Cracking Cuneiform: How AI Is Reading 5,000-Year-Old Clay Tablets

More than half a million cuneiform tablets sit in museum drawers around the world, and by most estimates fewer than one in ten has ever been translated by a modern scholar. At the pace trained Assyriologists have historically worked - a few thousand tablets a year - clearing that backlog could take centuries. So in the last two decades, historians have started doing something unexpected: teaching computers to read 5,000-year-old clay.

You will learn what cuneiform is and how it was physically written. You will learn why so many tablets remain untranslated even today. You will learn how digital humanities projects use imaging and machine learning to speed up translation. You will practice transliterating a short cuneiform-style phrase yourself.

Cuneiform was invented in Sumer, in what is now southern Iraq, around 3400 BCE - making it one of the earliest writing systems in human history. Scribes pressed a cut reed stylus into wet clay tablets, leaving wedge-shaped marks (the word cuneiform comes from the Latin cuneus, meaning wedge). Early cuneiform started as simple pictographs used to track grain, beer rations, and labor. Over centuries it evolved into a system of roughly 1,000 abstract signs that could represent whole words or individual syllables, eventually recording everything from legal codes to the Epic of Gilgamesh, one of the oldest surviving works of literature.

The British Museum alone holds more than 130,000 cuneiform tablets, and similar collections sit in Iraq, Germany, France, and the United States. Reading cuneiform requires years of training in dead languages like Sumerian and Akkadian, and translating a single damaged or fragmentary tablet by hand can take a specialist days. There are only a few hundred trained Assyriologists in the world - nowhere near enough to work through the backlog in any of their lifetimes, let alone catalog tablets still being excavated.

Digital humanities researchers now use Reflectance Transformation Imaging (RTI), which photographs a tablet under dozens of light angles to make the shadows inside each wedge mark stand out clearly, even on worn or damaged clay. Machine learning models are then trained on thousands of already-transliterated tablets to recognize sign patterns and suggest readings for new ones, dramatically speeding up the first-pass transcription that used to take a human scholar hours per line. Other projects, like Fragmentarium, use image-matching algorithms to help scholars physically rejoin broken tablet fragments scattered across different museum collections - a puzzle that would be nearly impossible to solve by memory alone, since matching pieces might sit in museums thousands of miles apart.