Indus Script Deciphered as Early Dravidian, with AI Assistance

The Indus script, written on seals, tablets and pots between about 2600 and 1900 BCE, has not previously been deciphered. We read it as an early Dravidian language, using Tamil, its best-preserved descendant, as our guide. We work in three stages. First, before assuming any language, we characterize the script from its own patterns. It is a logosyllabic (word-and-syllable) script, like early Sumerian, not an alphabet or a Brahmi-style abugida script. Its seal texts follow one fixed frame (opener, open middle, count, closer), used the same way in every city throughout the mature Indus period. Second, we ask which language fits. We ran a simulation identically for each language, with blind picture words and held-back seals. Early Dravidian, read through Old Tamil, fits the seals far better than chance, and far better than Sanskrit or Prakrit. Third, we judge each proposed sign value on three independent kinds of evidence: statistical fit, picture match and meaning. 100 sign values are now deciphered, graded by confidence. Together they account for 77% of all legible sign uses in the corpus. 756 inscriptions are fully read, meaning every sign in them has a value. That is 35% of the 2,165 distinct complete texts (and 46% of the 3,603 complete inscribed faces, counting duplicates). Data files: the readings of all 756 fully read inscribed objects (Indus Seal Readings), the sign readings with confidence levels (Indus Sign Readings), the full test register (300 tests), corpus statistics and their scripts, and 29 pre-registered test plans with SHA-256 hashes. See README.md. Use of AI: Claude, an AI model made by Anthropic, ran the analyses, wrote the scripts, drafted the test plans and much of the text, working with Siva Raj, who directed the research and made the decisions. The "blind judges" in the picture tests were separate Claude sessions.