How AI is Decoding the World's Lost Ancient Languages

The sheer volume of untranslated history is staggering. Museums and universities worldwide hold more than half a million clay tablets inscribed with cuneiform, the wedge-shaped writing system of ancient Mesopotamia. Yet, because the Akkadian language has not been spoken for 2,000 years, only a few hundred experts globally can read them.[3][4]

This creates a profound human bottleneck. The vast majority of these documents—detailing the political, economic, and daily lives of early civilizations—remain locked away, untranslated and inaccessible. But a new wave of artificial intelligence is beginning to clear the backlog, acting as a digital Rosetta Stone for antiquity and unlocking millennia of lost human history.[1][3][6]

The breakthrough relies on Neural Machine Translation (NMT), the same underlying technology that powers modern translation apps. A team of computer scientists and archaeologists from Tel Aviv University and Ariel University recently trained an AI model on a massive dataset of digitized cuneiform texts from the Open Richly Annotated Cuneiform Corpus.[1][4]

Translating Akkadian is notoriously difficult because the language is polyvalent; a single cuneiform sign can have multiple different readings depending on its function in a sentence. Traditionally, scholars must first transliterate the cuneiform into Latin script based on phonetics, and then translate that transliteration into English.[4]

The Israeli research team taught their AI to perform both tasks. The model learned to translate into English from both the Latin transliterations and directly from the Unicode representations of the cuneiform symbols themselves. The results were evaluated using the Best Bilingual Evaluation Understudy 4 (BLEU4) metric, a standard tool for measuring machine translation accuracy.[1][3]

The AI achieved a BLEU4 score of 37.47 when translating from transliterations, and 36.52 when translating directly from cuneiform. In the world of machine translation, a score around 37 is considered highly effective for an early-stage model, providing a translation that is entirely usable as a first-pass understanding of the text.[1][3]

However, the system is not flawless. The AI occasionally produces "hallucinations"—sentences that are syntactically perfect in English but completely misrepresent the original Akkadian meaning. Because of this, researchers advocate for a "human-in-the-loop" approach, where the AI rapidly processes thousands of tablets to find relevant texts, which human experts then refine for cultural nuance.[1][4]

While NMT is translating readable tablets, a different branch of AI is tackling ancient texts that cannot even be opened. In 79 AD, the eruption of Mount Vesuvius buried the Roman town of Herculaneum in volcanic ash, carbonizing an entire library of papyrus scrolls.[2][5]

For centuries, these scrolls presented an impossible puzzle. The papyrus is so fragile that any physical attempt to unroll it causes the carbonized layers to crumble into dust. To solve this, a global initiative called the Vesuvius Challenge was launched, offering substantial cash prizes to anyone who could read the scrolls using non-invasive technology.[2][5]

The process begins at a particle accelerator, such as the Diamond Light Source in the UK, which blasts the scrolls with powerful X-rays to create microscopic 3D scans of their internal structure. The challenge then shifts to software engineers, who use AI autosegmentation algorithms to digitally peel apart the crumpled layers of the 3D scan.[2][5]

Because the ancient Romans used carbon-based ink, it does not show up brightly on an X-ray like metal-based ink would. Instead, machine learning models are trained to detect the microscopic textural changes on the papyrus where the ink was applied. In 2023, a 21-year-old engineering student named Luke Farritor became the first to successfully train an AI to spot these faint Greek letters, winning a portion of the $700,000 grand prize.[5]

The technology is advancing at a staggering pace. In early 2025, the Bodleian Libraries and the Vesuvius Challenge announced a historic milestone: researchers had successfully generated an image of the inside of a scroll and identified its title. The text, "On Vices," was written by the Greek philosopher Philodemus, offering new insights into ancient ethical guidance.[2]

Beyond translating known languages and unrolling fragile scrolls, AI is also being deployed against the ultimate linguistic puzzle: completely undeciphered scripts. Languages like Minoan Linear A and the script of the Indus Valley Civilization have baffled linguists for decades because there is no bilingual text to serve as a translation key.[7]

To crack these codes, researchers at MIT and Google have developed algorithms that analyze the structural sparsity and distributional similarity of the unknown symbols. By mapping how frequently certain characters appear together, the AI can compare the mathematical structure of the mystery language to the structures of known, related languages.[7]

This method has already proven successful on scripts we do understand. When tested on Linear B—an early form of Greek deciphered by humans in the 1950s—the AI was able to algorithmically map the phonetics and accurately translate the text. The algorithm achieved similar success with Ugaritic, a 3,000-year-old cuneiform language.[7]

Applying this to isolated languages like the Indus Valley script remains a monumental challenge, as the AI must find a statistical match without knowing which language family the script belongs to. Yet, the sheer computational power of modern machine learning offers the first real hope of breaking these glyphs.[6][7]

The implications of these technologies extend far beyond academic curiosity. Language is the ultimate artifact of human consciousness. When a writing system is deciphered, an entire civilization speaks again, allowing us to discover ancient peoples through their own laws, poetry, and daily correspondence.[6][7]

As AI models grow more sophisticated and 3D imaging becomes more precise, the bottleneck of historical translation is rapidly disappearing. We are entering a new Renaissance of classical antiquity, one driven not by the discovery of new ruins, but by the silicon and code that can finally read the ruins we already have.[6]