The Indus script

Three things are secure. The signs form a bounded set with stable shapes across sites and centuries, which means the system was learned and transmitted⁠[1]. The texts have positional structure: n-gram models trained on the corpus find that sign order is far from random, that some signs are confined to the start or the end, and that the structure is strong enough to predict missing signs in damaged texts⁠[5]⁠[2]. And the texts are short: with no text over about seventeen signs on one surface, the corpus has far less connected material than any deciphered script had when it was read⁠[4].

What is contested is whether this structure is the structure of language. Rao and colleagues showed in 2009 that the conditional entropy of sign pairs in the Indus corpus lies in the range occupied by natural-language scripts and away from two artificial extremes, a rigidly ordered and a random sequence, and argued that this raised the probability that the signs encode language⁠[7]. Sproat replied that the two extremes were invented controls, that real non-linguistic systems such as Vinča signs, Mesopotamian kudurru emblems and heraldry can show similar entropies, and that the measure cannot discriminate; Rao and colleagues answered that they had never claimed entropy alone was decisive and that the comparison systems Sproat named were poorly documented⁠[8]⁠[13]⁠[14]. Sproat's 2014 study in Language tested a wider set of statistics on linguistic and non-linguistic corpora and concluded that such measures do not settle whether an unknown system is writing⁠[9].

Decipherments have been tested against four requirements: a published sign list and values, readings that apply across the corpus rather than to chosen texts, consistency with the known direction and positional structure, and some external check. The Dravidian hypothesis has a method, rebus reading anchored in Dravidian words and in later Indian iconography, and a small number of proposed readings, but no reading of a whole text that can be verified⁠[1]. Indo-Aryan and Sanskrit readings recur, most recently in 2024 and 2025 claims circulated online and dismissed on the Indology list as iterations of earlier ones⁠[6]. Farmer, Sproat and Witzel offered a US$10,000 prize in 2004 for the discovery of an Indus inscription of at least fifty signs, on the argument that a literate civilisation leaves long texts; none has been found⁠[4].

The non-script argument has itself been tested statistically, with the inconclusive result described above⁠[7]⁠[8]⁠[9]. Yadav and colleagues' n-gram analysis found syntactic regularities of the kind writing shows, while noting that the same regularities could arise in a structured non-linguistic system⁠[5]. The Tamil Nadu State Department of Archaeology, working with the Roja Muthiah Research Library, has catalogued more than 15,000 graffiti-bearing potsherds from Tamil Nadu sites and compared the marks with Indus signs; the Chief Minister's January 2025 keynote put the overlap at about 60 per cent of Indus signs, a figure that measures sign-shape resemblance, not reading⁠[11]⁠[10].

Four positions are live. First, the signs write an early Dravidian language, the view of Parpola, Mahadevan and most who think it is writing; its evidence is the rebus readings, the Dravidian substrate in later Indic languages and the survival of Brahui in Baluchistan⁠[1]. Second, the signs write an Indo-Aryan language, which would require Indo-Aryan speakers in the Indus valley before 1900 BC and has no reading accepted by specialists⁠[6]. Third, the signs are not writing: a system of emblems, clan or office marks, or religious symbols with structure but no phonetic content, which explains the brevity, the absence of long texts and the lack of writing on perishable media in a civilisation with extensive trade⁠[4]. Fourth, the signs are a logographic or logo-syllabic system that encodes names, titles and commodities in a language that cannot be identified from such short texts, a position compatible with the positional statistics and with the failure of every reading⁠[5]⁠[9].

A bilingual, which Indus trade with Mesopotamia makes possible in principle, or a long text of fifty signs or more, would settle whether the signs are writing and give a decipherment something to work on⁠[4]⁠[1]. Short of that, a decipherment must publish its sign values and read the corpus, not a sample, with results that can be checked against the positional statistics and the iconography⁠[1]. The 2025 prize has no published judging criteria; a credible award would need an expert panel and these tests⁠[10]⁠[15]. On the non-script question, the useful work is better comparison corpora of real non-linguistic systems, which is what Sproat's 2014 study began⁠[9].