An artificial-intelligence system has read a Babylonian law tablet at 98 percent character accuracy, raising hopes that tens of thousands of clay tablets still lying untranslated in museums could soon be opened to scholars and the public alike.
The breakthrough is described in a study uploaded on 7 May 2025 to the open-access arXiv server by University of Dubai researchers Shahad Elshehaby, Alavikunhu Panthakkan, Hussain Al-Ahmad and Mina Al-Saad. Their paper, Advanced Deep Learning Approaches for Automated Recognition of Cuneiform Symbols (ID 2505.04678), details how the team trained modern image-recognition software to spot the wedge-shaped impressions that record the world’s earliest written laws.
Unlocking clay archives
Cuneiform experts are few, and manually copying signs from tablets the size of a hand can take hours. By feeding the computer 14,100 cleaned images of 235 different signs, the team taught it to recognise nearly every mark on a test tablet that carries the first clause of Hammurabi’s Code—written around 1754 BCE.
How well does it work?
On held-back images the best network, an EfficientNet variant, misread just one sign in ten thousand. When faced with the real tablet, it got roughly two characters wrong in a hundred; a second model trailed behind at 89 percent.
Why archaeologists care
Fast, reliable transcription could let museums digitise extensive collections from Mesopotamia, Syria and Anatolia, revealing legal, economic and literary texts that have been effectively silent since antiquity. The authors add that comparing how individual signs change from city to city or century to century may sharpen dating and provenance studies.
Next steps
The Dubai group plans to blend several networks for even higher accuracy and to fine-tune the image-cleaning steps needed when tablets are chipped or burned. They also propose adapting the method to other scripts—such as Egyptian hieroglyphs—once similarly large image sets are available.
The preparation of this article relied on a news-analysis system.