AI-Generated Audio Summary

Verónica, the AI tutor built into our courses and applications, now responds in twelve languages. Arabic, Telugu, Tamil, and Korean join English, Spanish, Mandarin, Vietnamese, Tagalog, Hindi, Urdu, and Brazilian Portuguese.

That sentence took about two minutes to write and considerably longer to earn. The work was not what most people would expect, and the parts that turned out to be genuinely hard are worth explaining, because they say something about language that teachers already know intuitively.

Why these four languages

We did not choose them by looking at global speaker counts. We chose them by listening to the people who told us which families walk into their buildings.

District staff and teacher preparation faculty have been consistent about this. In North Texas in particular, the communities around Plano, Frisco, and Irving brought Telugu and Tamil into nearly every conversation about home language surveys. Arabic came up repeatedly across the Dallas and Houston areas. Korean came from the Carrollton and Plano corridor.

These were not requests for a feature. They were descriptions of who is already sitting in the classroom. The most useful product feedback we get is rarely phrased as a request. It usually sounds like someone describing their week.

Adding a language was never the difficult part

We wrote earlier this year about why modern AI was built multilingual from the start, and that piece is the foundation for this one. These models were not trained in English and then translated outward. They were built on many languages at once, which is why they can evaluate Spanish writing with real fluency rather than approximating it through English.

That has a practical consequence. Verónica did not need to be taught Arabic. The capability was already there, the same way it was already there for Vietnamese and Urdu. What she needed was permission, a way to introduce herself in that language, and above all a reliable way to know that Arabic was the language in front of her.

That last one is where the work was.

The first hard part: letters betray you

Here is the problem in one sentence. Before Verónica can answer in your language, something has to decide what your language is, and the evidence available is just the characters you typed.

That turns out to be far less reliable than it sounds.

Two languages can share an entire alphabet. Arabic and Urdu are written in the same script. A system that checks “are these Arabic letters?” will happily route an Arabic speaker into Urdu and feel confident doing it. The two languages are not interchangeable, and someone who writes in Arabic and gets greeted in Urdu learns immediately that the tool does not actually see them.

Accented letters wander across language families. Portuguese and Vietnamese share several accented characters. A perfectly ordinary Portuguese sentence can trip a Vietnamese check, and the greeting comes back in the wrong language while the rest of the answer is correct, which is somehow more unsettling than being wrong outright.

A single short word can belong to three unrelated languages. The word “para” is common in Spanish, common in Portuguese, and a real word in Tagalog. Any system that treats a shared word as proof of a language will misclassify a large number of Spanish speakers, which is the last group we can afford to get wrong.

And underneath the second one, something more basic was wrong. Why would an ordinary Portuguese sentence ever reach a Vietnamese test in the first place? Because the Portuguese checks could not fire. To check whether a sentence contains a particular word, a program has to know where words begin and end, and the standard built-in rule for that counts only the unaccented letters A through Z. We were looking for the Portuguese você and the Spanish está. Both end in a character the rule does not count as part of a word, so neither check could ever match, not once, not for anyone. Portuguese was falling straight past its own test and into somebody else’s. Spanish was usually rescued by a separate check for the ñ and the opening question mark, which is luck, not design.

You have seen a version of this outside of any chatbot. A school system that prints Hernández as Hernandez. A roster that will not accept the ñ in a child’s first name. A certificate that turns an accent into a question mark. It is the same failure, just visible instead of hidden.

The reason it keeps happening is worth saying plainly. Unaccented English letters are the default in most of the tools software is built from, and every other writing system is a special case somebody has to remember to handle. Nobody decides to exclude Spanish. Exclusion is what happens when the extra step gets skipped, and the software never announces that it skipped it.

We wrote this bug ourselves. We did not invent the assumption underneath it.

The second hard part: the page thought it knew you

Characters were only half of it. The other half had nothing to do with letters. It was a question of authority.

A chat window sits on a page, and that page carries assumptions about who is reading it. A Spanish course page announces itself as Spanish. So the moment someone types a sentence, there are two competing claims about what language they speak: what the page assumes about them, and what they just wrote. Our code checked the page and stopped there. On a Spanish page, a message written in Urdu would have been filed as Spanish without anything ever reading it.

That one never reached anybody. A wiring detail meant the page’s language was not actually being handed to the part of the system that decided, so the trap sat there loaded and never sprang. We found it by testing what would happen, which is the cheapest way there is to find anything.

The order we settled on is easy to say out loud: what you wrote, then what you have been writing, then what the page assumes about you. A Vietnamese conversation stays in Vietnamese when the person types “ok” on the fourth turn, instead of reverting to whatever the page believes. Evidence from the person outranks assumptions about the person, and the conversation you are already having outranks the room you happen to be standing in. When even that runs thin, the rule is restraint. Say less rather than guess loudly. The page setting is always available and always convenient, which is exactly why it cannot be allowed to win.

All of this mattered because of one small thing. That EN and ES switch in the corner of the chat window did nothing. Clicking it relabeled the button and told Verónica absolutely nothing. We could not simply connect it either, because wiring it up before settling the order would have let a visitor who clicked “ES” out of curiosity lock themselves out of Arabic. The unglamorous work of deciding who outranks whom is what made a one-line feature safe to ship. It is connected now, and relabeled to say what it actually does: it sets the language of the buttons and menus, and Verónica answers in whatever language you write in.

What this has to do with learning

That is the engineering. Here is why it mattered.

When a student can ask a question in the language they think in, the question gets asked. That is the whole mechanism. The barrier was never comprehension of the content, it was the cost of translating a half-formed thought into a second language before you are allowed to express it. Lower that cost and students ask more, ask earlier, and ask about the thing that is actually confusing them rather than the thing they have vocabulary for.

The same tutor, told to do the opposite

That principle has a limit, and it cuts the other way. The same technology has to behave in opposite ways depending on the goal.

We learned this from a student, not from a test suite. She was preparing for a Spanish proficiency exam, practicing conversation with Verónica, and Verónica kept drifting into English partway through. The reason is ordinary and human. The student would reach for one English word, the way bilingual speakers naturally do, and Verónica read that word as a request. She followed. Then both of them were code-switching.

In most of our apps that behavior is exactly right. If the goal is understanding content, meeting a student in the language they think in is the entire point. In proficiency practice it is the opposite of right. The exam penalizes code-switching. For this particular student, code-switching into English was the recurring note in her own evaluation feedback. The tutor was patiently modeling the exact habit the exam would mark her down for.

When we opened the code, the instruction was sitting there in plain Spanish: if the person gets stuck or speaks English, you may explain briefly in English, and do not insist that they switch back right away. That is a good rule. We wrote it for our English-learning apps, where English is the target and following the learner is the correct move. It had been carried across into a Spanish app, where it runs exactly backwards. Nothing was broken. A sound instruction was standing in the wrong room.

The fix was not to forbid English. A tutor who has only been forbidden things has no plan for a student who is genuinely stuck, and a tutor with no plan improvises, usually straight back into English. What she has now is a ladder: say it again more slowly, say it a simpler way, give a concrete classroom example, and finally hand the student the Spanish phrase to repeat. She never announces the rule and never points out the slip. She stays in Spanish and helps. Students can still write to her in English, and the activity tells them so; she answers in Spanish and helps them say it in Spanish.

We build for both cases deliberately, and they are not the same product decision wearing different clothes.

What we are not claiming

A large language model can produce text in far more languages than twelve. We could put a bigger number on this page tonight.

We are not going to. We list the languages we have tested, introduced properly, and checked against real sentences. The difference between “the model can probably handle it” and “we have verified this works for a teacher candidate studying at the kitchen table at eleven at night” is the entire difference between a claim and a product. Twelve is what we can stand behind today.

If you are learning with her, or deciding for a program

Verónica is already answering in these languages inside our courses and applications. There is no setting to change and nothing to enable. Write to her in your language and she will answer in it.

Most of the people who do are teacher candidates, practicing teachers, and adult learners preparing for an exam, and a good number of them are bilingual themselves. The languages matter twice over, once for the person studying, and again for the families that person will be serving on Monday morning.

If your district, university, or program needs a language we have not covered yet, tell us which one and roughly how many people. That is exactly the input that produced this update, and it is far more useful to us than a feature request.