Despite advances in machine learning, high-resolution imaging and large-scale data analysis, several ancient writing systems remain unread. A Deutsche Welle report published on December 25 profiles seven of them, including the Indus Valley script, Linear A from Bronze Age Crete and Rongorongo from Easter Island, and explains why decades of scholarship and newer computational tools have yet to produce a decipherment.
The report draws on Svenja Bonmann, a specialist in historical-comparative linguistics at the University of Cologne in Germany, whose research involves attempting to decipher unknown scripts. According to the report, these texts are often the only surviving window into societies that disappeared centuries ago, which is what keeps researchers working on puzzles that have defeated generations before them.
Where pattern recognition stops
AI systems are good at one part of the problem: sorting signs, counting how often they appear and spotting recurring sequences across a corpus. That kind of statistical analysis has helped scholars organize inscriptions and test hypotheses about sign order and frequency. The limits appear when the corpus is small.
The Indus script is the clearest case. It survives mainly as short strings of signs on seals and pottery, often only a handful of characters long. With so little text, researchers cannot even settle whether the signs encode a full spoken language or function as a symbolic system, such as marks of ownership. Software can arrange the data, but assigning meaning still depends on human interpretation and on evidence the inscriptions themselves do not provide.
No Rosetta Stone
The second obstacle is the absence of a bilingual text. Egyptian hieroglyphs were cracked because the Rosetta Stone carried the same decree in a known language, and the Elamite script from Iran was recently read with the help of comparable parallel texts, as Deutsche Welle has reported separately. For Linear A, Rongorongo and the Indus script, no such key exists. Without an anchor that ties signs to sounds or words, a model has nothing to check its guesses against, and plausible-looking decipherments cannot be confirmed or refuted.
That is why the field treats claims of a complete translation with caution. Several have circulated online in recent years, but none has been accepted by specialists, because they cannot be tested against independent evidence.
Older mysteries, same constraint
The report sets the scripts alongside other unresolved historical questions, from the 1908 Tunguska explosion in Siberia, which flattened trees across a vast area without leaving a confirmed impact crater, to the still-unlocated tombs of Alexander the Great and Cleopatra VII. In each case, modern techniques have added detail without supplying the one missing piece of evidence.
The lesson for AI in the humanities is practical rather than discouraging. Computational methods can speed up cataloging, suggest structure and rule out some hypotheses. Turning that into a reading of a dead language still requires new material from the ground: a longer inscription, or a parallel text in a language someone can already read.
