Big Tech's LLMs can speak Māori fluently because they scraped it without asking—now a New Zealand research team is building the antidote.
The Summary
- Te Taka Keegan and Kingsley Eng at University of Waikato built a text-to-speech model for te reo Māori with one non-negotiable constraint: the dialect speakers own everything
- ChatGPT, Claude, and Perplexity already speak fluent Māori by scraping community data without permission, then processing it offshore
- The project demonstrates that AI sovereignty isn't just possible, it's necessary when your culture's primary knowledge carrier is at stake
The Signal
Te reo Māori is spoken fluently by 4.3 percent of New Zealanders. About 30 percent can manage a few words or phrases. Yet the language is perfectly accessible to anyone with a ChatGPT account. OpenAI scraped Māori text and audio from communities and academics, processed it in American data centers, and now serves it back through an interface no Māori community controls.
For Keegan, this isn't a minor IP issue. "Our language is the most important conveyor we have for our knowledge," he told IEEE Spectrum. "Yet we see technology developed outside of Aotearoa get more and more control over the transfer of that knowledge."
"These companies overseas have the resources to produce AI models that work well. But they scraped all of that data with no input from us, and we don't own the output."
The response was a text-to-speech system built on what Keegan calls "sovereign digital systems." Every technical choice, from data collection to model architecture, was shaped by one constraint most AI labs never consider: the people who speak this dialect must own the synthetic voice and everything used to build it.
What sovereignty actually requires:
- No cloud training on AWS, Azure, or Google infrastructure
- No scraped datasets of unknown provenance
- Community consent at every stage, not just a terms-of-service checkbox
This is the collision point between Web2's extractive data economy and Web4's agentic future. Large language models are trained on billions of tokens scraped from the open web. For English, that's mostly public domain and fair use arguments. For indigenous languages, it's cultural artifacts uploaded by communities who never imagined their grammar guides and oral histories would become training data for a San Francisco startup.
Keegan and Eng's model is designed for a specific dialect, not the standardized Māori taught in schools and broadcast on TV. That specificity matters. When you homogenize a language into its most common form, you erase the local knowledge embedded in regional variations. You get fluency without cultural fidelity.
The project is part of a broader shift. Indigenous communities globally are realizing that if they don't build their own AI infrastructure, someone else will build it for them and charge rent forever. Sovereignty in the agent economy isn't about open source versus proprietary. It's about who controls the data, who runs the compute, and who decides what gets built.
The Implication
This is the template for communities who refuse to be data colonies. If you're building AI for a language, a culture, or a knowledge system that Big Tech didn't bother to ask permission to scrape, watch what Keegan and Eng did. Own the data. Own the model. Own the infrastructure. Don't outsource sovereignty to the cloud.
The agent economy will be built on training data. If you don't control yours, you're renting your own culture back from someone who scraped it for free.