How AI tests a proposed reading of a lost language

A machine can recognise the marks on an ancient tablet without knowing what its writer meant. It can also produce a plausible interpretation without establishing that the interpretation is right. Computational decipherment becomes interesting at the point between those achievements: can a proposed reading explain enough recurring evidence to survive a serious test?[1]

Reading marks and understanding language require different evidence

An influential 2010 statistical model approached the problem through related languages. It searched for recurring correspondences between Ugaritic and Hebrew, considering both individual letters and meaningful word parts such as stems and endings. The method relied on important preparation: the signs were electronically transcribed, words were separated, and a morphologically analysed Hebrew lexicon was supplied. The computer’s search began with that foundation.[2]

The useful idea is consistency across a collection. A proposed letter value affects every word containing it; a recurring ending can constrain several words simultaneously. A candidate reading that fits one inscription but requires different sound values elsewhere becomes less convincing. The 2010 model encoded some of those constraints statistically. It still operated at vocabulary level, and its authors explicitly identified the missing literary context as a limitation.[2]

What a celebrated neural result actually measured

In 2019, NeuroCipher combined a neural model of character correspondences with a global word-matching procedure. The latter favoured a coherent assignment across the vocabulary. Researchers tested it on Linear B, the script that records Mycenaean Greek, using already known word correspondences for evaluation. The often-repeated 67.3% result belongs to a particular names-matching condition.[3]

These experiments test whether a procedure can recover relationships without being given the pairings. But the setting supplies something an undeciphered corpus may lack: a productive comparison language. Recognisable relatives let one sound correspondence support another. Where that network is absent, searching faster cannot supply the missing linguistic evidence.[2]

The spaces between words are another unknown

A 2021 TACL paper tackled texts whose word boundaries were missing or incomplete. Its model jointly proposed segments and searched for related words, using phonetic features to favour plausible sound correspondences. Gothic and Ugaritic provided cases where researchers could check results against established readings. The experiments also varied how much sound and segmentation information was supplied, demonstrating why a score depends on what is known before the search starts.[4]

The researchers also examined Iberian, whose interpretation remains unresolved in that study. They could investigate personal names already supported by other evidence and compare candidate language relationships. Their method found no strong evidence for Basque as a close relative. That outcome was a limited test of a hypothesis: it neither deciphered Iberian nor proved that every proposed relationship was impossible.[4]

More realistic search includes uncertainty in the evidence

Fabio Tamburini’s 2025 research investigated a different computational approach, using constrained optimisation rather than a neural text generator. Its search could accommodate uncertain or damaged symbols, unequal sign mappings and partial knowledge. The objective was to find combinations of assignments that fit the available material while respecting the constraints, instead of requiring every inscription to arrive in an idealised form.[5]

A revealing part of the evaluation concerned information supplied to the algorithm. Tamburini reran the earlier neural baseline without giving it the expected number of cognate matches, because that count may be unknown in a genuine decipherment. The paper’s benchmarks still use known correspondences for checking. The methodological lesson is valuable: assess a system against the information a researcher would actually possess, including what remains uncertain.[5]

A reading becomes stronger when it predicts new evidence

The human decipherment of Linear B offers a striking example. After Michael Ventris and John Chadwick proposed their Greek reading, archaeologist Carl Blegen applied its sign values to a newly cleaned Pylos tablet in 1953. Words corresponding to tripods appeared beside drawings of three-legged vessels. Descriptions of vessels with different numbers of handles matched the accompanying signs. The proposed reading worked on material that had not supplied its original evidence.[6]

That is a powerful target for AI-assisted decipherment. A useful system can rank candidate sound values, expose repeated structures and show where a proposal breaks down. A convincing historical claim then needs the interpretation to hold together beyond the examples that suggested it. The prospect is a more systematic way to test readings—and, when the evidence supports them, to recover something people once chose to record.[6]

Report an error or suggest a correction