Our goal is to be the best app for serious learners who learn through comprehensible input. We have a second goal now too. Over time, we want to be the first app that lets you learn any language.
That is a long-term aim, and we are far from it today. The world has 7,170 living languages, according to Ethnologue, a catalog of the world's languages. About 3,000 of them may have no written form at all. The UN also counts more than 300 sign languages, which would need a different kind of app. Today we teach 122 languages.
So here is our plan. We keep adding languages when we can teach them well, and we tell you plainly where a language still has weak spots. Some of what stands in the way depends on the AI models we use, and some depends on us.
| Today | How many |
|---|---|
| Journeys | 174 |
| Reading levels in each journey | 6 |
A journey is one way to learn a language, such as Spanish from Spain or Spanish from Argentina. Poppy, your AI tutor, writes lessons and stories for you at your level in each one. If something in a lesson is hard to follow, you can ask her about it.
- Languages People SpeakMost journeys are languages people speak today, and some come in more than one dialect.
- Old LanguagesYou can read Latin, Ancient Greek, Biblical Hebrew and Sanskrit, which people mostly read today.
- Invented LanguagesYou can also learn Esperanto and Toki Pona, which people made up on purpose.
Before a new journey fully launches, it now goes through about 50 checks. Here are a few of them.
- A DictionaryWhen you tap a word, you see what it means.
- Grammar NotesEach grammar note explains a pattern with example sentences.
- AudioLessons and grammar examples come with audio you can play.
- All Six LevelsEach level needs at least three lessons that you can read and play.
Some of our newer journeys show a short warning in their lessons.
| Warning | Where you'll see it | What it means |
|---|---|---|
| Early-access audio | Some newer journeys, such as Scottish Gaelic, Shona, Lingala, Tok Pisin and Occitan | The AI voice may not sound like the people who speak the language, or it may say some words wrong. |
| Word meanings | Dzongkha | The meaning you see when you tap a word can be wrong, most often at Levels 4 to 6. |
| Word meanings | Papiamento | The meaning you see when you tap a word can be wrong. |
| Words from later Hebrew | Biblical Hebrew | At Levels 5 and 6, a lesson can use a few words from later Hebrew that the Bible does not use. |
| Word meanings and spelling | Cape Verdean Creole | Tapped meanings can be wrong, and a few words may use Portuguese spelling, most often at Levels 4 to 6. |
We use AI voices. Before a new language launches, AI graders listen to its voice. If they hear a problem, we still launch the language with the audio warning, because the lessons can help you even while the voice isn't right yet.
The AI models we use make more mistakes in some languages. In Dzongkha and Cape Verdean Creole, the higher levels use rarer words that our dictionary doesn't have yet. In Papiamento, mistakes show up at every level after Level 1, and most of them give a common word the wrong meaning. For example, a word that means "kitchen" was shown as "to cook".
We chose a warning over waiting because these lessons can still help you. Most Papiamento lessons passed our checks, for example, so you can still read and listen. Waiting would have meant no Papiamento lessons at all, with no date for when the rest would be fixed. With the warning, you know which part to trust less.
AI models learn a language from text, and many languages have very little of it. These counts come from different catalogs.
- 7,170That is how many living languages the world has.
- About 4,150They have a developed writing system.
- About 2,700They have a full grammar book.
- About 2,700They have at least a New Testament.
- About 270They have 1,000 or more entries in English Wiktionary.
- 183They have a Wikipedia with more than 10,000 articles.
Ethnologue counts 3,226 of these languages as endangered. More than 1,000 have nothing more than a short word list written about them.
Most people speak one of a few hundred languages. Over 88% of people grow up speaking one of the top 200. A grandchild who wants to talk with a grandparent needs that grandparent's language, wherever it ranks, and we'd love to help with that one day.
Each new language needs five things, and the first one is the hardest.
- A Way to Check LessonsSomeone or something has to catch the mistakes in each lesson.
- Enough TextLessons need a lot of writing in the language to start from.
- A VoiceEvery lesson needs audio you can listen to.
- Sources We May UseThe texts and voices we build on have to allow use in an app like ours.
- The Community's AgreementSome communities decide who may teach their language.
Poppy writes each lesson with an AI model, and we check lessons with AI models too. Models can be worse at checking text in smaller languages. In one study, GPT-4 agreed with human reviewers far less when it judged summaries in Bengali than in French. When no model can check a language well, the only answer is people who speak it.
English Wiktionary is one of the dictionaries we build on, and the numbers above show how few languages have much in it. The gap shows up in tests too. On a translated knowledge quiz, OpenAI's o3 model scored 91% in Spanish and 78% in Yoruba. The newest models are closing that gap on quizzes like this one, and a quiz is a different job from writing a good lesson.
Most of our voices come from Gemini, Google's AI model, which lists 132 languages and scripts for speech. Meta has research voices for 1,107 languages, but their license doesn't allow use in an app like ours without Meta's permission.
A lot of text in smaller languages can't be used freely. Some storybook libraries let anyone reuse their books, but many others allow only non-commercial use. A Bible that anyone may share is not always one that anyone may change. We also have a rule of our own: we write the example sentences in our grammar notes fresh, and we only use sentences we read elsewhere to check how a language works.
Publishing a language without the community's agreement has caused harm before, as in the dispute over a Hopi dictionary. This choice belongs to the community, and better AI models don't change that. For a language like that, we would want to work with the community before we teach it, and only with its agreement.
- Lots of Old WritingAbout a dozen old languages have enough text for real reading lessons. Classical Chinese and Classical Arabic have more writing than many living languages.
- Not Fully ReadableSome old languages, like Etruscan, can't be fully read by anyone today.
- Rebuilt by ScholarsSome languages, like Proto-Indo-European, were never written down.
Proto-Indo-European is the language that English, Latin, Greek and Sanskrit all grew from. Scholars rebuilt it by comparing the languages that came from it, and they still disagree about some of its sounds. The only texts in it are short modern ones, like the fable August Schleicher wrote in 1868, which scholars rewrite as the theory changes.
Studying Proto-Indo-European would mean reading those short texts word by word, and seeing how its roots became words you already know. We would label any journey like that clearly as a reading journey built on one scholar's version. It could not include conversation, or audio of how people really spoke. We have not decided to build one.
| Depends on the AI models | Depends on us |
|---|---|
| Writing and checking lessons in smaller languages | Keeping the app running until it pays for itself |
| Voices for more languages | Testing each new model on the languages that carry warnings |
In 2024, Google gave its Gemini 1.5 model a grammar book of about 500 pages for Kalamang, a language with fewer than 200 speakers. It also gave the model a word list of 2,000 entries and 400 example sentences. With all that, the model translated English into Kalamang about as well as a person who studied the same book. Translating out of Kalamang, it did clearly worse than that person. Without the book, it could not translate Kalamang at all. A later study found that almost all of the gain came from the example sentences, not the grammar explanations.
In an early research paper from Google Research in 2026, an experimental model learned to translate two rare languages into English one after the other without forgetting the first. That's a step toward models that keep learning new languages. It did not test writing in those languages, and the translation scores stayed low. None of these studies tested writing lessons at six levels.
We get an alert when Gemini changes its list of voice languages, so we can test a new one right away. What nobody knows yet is how many languages a model can write good lessons in. No public test measures it. We also don't know when models will reach the next few hundred languages. We hope to see big steps within the next few years.
Teaching more languages only matters if Bubblegum is still here to teach them. Our numbers are public on our Open page.
| On November 19, 2026 | Number |
|---|---|
| Installs | More than 208,000 |
| Learners each month | About 28,000 |
| Pro subscribers | About 330 |
| Yearly recurring revenue | About $28,000 |
| The money | Amount |
|---|---|
| Spent in the last 12 months | About $281,000 |
| Net revenue in the last 12 months | About $23,600 |
| From investors | $0 |
| From personal savings | $195,000 |
| Loans from friends and family | $149,000 |
We are not profitable yet, so smaller languages wait their turn behind keeping the app running. We expect each smaller language to bring fewer learners than a big one. Each journey also adds work to every later change, because Poppy's instructions are written once for every journey, and every improvement has to be made in each one. Until Bubblegum pays for itself, most of our work goes into making lessons better for the learners we have, and into helping more people find us.
Two rules stay the same, however fast we grow:
- We will not launch a language without all six levels. If the higher levels aren't good enough, we fix them, launch with a clear warning, or wait.
- We will tell you about a weakness we know about. When we launch a journey with a known weakness, its lessons tell you which part to trust less.
We will not say Bubblegum teaches any language until it is true, and we don't expect that for a long time. There will be languages we never teach. Some communities teach their own language best, and then we should point you to them.
We'd love to know which language you hope to learn. If your language isn't in Bubblegum yet, email us at support@bubblegum.la and tell us which one.