The unread cuneiform: half a million tablets and the limits of Sumerian

The corpus is every clay tablet, prism, brick, seal and stone inscribed with wedge-shaped signs between about 3300 BC and the first century AD. No official count exists. CDLI estimates that public and private collections hold more than 500,000 exemplars, and the 2026 York programme puts the figure above 550,000; CDLI has catalogued more than 400,000 and holds about two million lines of transliteration⁠[1]⁠[4]. The British Museum holds about 130,000 tablets and the Yale Babylonian Collection about 40,000, followed by the Penn Museum, the Vorderasiatisches Museum in Berlin, the Louvre, Istanbul and the Iraq Museum⁠[5]. The largest single block is administrative. The Ur III state (about 2100 to 2000 BC) left at least 120,000 tablets, of which the BDTNS database had catalogued 105,187 and transliterated 75,818 by May 2026⁠[6]⁠[7]. The literary corpus is small by comparison. ETCSL edited about 400 Sumerian compositions, roughly 50,000 lines, most of them known from Old Babylonian school copies of about 1800 BC, before its funding ended in 2006⁠[21]. ePSD2, the lemmatised Sumerian corpus behind the Pennsylvania dictionary, now draws on more than 110,000 manuscripts and 3.4 million word instances⁠[3]. For Akkadian literature and scholarship of the first millennium BC, the eBL Fragmentarium in Munich held about 27,000 transliterated fragments by 2024, most from the British Museum's Kuyunjik (Nineveh) and Babylon collections⁠[12]⁠[11]. The oldest literary tablets are about 500 Early Dynastic IIIa pieces from Abu Salabikh, excavated in 1963 and 1965, and a comparable group from Fara (ancient Shuruppak) dug in 1902 and 1903, now in Baghdad, Berlin and Philadelphia⁠[22]⁠[23]. Transliterations on CDLI may be reused with attribution and ORACC data is CC BY-SA; the photographs remain the museums' property, which limits what image-based work can be published⁠[24]⁠[2].

Cuneiform is read. Akkadian, a Semitic language with living relatives, was deciphered in the 1850s, and Sumerian was recovered through the bilingual word lists and translations that Babylonian scribes themselves compiled; a specialist reads an ordinary account or royal inscription at sight. The uncertainty is at the edges. Sumerian is a language isolate. Proposals to connect it with Indo-European, Dravidian, Uralic, Turkic, Austroasiatic, Sino-Tibetan and various macrofamilies have all been rejected, and Whittaker's suggestion of an early Indo-European language ("Euphratic") behind the Uruk texts has no following⁠[13]⁠[15]. Landsberger's pre-Sumerian substratum, meant to explain polysyllabic words and old city names, was taken apart item by item by Rubio in 1999⁠[14]. Sumerian phonology is reconstructed from Akkadian transcriptions and spelling variation, so the consonant transcribed ř, the velar nasal, the number of vowels and the existence of /h/ and a glottal stop are all disputed, and stress is unknown⁠[13]⁠[16]. The verbal prefix chain is analysed differently in the main grammars of Thomsen (1984), Edzard (2003) and Jagersma (2010)⁠[13]. Emesal, the dialect of laments and of goddesses' speech, was spelt out syllable by syllable because ordinary word signs would be read in the main dialect by default, so it is well attested but its relation to the main dialect is still argued⁠[13]⁠[25]. Deimel counted about 870 signs for Early Dynastic IIIa⁠[13]. The lexical lists, the scribes' own dictionaries, run from the archaic Uruk lists to first-millennium commentaries; the profession list Lú A was copied almost unchanged for a thousand years, and at least one archaic list, AD-GI4, may not be Sumerian at all⁠[26]. Whether proto-cuneiform encodes a language is unsettled⁠[13].

Two kinds of test have been run: on the hardest texts, and on the tools meant to speed up reading of the rest. On the Early Dynastic literary tablets, Krecher showed in 1978 that the Fara and Abu Salabikh compositions use a regular system of sign substitutions, UD for AN, GAL for EN, NUN for LIL and others, which has to be undone before the text can be read⁠[8]. Alster asked in 1982 whether the system hid an early form of Emesal; the question is still open⁠[25]. Biggs showed in 1971 that the Abu Salabikh Kesh Temple Hymn matches the Old Babylonian text eight centuries later except in spelling, which is why that one Early Dynastic composition can be read⁠[27]. Krebernik and Lisman's 2020 edition of the 70 Zame hymns, which have no later parallel, calls its translation not definitive because the orthography omits suffixes, postpositions and verbal morphemes and does not keep the signs in spoken order⁠[9]⁠[23]. Cooper's judgement is that the UD.GAL.NUN texts remain very hard to understand and probably went out of use before Ur III⁠[10].

On the tools, Fetaya and colleagues trained recurrent networks on Achaemenid Babylonian archival texts in 2020 to propose fillings for breaks, and Lazar and colleagues did the same with masked language models in 2021⁠[20]⁠[28]. Punia and colleagues built the first Sumerian machine translation, for Ur III administrative texts, in 2020⁠[29]. Gutherz and colleagues trained neural translation on 56,160 Akkadian sentences from five ORACC corpora and reached BLEU 37.47 from transliteration and 36.52 from cuneiform signs, against a translation-memory baseline of 23.51 and 27.09; they report hallucinated output and poorer results on long or logogram-heavy lines⁠[17]⁠[30]. DeepScribe, trained on 100,000 annotated signs from the Persepolis Fortification tablets, located signs with mAP 0.78 and classified them with top-five accuracy 0.89, falling to 0.80 end to end⁠[19]. The eBL Fragmentarium's matching, a bag of character n-grams scored by the overlap coefficient, surfaced three unidentified manuscripts of the Maqlû incantations when an expert checked the top 300 of some 25,000 candidates, and the project reports about 1,500 joins in five years⁠[12]⁠[18]. In 2026 a sign detector was run over 87,668 eBL fragments, giving 2.9 million detections, with the authors noting its sensitivity to damage⁠[31]. Heidelberg's HeiCuBeDa provides 3D scans and six-view images of Hilprecht Collection tablets as a benchmark⁠[32].

On whether machines can read the unread tablets, four positions are live. The first holds that neural translation already produces useful drafts for well-edited genres and will scale to the backlog as training data grows; it rests on the Akkadian results and on the fact that most of the backlog is formulaic administrative text⁠[17]⁠[29]. The second holds that reading is a pipeline from photograph to signs to transliteration to translation, and that the first stage, sign detection on damaged, curved, variably lit clay, is the bottleneck; DeepScribe's drop from 0.89 to 0.80 when its stages are chained, and the 2026 detector's sensitivity to damage, are the evidence⁠[19]⁠[31]. The third holds that the realistic role of machines is assistance: proposing joins, fillings and sign readings for a human to check, which is how the Fragmentarium's 1,500 joins were made and how the restoration models were evaluated⁠[12]⁠[20]⁠[28]. The fourth holds that the backlog is mostly administrative text whose difficulty is not translation but interpretation, which needs names, year dates and archive context that no model has, so that machines will catalogue and transliterate while the reading stays human⁠[6]⁠[7].

On the Early Dynastic literary texts, three positions. One treats UD.GAL.NUN as a learned orthography, a set of substitutions that can in principle be fully tabulated, after which the texts can be read like any Sumerian⁠[8]. A second takes it as evidence of a dialect, perhaps early Emesal, which would make the texts a different language variety from the later copies⁠[25]. A third holds that, whatever the system, texts without a later parallel cannot be read with confidence because the spelling omits what grammar needs, so only compositions like the Kesh hymn, with Old Babylonian duplicates, will ever be secure⁠[27]⁠[9]⁠[10].

On Sumerian's origins, the default is that it is an isolate with no detectable substratum⁠[16]⁠[14]; Whittaker's Euphratic is the only live minority view⁠[15].

For the machine question, a held-out test is the obvious instrument: take tablets that have been photographed but never transliterated, run the full pipeline, then have specialists transliterate the same tablets blind and score the agreement; no such test has been published, and every reported figure comes from texts already edited⁠[17]⁠[31]⁠[19]. A second test is genre transfer: models trained on Neo-Assyrian royal inscriptions or Persepolis Elamite applied to Old Babylonian letters or Ur III accounts, with the error rate reported. The data for both exist. CDLI's bulk dump gives every public transliteration, SumTablets gives 91,606 Sumerian tablets as glyph-transliteration pairs under CC BY 4.0, and the Akkademia repository gives the ORACC training sets⁠[33]⁠[34]⁠[30]. What is missing is a benchmark of unpublished tablets with later expert readings, which only a museum partnership can provide.

For the Early Dynastic texts, the decisive evidence would be new duplicates: an Old Babylonian copy of a Zame hymn or of a UD.GAL.NUN myth would do for those texts what the Old Babylonian Kesh hymn did for its Abu Salabikh ancestor⁠[27]. Short of that, a complete table of UD.GAL.NUN substitutions tested against the whole Fara and Abu Salabikh corpus would show whether the system is regular enough to be an orthography or variable enough to be something else⁠[8]⁠[25].

For the Sumerian question, nothing short of a related language could settle affiliation, and none is expected. What can be settled is phonology, by systematic study of how Akkadian and other scribes transcribed Sumerian, and by the kind of information-theoretic measurement of the script that Hermalin applied to Sumerian in 2023⁠[35]⁠[16].