Why AI Should Speak Your Language
Published
Most AI models barely speak many of the world's languages, and they fall short on what matters most locally, like health, learning, and farming. That gap leaves billions of people underserved. This essay explains why language coverage in AI is a fairness problem, and how speakers of underserved languages can help close it, a few taps at a time.
Ask a popular AI model a question in English and it will usually answer well. Ask the same question in one of hundreds of other widely spoken languages and the answer often gets thinner, stranger, or simply wrong. Most models barely speak many of the world’s languages, and where they do limp along, they fall short on what matters most locally, like health, learning, and farming. That is not a small inconvenience. It leaves billions of people with worse AI than their neighbors, for no reason other than the language they grew up speaking.
Why this is a fairness problem, not a technical footnote
AI is starting to sit between people and the things they need: a clearer explanation of a diagnosis, help with a child’s homework, advice on when to plant. When the tool works in one language and stumbles in another, it quietly hands an advantage to some people and withholds it from others. The same question, asked in two languages, gets two different qualities of help.
The cause is not that some languages are harder for machines in principle. It is that the human feedback used to make models good was gathered mostly in a few well-resourced languages. Models learn to be useful from people’s judgments about which answers are better. Where those judgments are abundant, the model improves. Where they are missing, it does not. The gap in coverage is really a gap in who got asked.
Why machines cannot just fill the gap themselves
A tempting shortcut is to take what a model already knows in one language and machine-translate the rest. This is exactly the failure mode Tendril exists to avoid. Automatic translation flattens tone, drops the diacritics and grammar that carry meaning, and misses the local context that makes an answer actually useful. A health or farming answer that is technically translated but locally wrong can be worse than no answer at all.
What closes the gap is the real thing: judgments from people who actually speak the language and live where the questions come from. A fluent speaker can tell, in a second, which of two answers sounds right, respects the politeness a situation calls for, and would genuinely help. No amount of compute substitutes for that.
How speakers help close it
This is where contribution becomes concrete. On your phone, you see two AI answers to the same question and you tap the one that is better. Each tap is a human-preference judgment, and judgments from fluent speakers of an underserved language are the most valuable signal there is for making AI better at that language. It takes seconds and needs no technical skill. This surface is Tap to Train, and you can find your language and start on the languages page.
A word on honesty. Tendril does not yet have data for these languages, and we do not pretend otherwise. The work described here is the path to creating it, contribution by contribution, with fluent native speakers in the lead. The judgments people make become open human-feedback data that can be used to build AI that understands the language, rather than a private asset locked away.
For each contribution you earn TEND Points, a way to recognize what you put in. TEND Points are not money. The point is bigger than that: AI that answers your real questions, in your own language, as well as it answers anyone else’s. If that future is worth a few taps to you, see Open AI as a Public Good for why this data should belong to everyone, then start with the language you speak.