The AI That Speaks Your Language, Not Just Your Words
Ayah Bdeir doesn't want AI to speak your language.
She wants it to hold your culture.
It's a quiet difference, but it's the whole point. You can teach a model to say "kuch bhi" or "mangwana" or "kabu"—but if it doesn't know what that word means to your grandmother, to your harvest, to your prayers, then it's just noise. Current AI is building the infrastructure to fix that.
Founded in February 2025 by Martin Tisné, the nonprofit is racing to create something no one else is: a public, open, community-owned alternative to the corporate AI monopolies. Not another chatbot. Not a faster model. A new kind of foundation.
"If AI is going to change every part of life," Bdeir told TechCrunch, "there has to be a public version. Like the World Wide Web—free for everyone."
And right now, the web of AI belongs to OpenAI, Google, Anthropic. Their models are trained on scraped data, often without consent. For Indigenous languages, that means missionary Bible translations become training data before communities have a say. English dominates—not because it's the richest, but because it's the easiest to scrape.
Half the world's spoken languages are vanishing. And with them, entire ways of knowing. The way a farmer in rural India identifies a dying plant by its leaf curl. The way an elder in the Amazon remembers flood patterns through song. That knowledge isn't in textbooks. It's in dialects. In silence. In gestures.
Current AI's answer? Don't train on data. Train with people.
Suno Sutra: AI That Doesn't Need the Internet
In February 2026, at the India AI Summit, Current AI partnered with Bhashini, India's government-led AI language initiative, to launch Suno Sutra—Hindi for "listening chronicles."
It's a pocket-sized device. No screen. No Wi-Fi. Just a button, a speaker, and a 22-language AI that works offline.
A farmer snaps a photo of a wilted crop. She presses the button. The device speaks back in her mother tongue: "This is fungal blight. Try neem oil and stop watering at dusk." No Google search. No English interface. No corporate tracking.
And it's open source. Any developer in a village in Odisha or a tech hub in Bengaluru can fork it, add a new dialect, or build a version for cattle herders in Rajasthan.
"In India, there are hundreds of languages," Bdeir said. "Right now, AI doesn't represent them. It pretends to."
Suno Sutra isn't a product. It's a platform. And it's already being rewritten by local coders who didn't wait for permission.
The $3.2 Million That Changed Everything
Last month, Current AI distributed $3.2 million in grants—not to startups, not to labs, but to four community-led projects across Kenya, Lebanon, and the Brazilian Amazon.
In Kenya, Masakhane is building AI datasets for over 50 African languages—focused on farming advice, maternal health, and school curricula. Not translated from English. Built from scratch, with elders, nurses, and agronomists as co-authors.
In Lebanon, the Institute for Worldmaking is digitizing Arab cultural history—poetry, oral legal traditions, folk medicine—into machine-readable formats controlled by local councils, not Silicon Valley servers.
In Brazil, Portal sem Porteiras (meaning "Gateless Portal") is working with Indigenous communities in the Amazon to create offline AI tools that help record and pass down ecological knowledge in their native tongues. No cloud. No API keys. Just a solar-charged tablet and a voice recorder.
And in Kenya again, the African Internet Rights Alliance is developing AI audit tools—so communities can ask: "Who trained this? What data was used? Can we stop it?"
These aren't pilot programs. They're prototypes of a new model.
"The usual default," Bdeir says, "is complexity as an excuse. Let the government or the tech company decide for everyone."
Current AI's rule? No project moves forward without community consent at every stage.
Data Ownership Isn't a Feature. It's the Foundation.
"Who owns the data?" I asked Bdeir.
She didn't pause.
"It shouldn't be a company in Silicon Valley trying to make a few thousand people wealthier."
Current AI doesn't collect your data. It doesn't even store it. Models and datasets live locally—on a tablet in a village clinic, a USB drive in a community center, a Raspberry Pi in a remote school.
Before any tool is built, community experts are invited in—not as consultants, but as co-designers. They decide what gets recorded, how it's labeled, who can access it.
And they can shut it down. Anytime.
That's radical. In tech, "user consent" usually means a checkbox you skip. Here, consent is a process. A relationship.
"We're not trying to build the best AI," Bdeir said. "We're trying to build the right one."
AlphaChat: The Open Stack That Wasn't Supposed to Exist
In seven weeks, a coalition of ten organizations—including Hugging Face, Mozilla, and MIT Media Lab—built AlphaChat.
An open-source AI chatbot.
No corporate backing. No venture funding. Just shared code, shared ethics, shared servers.
Each partner brought a piece: one provided the language model tuned for low-resource languages. Another added safety filters trained on local taboos. A third donated computing power from a university in Nairobi.
It's not perfect. Sometimes it mispronounces Swahili. Sometimes it gives overly literal answers. But it's alive. It's editable. It's yours.
And it's already being used by teachers in rural Ghana to explain math concepts in Twi.
"Big tech builds multilingual models to expand their market," Bdeir said. "We build them because the world is bigger than their business plan."
The Tokyo Connection: Sovereign AI for the Global South
Earlier this month, Current AI struck a deal with Sakana AI, a Tokyo-based startup known for its work on what it calls "Sovereign AI."
Their goal? Build a shared open-source stack—not just for Japanese, but for the Global South.
Why Tokyo? Because Japan has its own history of cultural erasure in tech. Its language models were trained on colonial-era texts, ignoring regional dialects and oral traditions. Sakana AI is trying to fix that.
Now, they're doing it together.
The stack will include:
- A lightweight language model trained on Indigenous Japanese poetry and Ainu oral histories
- A consent protocol designed for community-controlled data
- A decentralized storage layer that lets villages host their own models
This isn't about competing with GPT-5. It's about creating a new kind of AI—one that doesn't need to be big to be meaningful.
Scale Isn't the Goal. Presence Is.
"People ask me," Bdeir said, "how much impact can $3.2 million have?"
Her answer: "What if an Indigenous elder in the Amazon uses a tool built in Kenya to teach her grandchild how to read the stars?"
That's the metric.
Not user counts. Not revenue. Not model size.
Presence.
When a language dies, we don't just lose words. We lose ways of seeing. Ways of healing. Ways of remembering.
Current AI doesn't want to dominate AI.
It wants to make sure no culture gets left behind when AI finally arrives.
And for the first time, that's not a dream.
It's a device in a farmer's pocket. A chatbot in a schoolhouse. A protocol written by elders.
The web of AI is being built—not by venture capital, but by communities.
And it's already speaking.
