Your child’s AI might be safer in English than in their own language

Una passed every child safety test we gave her in English. A perfect score. Then we ran the exact same tests in French, translated word for word. She scored 83 percent. Same questions. Same guardrails. One language across, and she let through nearly one in six things she should have caught.

If your child does not speak English as their first language, that gap should bother you as much as it bothered us. It is the reason we now test Una against our child safety protocols in every language she speaks, before any child does.

Unomundi is a cultural-discovery app where children aged 6 to 12 explore countries and cultures through stories, games, and conversations with Una, a warm AI guide. Children talk to her in their own language. So her safety has to hold in their own language too.

Why an AI can be safe in one language and not another

The short version is this. The safety in most AI is taught mostly in English, and it does not fully carry across to other languages. In plainer terms, five things stack up.

1. It grew up reading English. Almost everything the AI learned from was written in English, so it is simply sharper in English than in any other language.

2. It was taught to be safe mostly in English. The rules that make it refuse something harmful were trained on English examples, so they are strongest in English and thinner everywhere else.

3. It thinks in something English-shaped. Underneath, the AI runs everything through an English-leaning core, so its caution fires most reliably when the words arrive in English.

4. Other languages come through blurry. The AI reads text in small chunks. English breaks into clean chunks, while many other languages get chopped into messier ones, which gives its safety checks a fuzzier signal to work with.

5. Even what counts as harmful was decided in English. The people who taught it what to block mostly worked in English, so risky phrasing, slang or nuance in another language can slip straight past.

   
   

When you ask the same model something harmful in another language, those English lessons do not fully carry over. The model understands the words. It just does not apply the same caution.

This is well documented. Researchers at Brown University showed in 2023 that translating a harmful request into a less common language could slip past GPT-4's safety filters around 79 percent of the time, when the same request in English was blocked. A 2024 study presented at the ICLR conference found the same pattern across many languages. Safety that looks solid in English gets thinner the further you travel from it.

There is a reason translating the test alone does not fix it. The safety is not just a list of banned words. It is a way of reasoning that the model learned in English. Swap the language of the question and you do not automatically swap in the same caution behind it. The words change. The judgement underneath does not always follow.

French is not even a hard case. It is one of the most resourced languages in the world, with huge amounts of training data behind it. If Una slips in French, the risk in languages with less of the internet written in them is far higher. The language-data firm Welo Data reported in 2025 that unsafe responses can rise from under one in ten in English to four or five in ten in some under-represented languages. The same model. A different language. A very different level of protection.

What most companies actually test

Here is the uncomfortable part. A lot of AI products are tested for safety in English and then shipped to the whole world. The assumption is that if it is safe in English, it is safe everywhere. Our own results show how wrong that assumption can be.

I understand why it happens. Testing in one language is faster and cheaper. The teams building these systems often work in English. And the gap is invisible unless you go looking for it, because the product feels fine. It answers. It is polite. It quietly lets more through.

For a general chatbot used by adults, that is a problem. For a product a seven-year-old talks to in Portuguese, or Arabic, or German, it is not a problem we are willing to hand to a parent.

Picture the same worried child typing the same hard sentence in two languages. In English, Una catches it, slows down, and responds with care. In a language we never checked, she might miss it entirely. That child did nothing different. They were simply born somewhere else. I am not prepared to build a product where that is the deciding factor in whether our guardrails hold.

What we changed

So we changed how we test. Every guardrail Una has is now checked in every language we release her in, against the same child safety and wellbeing protocols, with the same scoring. A language does not launch until it clears the bar we hold English to. We do not just translate the test and call it done. We look at how she reasons in each language, not only whether the final words came out clean.

That is also why, for now, Una speaks English only. It was a hard call. We built her for the world, and we are working through the other languages right now. Each one stays in testing while we adapt it, and it goes live the moment Una passes the same tests she passes in English. Not before.

It is slower. It is more expensive. It is also the only honest way to make the promise we make to parents, which is that Una is safe for your child. Not safe for your child if they happen to speak English. Safe, full stop.

A child should not be less protected because of the language they were born into. Safety that only works in English is not safety. It is a setting most of the world never gets.

Sign up to test our beta now

FAQs

Is an AI as safe for my child in our language as it is in English?

Not automatically. Most AI safety is trained and tested in English first, and it does not fully transfer to other languages. We test Una against the same child safety protocols in every language she speaks, and a language does not launch until it meets the same bar as English.

Why does an AI behave differently in different languages?

Because the safety rules are mostly learned from English data and English examples. In other languages, the model understands the words but applies less caution. Independent research has shown harmful requests can bypass safety filters far more often when translated out of English.

How does Unomundi test Una's safety across languages?

We run our full child safety and wellbeing tests in every language we release, with the same scoring, and we review how Una reasons in each language, not only the final wording. People stay in the loop alongside the automated tests.