A survey of published research, from the Essentia Memoria newsroom. It lists real papers, each looked up at its DOI registry; the game has not repeated their results.
Research survey
What AI research has found: ancient inscriptions and lost languages
Happened: · Published:
3 published works on ancient inscriptions and lost languages, each looked up at its DOI registry before it was listed here, with what each one reports in plain words.
Watch
Explanatory visualisation, drawn from the record's numbers. It is not footage of the event: nobody filmed it.
Listen
Transcript
News from the Essentia Memoria game universe. What AI research has found: ancient inscriptions and lost languages. 3 published works on ancient inscriptions and lost languages, each looked up at its DOI registry before it was listed here, with what each one reports in plain words. This survey lists what each paper reports. The game has not repeated these results. A search finds papers that match its words, so a listed paper may use AI only in part. The search is limited to the fifty most-cited matches, so the newest results may be missing until they are cited. This report was written from the world record by a fixed template. The full story and its source records are on essentiamemoria.com.
The question
Can a model fill the gaps of a broken inscription, or read a language nobody has spoken for thousands of years?
Claims, and where each one comes from
- Where this research lives in the game (a place the world grammar builds)museum → museum
Game simulation · Set in code
src/lib/world-grammar/place-grammar.ts · resolvePurpose('museum') - Published works in this survey, each resolved at its DOI registry3
Real-world knowledge · Source [1] [2] [3]
- A model called Ithaca restores missing letters of Greek inscriptions and suggests where and when each was carved.Assael 2022
Real-world knowledge · Source [1]
- A model matched words of Ugaritic and Linear B to their known relatives, the way a decipherer works.Luo 2019
Real-world knowledge · Source [2]
- A translation model turns Akkadian cuneiform into English as a first draft that scholars then correct.Gutherz 2023
Real-world knowledge · Source [3]
References, checked at their registry
[1] Assael (2022). Restoring and attributing ancient texts using deep neural networks. Nature. doi:10.1038/s41586-022-04448-z
Verified: the work exists and matches · checked at Crossref on Oct 28, 2026, 11:06 PM UTC · the registry holds “Restoring and attributing ancient texts using deep neural networks” (2022)
[2] Luo (2019). Neural Decipherment via Minimum-Cost Flow: From Ugaritic to Linear B. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. doi:10.18653/v1/P19-1303
Verified: the work exists and matches · checked at Crossref on Oct 28, 2026, 11:06 PM UTC · the registry holds “Neural Decipherment via Minimum-Cost Flow: From Ugaritic to Linear B” (2019)
[3] Gutherz (2023). Translating Akkadian to English with neural machine translation. PNAS Nexus. doi:10.1093/pnasnexus/pgad096
Verified: the work exists and matches · checked at Crossref on Oct 28, 2026, 11:06 PM UTC · the registry holds “Translating Akkadian to English with neural machine translation” (2023)
Review
Not peer reviewed. No human scientist has reviewed this survey. Its template and landmark summaries were written with an AI assistant (Claude); the papers it lists are real and each was checked at its registry, so anyone can follow the links and read them.
- Each listed DOI was resolved, and its title, year and first author matched the listing.
- Each work the search found is held by Crossref, a registry independent of the search.
- Each number in the headline and summary is accounted for by a listed claim.
- No candidate work was dropped: every one the desk found or named was confirmed by its registry.
Code version: deb216f03447
What the record does not say
This survey lists what each paper reports; the game has not repeated these results. A search finds papers that match its words, so a listed paper may use AI only in part. The search is limited to the fifty most-cited matches, so the newest results may be missing until they are cited.
How this was written
Landmark works named by the desk, then the most-cited works with a DOI that OpenAlex returns when titles and abstracts are searched for (inscription OR epigraphy OR cuneiform OR decipherment OR "ancient language") AND ("machine learning" OR "deep learning" OR "neural network" OR "neural networks" OR "artificial intelligence" OR convolutional OR "reinforcement learning" OR "language model" OR transformer); a hit is kept only if its own title names the field and its title or abstract names an AI method, as whole words, and the hits that name the most of these words in their title come first. Each candidate was looked up at Crossref (OpenAlex as second resolver) and its title, year and first author compared; a work the search found is shown only when Crossref holds it. Landmark summaries were written by the desk; the other works are listed by their own titles.
Source records: code:src/lib/world-grammar/place-grammar.ts#resolvePurpose · doi:10.1038/s41586-022-04448-z · doi:10.18653/v1/P19-1303 · doi:10.1093/pnasnexus/pgad096