The undeciphered Indus script of the Bronze Age Indus Valley Civilization has been cracked, revealing the lost language of Harappa

The case that it can be read

The optimistic case is not foolish, and its best version deserves a fair hearing. The strongest camp holds that the language behind the script is Dravidian, an argument developed over decades by scholars including Asko Parpola using careful comparative work and early computer analysis.

Its showpiece is the reading of the common fish sign. In Dravidian languages the word for fish is min, which is also the word for star. If the Indus scribes used the fish sign for its sound rather than its picture, a technique well attested in other early scripts, then fish-plus-modifier signs could spell the names of stars or planets, a reading that fits a culture plausibly concerned with the calendar and the sky. It is an elegant idea, and it is the kind of internal coherence that a real decipherment would show.

Supporters add that the signs behave statistically like a language. The 2009 study led by Rajesh Rao in Science found that the script's conditional entropy, a measure of how predictable each sign is given the one before it, sat closer to natural languages than to rigid or random nonlinguistic systems. On this view the material really is writing, the Dravidian key is promising, and decipherment is a matter of time and effort rather than a lost cause.

A large corpus, a recurring sign with a beautiful bilingual pun built into a living language, and statistics that look linguistic. If the Indus script is readable at all, this is roughly what the road to reading it would look like.

That is the case at full strength: not that the script has been translated, but that there is a serious, sober line of evidence suggesting it is a real language of a known family, and that patient work along these lines might eventually succeed.

Why no decipherment holds up

Here is the pivot. Everything above supports one modest word: promising. The rated claim needs a much stronger word: solved. Nothing in the record crosses that distance.

The decisive obstacle is that there is no bilingual text. Egyptian hieroglyphs fell to Champollion because the Rosetta Stone carried the same decree in a known language alongside them. The Indus corpus has no such key, and no long passages either. Inscriptions run about five signs on average, and the longest, the Dholavira signboard, has only around ten. With no grammar exposed by lengthy texts and no external anchor, a proposed reading cannot be checked. The proto-Dravidian solutions work by assuming the language first and then fitting signs to it, which is exactly the move that cannot validate itself.

The result is a field full of mutually incompatible solutions. The script has been read as Dravidian, as Indo-Aryan, as Sanskrit, and as other tongues, each by a confident author, each incompatible with the rest. A genuine decipherment does the opposite: it converges. When Michael Ventris cracked Linear B, independent readers could apply the values and arrive at the same words. Nothing like that reproducibility has happened here. Coexisting confident answers are the signature of an unsolved problem, not a solved one.

As for the computers, they have delivered real description and no translation. Modeling which sign follows which, or estimating the corpus's entropy, characterizes the patterns; it does not assign verifiable meanings when the language is unknown and no bilingual exists. Reports that artificial intelligence has read the script consistently claim more than the underlying work supports. Better tools have sharpened the puzzle; they have not finished it.

The deeper doubt: is it even writing?

There is a further problem that most decipherment claims quietly assume away. It is not settled that the Indus signs record spoken language at all.

In 2004, Steve Farmer, Richard Sproat, and Michael Witzel argued that the system is nonlinguistic. Their case rests on features that sit oddly with true writing: the inscriptions are extraordinarily short, a striking number of signs appear only once, and there are no long texts anywhere, across a civilization that lasted centuries and traded over long distances. Writing systems used to record speech, they note, tend to leave behind at least some lengthy inscriptions. The Indus system leaves none. On their reading, the signs may have functioned like heraldry or religious and economic emblems, marking people, goods, or deities without spelling out words.

This is not a settled defeat for the linguistic view. The 2009 entropy paper was a direct response, and its authors argue the statistics favor language; skeptics such as Sproat reply that the entropy measure does not cleanly separate language from non-language when proper controls are applied. The dispute continues. But notice what its mere existence does to the decipherment claims. Every proposed reading takes for granted that there is a language to read. If the nonlinguistic thesis is right, no key will ever fit, because there is no lock.

You cannot translate a text into a language until you have shown it is a text in a language. That first step, obvious as it sounds, has never been securely taken for the Indus script.

The honest position is symmetrical. The nonlinguistic camp has not proven the signs are meaningless as speech, and the linguistic camp has not proven they encode it. That basic uncertainty, unresolved after a century, is itself a reason to treat any confident decipherment with caution.

Why the script keeps getting cracked

New decipherments of the Indus script arrive with almost seasonal regularity, and they keep coming for reasons that have less to do with the seals than with us.

They ride the romance of a silent civilization. The Indus built great cities and then went quiet, and the promise of suddenly hearing its voice is intoxicating. The reading of Egyptian and Maya scripts supplies a heroic template, the lone genius with the key, that makes solving the Indus feel like a prize waiting to be claimed, most concretely in the one-million-dollar reward announced in 2025.

They carry a political charge. Whether the language was Dravidian or Indo-Aryan touches deep and sensitive questions about the origins of South Asian peoples, so a decipherment is never only academic; it can seem to settle who was there first. That gives advocates on each side a powerful motive to believe a reading that favors their story, and to present a hypothesis as a proof.

And they exploit a gap that is easy to miss. Between describing the sign patterns, which computers do well, and reading them, which no one has done, lies a chasm that technical language papers over. A statistical result about entropy or sign order is real work, but reported through the frame of “cracking the code” it becomes, in the retelling, a solution. The boldest claim is always the most satisfying, and “we still do not know” makes for a poor headline.