Can AI crack the ancient writing systems that stumped linguists for a century?

Key points

• Linear A, the writing system of the Minoan civilisation on Crete, has never been deciphered and has no confirmed link to any known language.
• Etruscan, spoken in central Italy before the Roman Empire, has a partial vocabulary but its grammar remains unknown.
• Most ancient scripts were cracked using bilingual texts like the Rosetta Stone; neither Linear A nor Etruscan has one.
• Researchers are now testing whether large language models, the AI technology behind tools like ChatGPT, can find patterns humans have missed.

Every ancient writing system ever deciphered had something to grab onto. Usually that's a bilingual text like the Rosetta Stone, which showed the same decree in three scripts side by side. Sometimes it's a close linguistic cousin, a known language similar enough to help fill the gaps. Linear A, the Minoan script used on Crete, has neither, and a century of scholarship hasn't cracked it.

Linguists describe it as a "language isolate": no confirmed connection to anything else, alive or extinct.

Etruscan, spoken across central Italy before Rome took over, is in slightly better shape. Short funerary inscriptions have given researchers a partial vocabulary. But the grammar and deeper meaning are still missing.

Why are these two languages so hard to crack?

Without a bilingual anchor or a related language, there's nothing to cross-check against. It's like trying to solve a crossword with no clues and no grid.

Linear A tablets have been found, but most are inventory lists and accounting records, not the rich narrative texts that helped researchers work through Egyptian hieroglyphics. Etruscan inscriptions tend to be short and formulaic. Neither gives researchers the volume or variety of text that traditional methods need.

As reported by Ars Technica, linguists are now exploring whether AI, specifically large language models trained on vast amounts of human text, can spot structural patterns across thousands of fragmentary inscriptions faster than any human team could.

Could AI actually succeed where humans have not?

Possibly, for some pieces of the puzzle. AI is very good at finding statistical patterns across large datasets. Feed it every known Etruscan inscription and it can map which words cluster together, which grammatical forms repeat, which sounds appear at the start or end of words. That's genuinely useful groundwork.

But there's a real limit. A large language model needs something to learn from. With only a few thousand short Etruscan texts and zero confirmed translations of Linear A, the model's working with thin material. Garbage in, uncertain out.

Imagine asking a friend to learn a language from a stack of till receipts and nothing else. That's roughly the situation here.

This is the first time AI2Day has covered this particular corner of computational linguistics, picking up the thread on 29 July 2026 when we began tracking AI's reach into research domains far outside its original brief.

What does this mean for ordinary people?

No app will hand you a Linear A translation next Tuesday. This is long-horizon research. But it does matter: if AI helps linguists narrow the possibilities for even one of these scripts, it reframes what we understand about Bronze Age trade and religion across the Mediterranean.

Common questions

Has AI successfully deciphered any ancient language before?

Researchers used machine learning, a form of AI that finds patterns in data, to help confirm the relationship between the extinct language Ugaritic and Hebrew, and to assist with Mayan glyphs. Full decipherments still relied heavily on human expertise alongside the AI tools.

Is there a free tool anyone can use to explore ancient scripts?

Not for serious decipherment work. Academic projects and university databases hold the inscription records, and access is mainly for researchers. General AI chatbots can explain known scripts but can't crack unknown ones.

What would it take to finally crack Linear A?

A bilingual inscription showing Linear A alongside a known language would be the fastest route. Short of that, a much larger body of varied text would help enormously. Neither has turned up yet.