Linear A still has no Rosetta Stone. AI isn't going to build one.

A self-taught engineer and amateur linguist reported in June 2026 that he had assigned values to 40 Linear A signs and compiled a 408-word lexicon, working from a single guess about one word in a prayer inscription and scaling it with AI-built pattern-matching scripts. Linear A, the Bronze Age Minoan script from Crete, has no Rosetta Stone and no known relative language (Ars Technica).

Linear A has resisted decipherment for a century. Every other script ever read had a key outside itself: a bilingual text, or a known relative language. Champollion read Egyptian in 1822 because the Rosetta Stone carried the same decree in hieroglyphs, demotic, and Greek. Linear A has no bilingual inscription and no confirmed relative language. One useful parallel is Etruscan, a pre-Roman language of Italy known from short funerary inscriptions. Linguists have built a partial vocabulary from those texts, but the grammar and deeper meaning remain unresolved (The Conversation).

The June 2026 claim began with a hunch, not a discovery. The engineer guessed that one unknown word in a prayer inscription derived from a Semitic root meaning "to dwell" or "to inhabit." He then wrote programming scripts that propagated that single sound-pattern guess across a collected Linear A corpus, looking for clusters of signs that might match other Semitic roots and assembling those matches into a working vocabulary. The reported result was a tentative 40-sign table and a 408-word lexicon, plus a hypothesis that Linear A is an extinct member of the Semitic family that includes Hebrew, Aramaic, and Arabic (Ars Technica).

The hypothesis is the engineer's, not the field's. The claim has not been peer-reviewed, and Ars Technica frames it with hedging language like "reportedly" and "in his view" (Ars Technica). No major linguistics society has endorsed it. Prior work in the space, including a 2020 MIT News feature on machine learning for lost languages, treats AI in this field as pattern-propagation tooling under human linguistic priors. The engineer describes his own process in similar terms on a self-published site: the human had the idea, the AI did the testing (aiclambake.com).

A finite corpus, thousands of signs, and a single seed hypothesis can be searched with scripts at scale, and that is the part the method handles well. What the corpus cannot produce is the seed itself. Every decipherment in the historical record began with an anchor from outside the script: a bilingual inscription like the Rosetta Stone, a borrowed place name, or a structural hint from a related tongue. Linear A has no equivalent. The AI in this case is grading the engineer's guess, not the script itself.

The Semitic-link hypothesis, if correct, would make Linear A the oldest attested member of the Semitic family by a wide margin, since Hebrew, Aramaic, Arabic, and Akkadian are separated by thousands of years and major structural shifts. Linguists testing the claim would look for Semitic fingerprints in the corpus: triconsonantal roots shared across word families, non-concatenative morphology where vowels slot into consonantal skeletons, and grammatical features like verb-initial word order. None of that has been independently verified. The claim is either very important or very wrong, and the burden of proof sits with the claimant.

A Linear A tablet with parallel text in a known tongue would be the field's Rosetta Stone, and the discovery would shrink the entire category of AI-decipherment claims. Until one turns up, 408 words and 40 signs remain a working hypothesis, not a decipherment.