
Because models charge and think in tokens, and the same sentence becomes far more tokens in languages the tokenizer was not built around. Practitioners call it the token tax, and it has nothing to do with quality.
Analysis comparing several commercial and open tokenizers found Arabic text needing anywhere from around 68% more tokens in one model to over 340% more in another, for the same paragraph. Nothing about the idea got bigger. Only the accounting did.
| Consequence | Why it happens |
|---|---|
| Cost rises | APIs bill per token in both directions |
| Context shrinks | A document that fits in English may overflow the same context window in Telugu or Japanese |
| Latency grows | There is simply more to process |
The uncomfortable part is who pays it. Research examining tokenization across many languages has found the penalty tends to fall hardest on languages spoken in lower-income regions, which means the markets least able to absorb the cost are charged the most for the same intelligence. Better multilingual training data is one of the few structural fixes, because a tokenizer trained on genuinely diverse text encodes that text more efficiently from the start.
Why do safety guardrails weaken in other languages?
Because safety alignment is trained, not inherited. A guardrail exists only in the languages it was taught in, so a model can be strict in English and permissive in Zulu at the same time.
Researchers at Brown University showed that taking unsafe English prompts, translating them into low-resource languages such as Zulu with a free translation tool, and sending them to GPT-4 produced harmful responses roughly 80% of the time — against a model that refused the same requests in English. No technical skill was required.
Later work tracks how the picture has evolved. A 2026 study testing African languages including Kiswahili, isiXhosa and isiZulu found that straightforward translation attacks no longer work as easily as they once did, which is real progress, but that conversations spread over several turns still succeed at high rates. Separate evaluation across 79 languages found unsafe response rates climbing by as much as 25 percentage points as prompts moved from English into low-resource languages.
The strategic point is the one that gets missed. This is not only a problem for speakers of those languages. A guardrail that fails in any language is a guardrail anyone can route around with a translation tool and thirty seconds. Multilingual safety data is a security control, not a diversity gesture.
Can translation just do the job instead?
Translation moves words. It does not move meaning, context, register or intent, and it inherits every assumption baked into the original English. It is genuinely useful as a bridge, and the mistake is treating it as a substitute.
Three things break. Register and politeness get flattened — many languages encode formality, seniority and social distance in grammar, so a translated reply can be technically correct and socially wrong, which in a customer-facing product reads as rudeness. Local reference points disappear — prices, legal terms, document names, payment methods, holidays and units all carry local meaning a translated English answer quietly gets wrong. And the original framing survives, so a model built on English assumptions about how banking, healthcare or family structure works keeps those assumptions and simply expresses them in a new language.
There is also a compounding effect. When translated text is used to train or evaluate, translation errors become training signal, and the model learns a slightly warped version of the language that then looks correct to anyone checking with the same translation tool. Native speakers break that loop. Nothing else does.
Does multilingual data make a model better in English too?
Increasingly, yes, for two independent reasons.
Cross-lingual generalisation. In a 2025 study, a model trained with reinforcement learning on Chinese reasoning data improved not only on Chinese but substantially on German, Spanish and Bengali evaluations — well beyond what supervised fine-tuning on the same data achieved. Other work finds that reasoning skill, as distinct from factual recall, transfers between languages remarkably well. Training in more than one language appears to push a model toward strategies that work generally rather than tricks that fit the training language.
Supply. Epoch AI estimates the effective stock of quality, human-written public text at roughly 300 trillion tokens, and projects that frontier training runs will fully consume it somewhere between 2026 and 2032. English is the part being exhausted fastest, because it was mined first. Meanwhile most of the world's linguistic output has never been digitised at all — it sits in speech, in local platforms, in messaging, in undocumented dialects.
Put those together and multilingual data stops looking like a cost centre. It is simultaneously the largest untapped reserve of human-generated training data and a route to better general capability.
What should a team building AI actually do?
- Pick your languages deliberately, and early. Language coverage affects tokenizer design and data mix, both painful to change later. Retrofitting a language after launch costs far more than including it in the plan.
- Evaluate per language, never in aggregate. A single averaged score hides exactly the failures that matter. Accuracy, refusal behaviour, tone and safety should each be measured language by language, against benchmarks written by speakers rather than translated into their language.
- Red-team in every language you ship in. Given how guardrails degrade, a safety evaluation conducted only in English tells you almost nothing about your actual exposure.
- Use translation as a bridge, not a foundation. It is a reasonable way to bootstrap coverage, and not a substitute for in-language data when accuracy, safety or brand tone are on the line.
- Keep speakers of the language in the loop. Automated checks confirm format and completeness. They cannot tell you the model was subtly condescending, used the wrong register for an elder, or invented a legal term that does not exist in that country.
How Lifewood approaches this
Lifewood's multilingual work sits at the stages where authorship cannot be substituted: collection, annotation, evaluation and red-teaming carried out by native speakers rather than translated in. The practical consequence is that a model's behaviour in Bengali is judged by someone who speaks Bengali, rather than by a score averaged across a dozen languages.
That is a delivery-network question rather than a tooling one. 50+ languages including underrepresented dialects, 40+ delivery centres across 30+ countries, and 56,788 registered contributors are what make per-language red-teaming and per-language evaluation practical at the point in a programme where they matter — before launch, not after a safety incident.
Quality is verified under a human-in-the-loop model against a customer-approved gold set, at a 95%+ accuracy SLA, with per-language reporting rather than an aggregate figure. See AI data services, multilingual data collection and high-resource vs low-resource languages.
Sources and further reading
- Yong et al., Low-Resource Languages Jailbreak GPT-4, Brown University.
- Multilingual jailbreaking of LLMs using low-resource languages (2026), arXiv.
- Welo Data, Global Security Blind Spots: LLM Safety Failures in Low-Resource Languages (2026).
- Predli, Token Tariffs and the Case for Custom Tokenizers.
- Petrov et al. and Ahia et al. on tokenization inequality, summarised in Measuring the Tokenization Premium, arXiv.
- Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs (2025), arXiv.
- Epoch AI, Will we run out of data? Limits of LLM scaling based on human-generated data.