The Machines Learning to Read Cuneiform

Probably composed near the start of the first millennium BCE, the Hymn to Babylon later became a school text that Babylonian students copied as a writing exercise. It was a hit in its day, reproduced again and again on clay. Then Babylon fell, the clay broke, the pieces scattered across museums and excavation dumps, and the poem disappeared for the better part of two thousand years.

In 2025 it came back. Not because someone dug up a complete copy, but because a computer in Munich recognized that thirty broken fragments, sitting in different drawers, were all pieces of the same song.

The library that was one giant broken tablet

There is an old joke among Assyriologists: one day someone will realize there was only ever a single cuneiform tablet, and it just broke into a million pieces. The joke is really a description of the job. Cuneiform, the wedge-writing pressed into wet clay across the ancient Near East, survives mostly as fragments. A scholar might spend a career hunting for the shard that continues the text on the shard in front of them.

That is the problem Enrique Jiménez set out to attack at the University of Munich. Since 2018 his project, the Electronic Babylonian Library, has been photographing and transliterating the world's cuneiform, building a database that now holds something like 25,000 tablets and more than 350,000 lines of text. The clever part is what the software does with it. Borrowing a technique from the way biologists align strands of DNA, it compares every line against every other line and flags fragments that probably join.

The results are not subtle. Jiménez puts it plainly: in the first 150 years of cuneiform studies, scholars found roughly 6,000 places where fragments join. His project has found around 1,500 in just five years. That is the difference between a lifetime of squinting and a search that runs overnight.

The Hymn to Babylon was one payoff. The algorithm gathered thirty fragments from twenty different tablets, and out of them rose about two-thirds of a lost 250-line poem, including the first written testimony we have to the roles of Babylon's priestesses. A text that children once knew by heart, unknown to every modern scholar until a machine reassembled it.

What the machine can and cannot do

It is worth being exact about what "AI reads cuneiform" actually means, because the honest version is more interesting than the hype.

The machine is very good at matching and prediction. Alongside the fragment-joining, other teams have trained models to fill in small breaks in damaged tablets, guessing the missing signs from context with surprising accuracy, and even to produce rough English translations of Akkadian. But "rough" is the operative word. The best translation models score, by the standard automated measure, well below mature commercial translation, and they will happily produce fluent English that is simply wrong. They work best on short, formulaic sentences and stumble on literary flights.

And there is a deeper limit. Cuneiform signs are polyvalent: a single wedge-cluster can be read several different ways, and only context decides. That disambiguation, the actual reading, is the part humans still do. The machine finds the pieces and suggests the joins. A scholar decides what the text means.

Jiménez is refreshingly clear-eyed about this. Technology, he says, will never fully replace human scholarly work; its real gift is acceleration. The Euphrates passage his team recovered from the hymn makes the point better than any argument:

The Euphrates is her river, established by wise lord Nudimmud.
It quenches the lea, saturates the canebrake,
its meadows, in brilliant bloom, sprout barley.

No algorithm wrote that. An algorithm just found the barley, scattered across twenty broken tablets, and let a human plant it back in the field.

The quiet rescues

The best stories from this work are almost domestic. When the long-lost transliterations of a nineteenth-century scholar, Carl Bezold, were presumed gone forever, they turned up in a library in Heidelberg and were fed straight into the database, a dead man's handwriting given new life by software. Photographers work through the British Museum's holdings one tablet at a time, thousands of them, so the machine has something to read. The romance of the field is not a robot cracking a code. It is patient people and patient code, together, coaxing a hymn back into existence.

Five words from the clay

The language under all of this is Akkadian, the Semitic tongue of Babylon and Assyria, and it is more approachable than the wedges make it look:

• ṭuppu (TUP-pu), a clay tablet, the thing being reassembled.
• ṭupšarru (tup-SHAR-ru), a scribe, literally a "tablet-writer."
• Bāb-ili (BAB-ih-lee), Babylon, "the gate of the god."
• Purattu (pu-RAT-tu), the Euphrates, the river of the hymn.
• šarru (SHAR-ru), king, one of the first words any cuneiform student meets.

The machines are astonishing, and they are also a reminder of why the human skill still matters. Someone has to actually know Akkadian to turn a found fragment into a poem. That someone could be you, and it starts with a handful of words. PRAVIEL exists to carry you from those first signs toward reading the real thing, the way the scholars in Munich do, one recovered line at a time.

Sources

• Fadhil and Jiménez, "A Hymn in Praise of Babylon and the Babylonians" (Iraq, 2025)
• The Hymn to Babylon, reassembled with AI (phys.org)
• The Electronic Babylonian Library project (The Conversation)
• Translating Akkadian with neural machine translation (PNAS Nexus, 2023)
• Restoring broken Babylonian texts with neural networks (PNAS, 2020)