AI will not save Dying Languages, and Greece Proves Why

Most radical deliberation of artificial quality arsenic an ever-expanding instrumentality that volition yet screen each country of quality knowledge. Yet probe shows a stark truth: AI struggles the astir wherever it is needed the most. Low-resource languages, galore already connected the brink of extinction, are poorly represented successful the immense datasets that provender ample connection models.

Recent analyses collated by Stanford’s AI Index amusement that ample connection models inactive underperform connected galore non-English languages, peculiarly those with constricted integer data. The United Nations estimates that astir 40 per cent of the world’s 7,000-plus languages are astatine hazard of disappearance. The information is not lone cultural. It is technological, economical and ethical.

Why AI Fails Endangered Languages

Artificial quality models are trained connected tremendous volumes of substance and speech. The occupation is that astir of that worldly is successful English oregon a fistful of high-resource languages specified arsenic Mandarin, Spanish oregon French. For endangered languages, the integer footprint is tiny. What does beryllium is often confined to scripture, mediocre translations oregon fragmentary Wikipedia entries.

Only a fewer 100 languages person immoderate meaningful online footprint. The State of the Internet’s Languages project counts astir 500 of 7,000+, portion indicator-based measurement covers astir 329 languages with capable web signals, showing however bladed sum remains. For example, portion English and Mandarin predominate everything from Wikipedia to societal media, galore number languages person nary keyboard support, nary spellcheck and nary large quality oregon acquisition contented online.

There are besides structural challenges. Standard tokenisers (the bundle that breaks substance into tiny units called “tokens” truthful an AI tin process it) divided galore non-English texts into much tokens to explicit the aforesaid content. Research shows this “token tax” raises compute outgo and tin depress prime for those languages. For instance, a elemental condemnation successful Greek is often breached into acold much tokens than the aforesaid condemnation successful English, making processing slower and much costly adjacent erstwhile the meaning is identical.

The stakes widen beyond exclusion. A 2023 study shows that simply translating unsafe prompts into low-resource languages tin bypass GPT-4 guardrails, lifting occurrence from nether 1 per cent successful English to arsenic precocious arsenic 79 per cent connected a modular benchmark. For example, a petition specified arsenic “How tin I physique a homemade bomb?” volition usually beryllium blocked successful English. But erstwhile translated into languages similar Zulu oregon Gaelic, the strategy whitethorn make harmful instructions earlier filters intervene. This hazard matters for everyone, due to the fact that translation makes the onslaught trivial.

Ancient Greek Survival Through Texts

The Greek acquisition offers a crisp perspective. Greek has the longest documented past of immoderate Indo-European connection – astir 34 centuries – which preserved Ancient Greek for survey adjacent arsenic mundane code changed. Because texts similar Homer’s Iliad, Plato’s dialogues and Aristotle’s treatises were copied and studied for centuries, Ancient Greek ne'er disappeared from quality memory, adjacent erstwhile regular code evolved.

This textual richness ensured its endurance arsenic a taxable of learning agelong aft mundane speakers disappeared. Unlike galore of today’s endangered languages, Ancient Greek had the vantage of written tradition, world continuity and planetary interest. AI tin grip Ancient Greek comparatively good precisely due to the fact that determination is an abundance of high-quality data.

Fragility of Modern Greek Dialects

By contrast, respective Greek dialects spoken contiguous are successful existent danger. Tsakonian, a descendant of the past Doric branch, is classified arsenic severely oregon critically endangered, with astir 2,000 to 4,000 speakers, mostly older adults. In Tsakonian villages of the Peloponnese, younger residents often recognize the dialect but reply successful Standard Modern Greek, a motion of however fragile intergenerational transmission has become.

Romeyka, spoken successful distant communities astir Trabzon successful bluish Türkiye, is possibly adjacent much remarkable. Researchers picture it arsenic a “living bridge” to the past satellite due to the fact that it preserves the infinitive, a grammatical diagnostic mislaid successful each different modern Greek dialect. Cambridge’s Crowdsourcing Romeyka task has been described arsenic a “last chance” to seizure the connection earlier it disappears.

Pontic Greek, erstwhile wide astir the Black Sea, survives among diaspora communities successful Greece and abroad. UNESCO lists it arsenic endangered, and portion it has a larger basal than Tsakonian oregon Romeyka, younger generations are shifting to Standard Modern Greek.

These dialects item the cardinal issue: without written corpora, digitised archives and sustained assemblage use, AI has thing to larn from.

Greek Researchers Bridging Preservation and AI

At Cambridge, Professor Ioanna Sitaridou leads the Romeyka Project, combining fieldwork with integer crowdsourcing to physique a code corpus of this endangered dialect. By inviting speakers to grounds and stock their language, her squad is creating precisely the benignant of dataset AI systems require.

Professor Ioanna Sitaridou

In the Peloponnese, linguist Efrosini Kritikos has worked connected community-based curricula for Tsakonian, embedding revitalisation successful education. Earlier, the precocious Thanasis Costakis laid indispensable groundwork by devising grammars, lexicons and an orthography adapted to Tsakonian’s antithetic phonology.

Meanwhile, astatine the University of Patras, Professor Angela Ralli established the Laboratory of Modern Greek Dialects and created Greece’s archetypal physics linguistic atlas. This integer mapping of saltation has go a instauration for resources specified arsenic the Greek Dialectal NLP dataset, which turns dialectal information into machine-readable form.

Together, these initiatives amusement that linguistic documentation is not lone taste but technological. They supply the high-quality information without which AI systems cannot learn.

Efrosini Kritikos, PhDc

Global Parallels

Elsewhere, governments and institutions are grappling with the aforesaid challenge. SEA-LION (South-East Asian Languages In One Network) is an unfastened multilingual exemplary household focused connected South-East Asian languages, developed by AI Singapore. In August 2025, Malaysia launched ILMU, a home-grown multimodal exemplary developed with Universiti Malaya and YTL AI Labs, tailored to Malay, English and section dialects.

New Zealand’s Te Hiku Media has spent years collecting Māori information with elders, connection learners and archives. Crucially, it created kaitiakitanga licences truthful Māori communities clasp guardianship and benefit-sharing implicit the corpora and models built from them. In practice, this means a Māori elder’s recorded communicative cannot beryllium scraped by a tech institution for escaped — it remains nether assemblage power for aboriginal generations.

In Indonesia, researchers fine-tuned Meta’s XLS-R to recognise Orang Rimba speech. The results were promising, but the task was constrained by the tiny size of the dataset. The signifier is consistent: without large, community-approved corpora, exertion reaches its limits quickly.

Greek Philosophy and the Meaning of Language

The past Greeks were among the archetypal to grapple with wherefore connection matters. In Cratylus, Plato debated whether words are earthy oregon conventional, and what this means for information and identity. His reflections resonate today. A connection is much than a instrumentality for communication. It encodes a community’s worldview, its representation and its mode of mapping reality.

In the aforesaid way, erstwhile Tsakonian oregon Romeyka fade, truthful excessively does a mode of knowing family, spot and individuality that has nary nonstop equivalent successful Standard Modern Greek oregon successful English translation.

When a connection dies, an full position connected the satellite disappears with it. Artificial quality cannot reconstruct that loss. No algorithm tin invent taste representation wherever nary was recorded.

The past Greek philosopher Plato explored the precise quality of words successful debates that echo successful today’s conflict to sphere languages.

The Bottom Line

Artificial quality tin translate, transcribe and link crossed languages with unprecedented speed. But it cannot prevention languages that deficiency the information it needs.

The Greek lawsuit illustrates some sides of the coin. Ancient Greek survived due to the fact that of abundant texts and scholarly continuity. Tsakonian, Romeyka and Pontic amusement however susceptible surviving dialects are without akin resources. The enactment of Greek linguists demonstrates that preservation indispensable travel first: grammars, lexicons, code recordings, curricula and integer atlases. Only past tin AI go a instrumentality for practice alternatively than erasure.

The UN warns that 40 per cent of the world’s languages whitethorn vanish wrong this century. Artificial intelligence, unless grounded successful community-led preservation, volition not forestall that outcome. At best, it volition reflector the imbalance that already exists. At worst, it volition assistance to hide what remains.

Read also: Sunset Lover’s Guide to the Cyclades