How Indigenous Engineers Are Using AI to Revitalize Endangered Languages

Language is the ultimate vessel of human culture, carrying unique worldviews, ecological knowledge, and ancestral histories. Yet, the world is currently experiencing a silent mass extinction of human expression. Of the roughly 6,700 languages spoken globally today, linguists estimate that at least 40 percent are endangered. The statistics are grim: on average, one Indigenous language dies every two weeks as its last fluent elder passes away. For decades, the standard response to this crisis was manual documentation—linguists recording elders on cassette tapes and painstakingly transcribing the audio into dictionaries. But this analog approach is fundamentally mismatched to the speed of the crisis. Now, a new generation of Indigenous engineers and computational linguists is turning to artificial intelligence to reverse the tide, transforming machines from tools of global homogenization into instruments of cultural resilience.[2][3]

The traditional bottleneck in language preservation has always been the sheer volume of labor required to process archival data. When media organizations or universities digitize old recordings of native speakers, the audio often sits in servers for years because only a handful of living people possess the hyper-specific regional fluency required to transcribe it accurately. Manual transcription can take up to ten hours for every one hour of audio. Artificial intelligence, specifically natural language processing (NLP), offers a way to clear this backlog at an unprecedented scale. By training machine learning models to recognize the phonetic patterns and grammatical structures of endangered languages, technologists are building automated speech recognition systems that can process decades of archival tape in a matter of days, unlocking a treasure trove of linguistic data for younger generations.[1][7]

The technological mechanism driving this revitalization is a shift away from massive, generalized AI toward Small Language Models (SLMs). Large Language Models (LLMs) like GPT-4 or Gemini rely on billions of parameters and require massive datasets—often millions of parallel sentence pairs scraped from the internet. This brute-force approach works flawlessly for high-resource languages like English or Spanish, but it fails spectacularly for Indigenous languages that have little to no digital footprint. SLMs, by contrast, are designed to be highly efficient and targeted. They can be trained or fine-tuned on much smaller, meticulously curated datasets. This allows developers to build functional translation tools, spellcheckers, and word predictors using only a few hundred hours of audio or a few thousand pages of text, making AI accessible to communities that Big Tech's massive models have historically left behind.[3]

The efficiency of Small Language Models also solves a critical infrastructure problem for remote Indigenous communities. Because SLMs require significantly less computational power than their massive counterparts, they can be deployed locally on affordable hardware or mobile devices without requiring a constant, high-bandwidth internet connection. This is particularly vital for communities in rural areas or low-and-middle-income countries where cloud-based AI is impractical. Furthermore, technologists are using a technique called chain-of-thought distillation, where a larger model's reasoning capabilities are compressed into a smaller, language-specific model. This creates an AI that is not only highly accurate within its specific linguistic context but also transparent—allowing human linguists to verify the model's internal processing steps and ensure that cultural nuances are being handled correctly.[3]

One of the most successful blueprints for this approach comes from New Zealand, where the Māori media organization Te Hiku has pioneered the use of AI for language revitalization. Decades of colonial repression had severely threatened Te Reo Māori; by 1960, only one in four Māori spoke their native tongue. In 2014, Te Hiku began digitizing a vast archive of recordings from Māori elders born in the late 19th century. Realizing that manual transcription would take lifetimes, Chief Technology Officer Keoni Mahelona and CEO Peter-Lucas Jones decided to build their own bespoke natural language processing tools. They launched an open-source app called Kōrero Māori, crowdsourcing voice donations from the community to build a robust dataset that accurately reflected the language's unique cadence and regional variations.[1][4]

The community response to Kōrero Māori was overwhelming. In just the first ten days of the app's launch, Te Hiku's data science team collected over 300 hours of annotated audio recordings. Using this hyper-local dataset, they trained an automatic speech recognition system that vastly outperformed commercial models built by Silicon Valley giants. Today, Te Hiku uses this AI to transcribe historical broadcasts, power pronunciation apps like Rongo, and develop voice assistants that allow New Zealanders to interact with everyday technology entirely in Te Reo Māori. By proving that high-quality AI can be built from the ground up by the community it serves, Te Hiku has provided a scalable template for other Indigenous groups worldwide.[1]

In North America, similar community-led initiatives are racing against a ticking clock. Michael Running Wolf, a Northern Cheyenne engineer and founder of the First Languages AI Reality (FLAIR) initiative, notes that the United States stands to lose the majority of its Native American languages within the next decade if drastic action is not taken. FLAIR, operating in collaboration with the Mila-Quebec Artificial Intelligence Institute, is currently building speech recognition models tailored for over 200 endangered Indigenous languages across the continent. Running Wolf's team is actively training a new cohort of Native American, Alaska Native, and Native Hawaiian computer scientists, ensuring that the people building the technology intimately understand the cultural weight of the data they are processing.[2]

The movement is also gaining momentum in South America. In Colombia, researchers at the Universidad de los Andes are developing machine translation models for low-resource languages like Wayuunaiki, Nasa Yuwe, Inga, and Arhuaco. Working directly with native speakers and expert translators, the team gathers high-quality translations of short phrases to fine-tune high-resource language models. In 2025, their datasets were officially included in the AmericasNLP initiative, marking a historic milestone for Colombian Indigenous representation in computational linguistics. These tools are not designed to replace human learning, but to create digital bridges—allowing a young Wayuu student to instantly translate educational materials or government documents into their ancestral tongue.[5]

However, the intersection of artificial intelligence and Indigenous heritage is fraught with ethical tension. The primary concern is data sovereignty—the right of Indigenous communities to control how their languages and cultural knowledge are collected, stored, and monetized. Historically, Western academic and corporate institutions have treated Indigenous data as a public resource to be extracted. Today, many communities fear a new wave of digital colonialism, where Big Tech companies scrape Indigenous language data to train commercial AI models without providing compensation or attribution. When a massive tech conglomerate absorbs an endangered language into a global LLM, the community loses control over how their ancestors' voices are used, potentially exposing sacred or restricted knowledge to the public domain.[6]

To combat this, Indigenous technologists are pioneering new legal and ethical frameworks for AI development. Te Hiku Media, for example, licenses its technology under the Kaitiakitanga license. Rooted in the Māori concept of guardianship, this license explicitly prohibits the use of the community's data by commercial tech giants or for any applications that violate human rights. It ensures that the data remains the sovereign property of the Māori people. By rejecting the Silicon Valley ethos that all data wants to be free, these communities are proving that AI development can be rigorous, innovative, and deeply respectful of cultural boundaries. They are building walled gardens where their languages can flourish safely.[4]

The necessity of community oversight was starkly illustrated in December 2024, when a series of AI-generated language-learning books appeared on major online retail platforms. The books claimed to teach endangered languages like Abenaki, Mohawk, and Omok—a Siberian language that has been extinct since the 18th century. However, Indigenous linguists quickly discovered that the AI had hallucinated the content, filling the books with fabricated words, incorrect grammar, and nonsensical translations. For communities fighting desperately to preserve the authenticity of their heritage, these AI-generated fabrications were not just inaccurate; they were actively harmful, threatening to pollute the limited pool of learning resources with algorithmic noise and undermining the trust of new learners.[4]

This incident underscores a fundamental truth recognized by Indigenous engineers: artificial intelligence is a tool for preservation, not a replacement for human connection. Language is a living, breathing ecosystem that requires human relationships, cultural context, and intergenerational transmission to survive. An AI chatbot can simulate a conversation in Cherokee or Navajo, but it cannot impart the spiritual significance of the words or the ancestral stories attached to them. Technologists emphasize that digital tools must be designed to complement, rather than supplant, traditional practices like immersive classroom learning and time spent with community elders. The goal is not to create a perfect digital archive of a dead language, but to lower the barrier to entry for the living.[2][6]

Looking ahead, the synthesis of AI and Indigenous knowledge promises a profound shift in how marginalized languages interact with the modern world. Researchers are developing real-time translation tools that could soon allow a native speaker of an endangered language to navigate the internet, access healthcare, or interact with government services without having to default to English or Spanish. Educational chatbots are being designed to act as tireless, interactive tutors, providing young learners with a safe space to practice their ancestral language without the fear of making mistakes in front of fluent elders. These innovations are transforming the digital landscape from a threat to linguistic diversity into its most powerful ally.[3][6]

Ultimately, the application of artificial intelligence to endangered languages represents a paradigm shift from preservation as static archival storage to preservation as active, participatory engagement. By taking ownership of the technology, Indigenous communities are ensuring that their languages do not merely survive in museum databases, but thrive in the digital media ecosystems of the 21st century. The work of these engineers and linguists proves that the future of artificial intelligence does not have to be an English-only monolith. When guided by the communities themselves, AI can help resurrect the voices of the past and secure them for the generations to come.[6][7]