Deciphering Linear A
The Minoan script of Crete (about 1800-1450 BC) can be sounded out using Linear B values, but its language is unknown and no text can be read.
Open in the interactive tree →Many Linear A signs resemble Linear B, which writes Greek, so they can be pronounced, but the resulting words are meaningless in any known language. No bilingual text exists, and the surviving tablets are short lists of names, goods and numbers. The grammar suggests the language is not a close relative of any known one.
As of October 2026
A supplement published in 2024 (RILA-S1) raised the corpus to 1,534 inscriptions and 7,574 signs, against about 6,000 inscriptions for Linear B. In June 2026 an engineer publicly claimed a Semitic reading found with AI-written analysis scripts; no peer-reviewed paper exists, and a Semitic link (Cyrus Gordon, 1957) had been rejected before. Neural methods have matched 67.3% of Linear B cognates, but only because Greek was a known relative.
What is missing
- More and longer inscriptions, ideally a bilingual text
- An independent check of sign sounds that does not borrow from Linear B
- A candidate language family with a test that can reject it
- Agreed statistical controls so that claimed readings can be scored against chance
Becomes possible once solved
- Knowing what language the Minoans spoke
- Reading palace and town records from across Crete
- Comparing Minoan with Mycenaean administration
Open steps
- More and longer inscriptions Low AI leverageThe corpus is 1,534 inscriptions with 7,574 signs, against about 6,000 inscriptions for Linear B; a longer text or a bilingual one would change what can be tested.
- Check sound values independently Medium AI leverageSound values come from similar Linear B signs; they are widely accepted but not incontrovertible. Find tests that do not reason in a circle.
- Which language family? Medium AI leverageTest candidate relatives by neural cognate matching. It matched 67.3% of Linear B cognates because Greek was known; for Linear A no relative is known.
- Tests that can reject a reading Medium AI leverageAgree shared random-control tests so claims such as the June 2026 Semitic reading can be scored against chance on the same 7,574 signs.
Where AI could help
Medium AI leverage. Computers can test thousands of sign-value and language hypotheses quickly, but with about 7,500 signs and no bilingual any fit is weak evidence.
- Search sign-value assignments against the whole corpus
- Test candidate languages with neural cognate matching
- Build random-control tests to see how easily chance gives a 'fit'
- Read worn signs from photographs and 3D scans
Shown so far
- A 2019 ACL paper deciphered Ugaritic and matched 67.3% of Linear B cognates with a minimum-cost-flow neural model; both scripts had a known related language. source