How Indigenous Technologists Are Rewiring AI to Save Endangered Languages

For decades, the internet has operated as a linguistic homogenizer. With more than half of all web domains written in English, the digital age has inadvertently accelerated the decline of minority languages. According to the United Nations, roughly 40 percent of the world's 7,000 languages are currently at risk of extinction, with one Indigenous language lost every two weeks. But a new generation of Indigenous technologists is flipping the script, transforming artificial intelligence from a tool of cultural erasure into an engine for language revitalization.[1][2][3][5]

One of the most visible examples of this shift sits on the shoulders of children. Danielle Boyer, a 24-year-old Anishinaabe roboticist, recently designed the "Skobot"—a small, brightly colored wearable robot shaped somewhat like a parrot. Equipped with an internally developed AI model, the motion-activated toy converses fluently in Anishinaabemowin, the endangered language of the Anishinaabe nation in North America. When a child asks the robot how to say a specific word, the AI interprets the audio and responds in real-time, simulating a natural, immersive conversation that is often missing in modern digital environments.[2]

Innovations like the Skobot are necessary because the broader AI industry has largely left the global majority behind. A 2025 paper from the Stanford Institute for Human-Centered Artificial Intelligence highlighted a structural disparity known as the "global language data gap." Mainstream large language models (LLMs) developed by major tech firms rely heavily on publicly available text scraped from the internet. Because the web is overwhelmingly English-centric, these models perform exceptionally well in dominant languages but fail spectacularly when tasked with anything else.[1][3][4]

In the field of artificial intelligence, languages outside this dominant cluster are termed "low-resource." This designation has nothing to do with the number of native speakers a language has; even widely spoken languages like Urdu fall into this category. Instead, "low-resource" refers strictly to a scarcity of machine-readable, digitized, and annotated text available to train algorithms. Without this massive corpus of data, standard AI transcription and translation tools struggle with cultural nuance, introduce inherent biases, and frequently hallucinate incorrect grammar.[3][4][6]

To bridge this gap, Indigenous communities are taking data collection into their own hands, refusing to wait for Silicon Valley to notice them. In New Zealand, Te Hiku Media, an iwi-led (tribal) broadcasting organization, recognized that te reo Māori needed a robust digital presence to survive. Rather than relying on existing tech giants, they launched "Kōrero Māori," a massive crowdsourcing initiative designed to build a custom automatic speech recognition (ASR) model from scratch.[1][5]

The response from the Māori community was unprecedented. In just ten days, over 2,500 individuals signed up to read more than 200,000 phrases, generating over 300 hours of highly accurate, labeled speech data. Using the open-source NVIDIA NeMo toolkit and advanced tensor core GPUs, Te Hiku Media trained a speech-to-text model that now transcribes te reo Māori with 92 percent accuracy. It can even seamlessly transcribe bilingual speech, switching between English and te reo with an 82 percent accuracy rate.[5]

The success of Te Hiku Media is not just a technical triumph; it is a blueprint for Indigenous data sovereignty. Historically, marginalized communities have seen their cultural artifacts and knowledge extracted and monetized by outside entities. To prevent this, Te Hiku Media collected its data under a strict "Kaitiakitanga" (guardianship) license. This legal and cultural framework ensures that the data, and the AI models built from it, remain under Māori control and are used exclusively for the benefit of the Māori people.[1][5][6]

This insistence on sovereignty is reshaping how AI research is conducted globally. At the prestigious NeurIPS AI conference, recent workshops have centered entirely on building LLM architectures tailored to low-resource linguistic features through ethical, community-centered dataset collection. Researchers are moving away from brute-force data scraping and toward "transfer learning"—a technique where an AI model applies the underlying structural knowledge it learned from a high-resource language to a low-resource one, drastically reducing the amount of native data required.[6]

Institutions like Mila, the Quebec AI Institute, are pushing this further through their First Languages AI Reality (FLAIR) initiative. Developing an ASR model for a new language typically requires hundreds of hours of pristine audio. FLAIR is pioneering foundational research to create custom voice models for endangered languages using a fraction of that data. These lightweight models can then power voice-controlled technology, audio transcription, and immersive virtual reality experiences for Indigenous youth.[8]

The shift from static preservation to dynamic interaction is critical for intergenerational transfer. Younger speakers engage primarily through smartphones and interactive media, making traditional, static digital archives less effective. To address this, New Zealand-based software company Kiwa Digital partnered with Amazon Web Services to launch CultureQ, a generative AI platform. By embedding conversational AI into cultural archives, users can ask questions and hear the language spoken aloud, turning historical texts into living dialogues.[7]

Despite these breakthroughs, significant technical and ethical uncertainties remain. AI models, by their nature, recognize patterns and calculate probabilities; they do not "understand" culture. Linguists warn that general-purpose AI can inadvertently simplify or misrepresent Indigenous knowledge, stripping away the deep contextual tones and morphologies that give these languages their meaning. There is a persistent fear that synthetic data generation—using AI to create artificial training text—could slowly dilute the authenticity of the language over time.[3][4][6]

To mitigate these risks, the consensus among researchers and community leaders is that AI must remain a "human-in-the-loop" technology. In newsrooms and classrooms experimenting with these tools, "hybrid translation"—where AI outputs are rigorously reviewed by native speakers before publication—is becoming the gold standard. The goal is not to replace human teachers or elders, but to give them infinitely scalable tools to amplify their reach.[3][7]

The implications of this work extend far beyond linguistics. Studies have shown that a strong connection to linguistic heritage correlates with tangible public health benefits in Indigenous communities, including lower rates of teen suicide, diabetes, and excessive alcohol consumption. Language is the vessel for identity, and preserving it has a profound stabilizing effect on a community's social fabric.[2][6]

By forcing cutting-edge technology to adapt to their needs, Indigenous technologists are proving that the future of AI does not have to be a monolith. From wearable robots in Michigan to sovereign data centers in New Zealand, these initiatives demonstrate that with the right ethical frameworks and community leadership, artificial intelligence can be harnessed to protect the very diversity it once threatened to erase.[1][2][5]