Deciphering the Indus Script
Thousands of short inscriptions on seals from about 2800-1900 BC remain unread, and experts disagree even on whether the signs record a language.
Open in the interactive tree →The Indus civilisation left seals, pottery marks and tablets with 400 or more signs, and the average text is about five signs. No bilingual text exists, and over 100 mutually exclusive decipherments have been published since the 1920s; a 2004 paper argued the signs are non-linguistic, while others read a Dravidian language. The Easter Island script rongorongo (26 texts, over 15,000 glyphs) is another unread script with similar limits.
As of October 2026
In 2025 Tamil Nadu offered a US$1 million prize for a reading, and no reading has gained scholarly acceptance. A July 2026 conference paper on 6,579 inscriptions finds order in sign sequences that random generation does not easily explain, which says nothing about meaning. Sign lists range from 419 (Mahadevan, 1977) to about 694 (Wells, 2015). A 2009 entropy argument for language was challenged by Sproat in 2014 because the test also passed non-linguistic systems.
What is missing
- A bilingual or much longer inscription
- A settled sign list that scholars share
- Proof that the signs record a language rather than names or symbols
- Statistical tests that tell language apart from non-linguistic sign systems
- More excavated texts from more sites
Becomes possible once solved
- Knowing the language of the Indus cities
- Reading seals as names, titles or trade marks
- Linking the civilisation to later South Asian languages
Open steps
- Does it record a language? Medium AI leverageSettle whether the signs write a language or a non-linguistic system, as a 2004 paper argued; a test must work on texts of about five signs.
- A settled sign list Medium AI leverageCounts range from 419 (Mahadevan, 1977) to about 694 (Wells, 2015); classify signs and variants from seal images in a consistent, shared way.
- Longer or bilingual texts Low AI leverageNo bilingual text exists and the average inscription has about five signs; new finds or paired texts abroad (Indus seals turn up at Ur and Susa) would give a test.
- Rank language hypotheses fairly Medium AI leverageCompare Dravidian, Indo-Aryan, Munda and others by predictions made before testing; over 100 mutually exclusive decipherments have been published.
- Prove methods on solved scripts Medium AI leverageCheck any method on solved scripts (Linear B, Ugaritic) before using it on the Indus signs or on rongorongo (26 texts, over 15,000 glyphs).
Where AI could help
Medium AI leverage. AI can cluster signs and model sign order, but short texts and no bilingual keep the statistics from saying anything about meaning.
- Cluster sign shapes and variants from seal photographs
- Model sign order and compare it with solved scripts under shared controls
- Score rival language hypotheses on every inscription with identical rules
Shown so far
- A July 2026 NLP4DH paper used visual clustering, entropy analysis and a BiLSTM on 6,579 inscriptions and reports structure not easily explained by random generation. source
- A 2009 entropy result for language was challenged by Sproat, who argued the model also fits Mesopotamian deity symbols and so cannot tell language from non-language. source
Prerequisites
- Writing~3200 BC
- Cuneiform Deciphered1857
- Deep Learning2012