ИИ берётся за расшифровку мёртвых языков: находит паттерны, смысл — за учёным

Neural networks are already helping historians read texts that have remained incomprehensible for millennia. Yet for all the power of algorithms, decipherment still requires a human: AI is exceptionally good at finding patterns in symbols, but giving them meaning is something only a linguist can do.

What neural networks can do in linguistics

The main advantage of neural networks over a researcher is speed and scale of statistical analysis. In a matter of hours, an algorithm processes a corpus that would take a scholar years: counting symbol frequencies, building co-occurrence matrices, identifying structural patterns.

One of the most telling examples is the Ithaca project, developed by a DeepMind team together with historians and published in the journal Nature in 2022. The system restored damaged fragments of ancient Greek inscriptions. The algorithm's accuracy was approximately 62%, whereas specialists without AI guessed the correct variant in only 25% of cases. When historians worked in tandem with the neural network, the result rose to 72%.

Ithaca worked effectively precisely because a large corpus of deciphered texts already existed for ancient Greek. The algorithm relied on thousands of inscriptions previously studied by scholars — and built predictions for unknown fragments on that foundation.

• Ithaca (DeepMind, Nature, 2022) — 62% accuracy in restoring inscriptions
• Historians without AI — approximately 25% correct variants
• Historians + Ithaca — 72%
• The largest undeciphered writing systems: Linear A (Crete), Proto-Elamite (Iran), Rongorongo (Easter Island), Indus script (Mohenjo-daro)

Why a linguist is indispensable

AI can describe the structure of a text, but cannot connect it to meaning. This is especially apparent where there is no "anchor" — a parallel text like the Rosetta Stone for Egyptian hieroglyphics. For Linear A, Proto-Elamite, or Rongorongo, no such entry point exists, and without one decipherment turns into guesswork.

An algorithm will detect that certain symbols occur together more frequently than others — but labeling them as words without cultural context is impossible. A linguist brings what is absent from the data: knowledge of language typology, an understanding of how trade records differ from religious texts, intuition about grammatical categories — all of that which cannot be directly learned from a corpus.

"AI is fantastically good at pattern recognition, but human intuition remains key" — this is exactly what researchers in computational linguistics interviewed by

Ars Technica emphasize.

How scientists' work is changing

The pairing of "neural network + linguist" does not replace the scholar — it changes the pace of their work. The algorithm takes over the routine: frequency analysis, restoration of missing symbols, searching for structural matches in parallel corpora. The specialist concentrates on interpretation — the stage where understanding of the world is needed, not just of the text.

A number of researchers compare this shift to how MRI changed diagnostics: the doctor still makes the diagnosis, but sees details previously inaccessible. Teams that formerly spent months on manual corpus encoding now receive initial patterns within days. Several dozen undeciphered writing systems remain in the world — some of them, according to specialists, may be read in the coming years precisely thanks to this partnership.

What this means

The application of AI in linguistics is an example of how neural networks augment expertise rather than displace it. Where parallel texts and sufficient data exist, the algorithm acts as an accelerator. Where cultural context is opaque or data is scarce, the final word belongs to the scholar.