How AI and Digital Tools Are Racing to Save the World's Endangered Languages

The world is currently speaking roughly 6,700 distinct languages, but linguists are sounding a quiet, urgent alarm: nearly 40% of them are at risk of extinction by the end of the century. [1] When a language fades, it takes with it centuries of ecological knowledge, unique cultural worldviews, and ancestral history. [8][1][8]

For decades, the work of language preservation was a painstaking, analog process of recording elders on cassette tapes and compiling physical dictionaries. Today, a new wave of technologists and indigenous communities are turning to artificial intelligence to radically accelerate that timeline. [2][2]

Generative AI and machine learning models are now capable of transcribing, translating, and teaching languages that were previously entirely excluded from the digital realm. [2] However, the intersection of Silicon Valley technology and indigenous heritage has sparked a complex debate about who owns linguistic data and whether code can truly save a culture. [1, 7][1][2][7]

The fundamental technical hurdle for endangered languages has always been data scarcity. [5] Large language models like those powering ChatGPT or standard Google Translate require billions of words of text to learn a language's grammar and vocabulary. [1][1][5]

Languages that are primarily oral, or those spoken by marginalized communities, simply do not have that volume of digitized text. [1] In the AI industry, these are known as "low-resource languages," and until recently, they were effectively invisible to modern algorithms. [2][1][2]

That bottleneck is beginning to break. Meta's "No Language Left Behind" (NLLB) project recently released an open-source AI model capable of direct translations across 200 languages, including many low-resource tongues like Luganda and Māori. [2][2]

To achieve this without massive datasets, researchers utilized a "sparse mixture-of-experts" architecture, which routes translation tasks through specialized sub-networks to prevent the AI from overfitting on limited data. [2] Crucially, NLLB removes the need to translate a language into English first before translating it into a third language, preserving local idioms and cultural nuances. [2][2]

But while big tech companies are building massive, universal models, indigenous communities are increasingly demanding control over how their languages are digitized. [1] This movement, known as Indigenous Data Sovereignty, argues that communities—not for-profit corporations—must own and benefit from their linguistic heritage. [1, 8][1][8]

A pioneering example is unfolding in New Zealand. Te Hiku Media, a Māori broadcasting organization, refused to hand over its decades of archival audio to global tech giants. [3] Instead, they built their own automatic speech recognition (ASR) model using the open-source NVIDIA NeMo toolkit. [3][3]

By crowdsourcing voice recordings from the community under a strict "Kaitiakitanga" (guardianship) license, Te Hiku developed an AI that transcribes spoken Te Reo Māori with 92% accuracy. [3] The license legally ensures that the data can only be used in ways that benefit the Māori people, preventing commercial exploitation. [3][3]

The integration of AI into daily life is also proving vital for language visibility. In New Zealand, the media organization Stuff partnered with Microsoft and Straker Translations to build a custom AI tool that rapidly translates entire news articles into Te Reo Māori. [4][4]

Because Te Reo Māori is highly contextual—where a single word can carry multiple meanings depending on the speaker's intent—the AI does not publish directly. [4] Instead, it generates a high-speed draft that is then reviewed and refined by a human kaiwhakamāori (translator). [4] This human-in-the-loop system allows the newsroom to scale its bilingual output dramatically without sacrificing cultural accuracy. [4][4]

Beyond translation, AI is being used to resurrect scripts that are already on the brink of vanishing. At Dartmouth College, researchers developed "NüshuRescue," an AI framework dedicated to preserving Nüshu, a centuries-old script created by women in China's Hunan province to communicate in secret. [5][5]

Using just 35 pairs of matching sentences, the researchers trained a large language model to understand and translate the rare script. [5] The project demonstrated that modern AI can rapidly produce valuable linguistic resources even from microscopic amounts of training data. [5][5]

In Australia, digital tools are stepping in where human speakers have almost entirely disappeared. The Ngalia language, native to the Western Desert, currently has only three known fluent speakers remaining. [6][6]

To prevent the language from dying with them, developers launched the Mamutjitji Story app, an interactive educational tool that uses a local Dreamtime story about an antlion to teach Ngalia vocabulary and science concepts to children. [6][6]

Despite these technological triumphs, linguists and anthropologists caution against viewing AI as a panacea. [7] Critics argue that artificial intelligence cannot solve the material, social, and political problems that drive language endangerment in the first place. [7][7]

As anthropologist Gerald Roche points out, languages do not simply "die" of natural causes; they are often actively suppressed by state policies that defund minority schools, mandate dominant languages in workplaces, and marginalize indigenous populations. [7][7]

An AI app cannot build a bilingual school, pass protective legislation, or erase the systemic discrimination that forces a community to abandon its mother tongue for economic survival. [7, 8][7][8]

Ultimately, technology is a powerful amplifier, but it is not a substitute for human agency. [8] AI can archive the past and build bridges for the future, but the true survival of the world's endangered languages will depend on the communities fighting to speak them, and the political will to let those voices be heard. [1, 7, 8][1][7][8]