Moderating a language you don't speak

Moderators are the people who review posts and decide whether they break a platform's rules. On any global platform, they review posts in languages they don't speak. There are thousands of languages and no team covers them all. So the real work runs on machine translation, a few scarce language experts, and a lot of educated guessing.

People in the field don't say this out loud much, which is why it's worth writing down how to do it honestly. When it's done carelessly, the people who pay are usually in countries that already get the least attention.

When a translation stops being proof

Machine translation is better at some things than others. What usually comes through:

• Direct threats and obvious slurs
• Spam and scams
• Names of people, places and groups
• Roughly what a post is about
• Long posts in widely spoken languages

What usually gets lost:

• Code words, hidden hate symbols, and numbers that stand in for phrases
• Sarcasm, irony, and jokes that mock the harmful thing
• Slurs a group has taken back and uses about itself
• Slang, local dialects, and spelling tricks meant to sneak past filters
• Short posts, song lyrics, and anything that depends on tone

See the pattern? Translation keeps the words and drops the meaning behind them, and most hard calls are about meaning. A sentence that reads like praise can be well-known sarcasm. A string of harmless words can be a known call to violence. Treat the translation as a clue, and keep looking.

Handling a case honestly

• Look at the behavior before the words: who posted, to whom, replying to what, and how often. Behavior doesn't need translating.
• Treat the translation as one witness, not the whole story. If the decision depends on a small shade of meaning, that shade is exactly what you don't have.
• Check the team's list of code words first. If your team keeps one for this country or language, it settles cases faster than translation ever will. If there isn't one, start it today.
• When the call is close and the harm is serious, pass it to someone who reads the language. "I can't judge this reliably" is what experienced people say. It isn't failure.
• Note that the decision was made through a translation. It's one more honest line in the record, and whoever reviews an appeal deserves to know.

In a language you don't speak, confidence is the mistake. The skill is knowing which half of the picture you actually have.

Making the most of language experts

People who read the language are rare, so save them for cases where the language is what decides it. And learn from them every time. Each case they settle should leave something behind: a phrase added to the code word list, an example added to training, a rule reworded so the next reviewer doesn't need to ask.

An expert whose knowledge lives only in their head is a problem the day they go on vacation, and they will go on vacation. One whose knowledge ends up in the team's shared tools makes everyone better. And when they tell you coverage for a language is genuinely unsafe, put it in writing in the team's formal review of risks, with their name on it if they're willing.

If you plan the coverage

Companies that sell automatic detection tools mostly measure their accuracy on the big languages. That's why the guide on testing tools says to try your smallest language first. The same trap hides in your own numbers. If you never measured how much bad content you catch in each language, you don't actually know it for the languages that need it most.

An honest coverage map is uncomfortable to draw. It shows which languages get a native speaker, which go through translation, and which barely get looked at. It's also the most useful thing a global team can hand to its rule writers and its bosses.