Safety that only works in English isn't safety

For 3 months, our users in a non-English market were invisible to our safety system. They just did not speak English.

I was testing crisis detection and tried a non-English phrase that clearly expressed suicidal ideation. The response was a normal, cheerful coaching reply. No escalation. Nothing.

Tracing the blind spot

1. The toxicity classifier we used was English-first. Non-English text with no English keywords scored below the moderation threshold.
2. The local crisis detector was a hard-coded English keyword list, with no non-English patterns at all.
3. The result: non-English suicidal ideation passed both safety layers undetected.

And this product was specifically built for a non-English-speaking market. The most vulnerable users were the least protected, since day one.

The fix

I added a crisis intent to the LLM classifier that already handled routing. The LLM understands intent regardless of language. I made every reply path escalate on crisis, and added a guard that can never downgrade a confirmed crisis once it is raised.

Verified: crisis caught across languages and even misspellings, while benign idioms like “dying for a cigarette” correctly did not trigger the wall.

The lesson, hard-earned: a safety system that only works in one language is not a safety system. It is a safety system with an exception, and the exception is a person. Test your safety in every language you support, before you ship.