Legible at Last
AI is making inroads on centuries of accumulated historical artefacts that humans never yet managed to analyse.
Every archive holds material that has been catalogued, stored, and left unread, and the obstacle differs from case to case. A carbonised papyrus roll crumbles when opened. A fragment is perfectly legible and cannot be placed in time. A script survives with most of its meanings lost, or a collection outruns the number of people qualified to read it. Four projects reported over the past eighteen months each took on one of these problems.
Scanning the scrolls
The Vesuvius Challenge announced in June that it had virtually unwrapped PHerc. 1667 and read it from beginning to end. Vesuvius carbonised the scroll when it buried the Villa dei Papiri in 79 AD, preserving it on terms that kept it shut, because opening one of these rolls destroys it. Scans on the BM18 beamline at the European Synchrotron Radiation Facility in Grenoble resolved the layers, and the recovered writing surface ran to 1.4 metres of papyrus and 22 columns of Greek. The text is a Stoic treatise on ethics, and its final preserved column names Aristocreon, the nephew of Chrysippus, which places the work in the second century BC.
Carbon ink is close to invisible against papyrus in an X-ray scan, so the models learn to detect it on opened fragments where the ink can be seen, then infer its presence inside sealed rolls. The same pipeline recovered the title and attribution of a third scroll, PHerc. 139, identifying it as Philodemus, On Gods, Book 8. Reading a title tells a papyrologist what a closed roll contains before a single column of its body has been studied, which turns several hundred unopened scrolls from a backlog into a collection that can be sorted.
Signs of the times
Paleography dates a manuscript by the shape of its letters, and the method needs securely dated examples against which to calibrate. Almost none survive from the Judean Desert, leaving a gap between dated manuscripts of the fifth and fourth centuries BC and those of the first and second centuries AD. A team at the University of Groningen radiocarbon-dated 24 scroll samples and trained a model called Enoch on the relationship between those dates and measured features of the handwriting, down to the curvature and angle of individual strokes. Shown carbon-dated fragments whose dates had been withheld, Enoch guessed correctly 85 per cent of the time, occasionally within a narrower range than the carbon dating itself.
Applied afterwards to 135 scrolls that had never been tested, the model returned dates mostly older than paleography had proposed, with a mean error of around three decades. Fragments of Daniel and Ecclesiastes now fall within the lifetimes of their presumed authors, and the Ecclesiastes manuscript moved from a paleographic estimate of 175 to 125 BC back to somewhere between 300 and 240 BC. Several texts associated with the Qumran community turned out to predate the settlement at Qumran, which means that they were copied elsewhere and carried there.
Reading the bones
Excavations have produced more than 150,000 oracle bone fragments, the divination records of the Shang dynasty, carrying at least 4,500 distinct characters. Around 3,000 of those characters remain undeciphered. Fragmentary inscriptions and sparse attestation leave most of them without enough evidence to fix a meaning, and each stands for a lost word for some institution, technology, or belief of a civilisation three thousand years gone.
AlphaOracle, published this month in The Innovation, assembles the largest digitised corpus yet gathered for the problem, comprising 73,883 rubbings, 45,364 scans, and 23,755 scholarly studies. The system reproduces the sequence a philologist follows, working through the form of a glyph, then its context in the surrounding inscription, then the transmitted classical literature. Each candidate reading arrives as a report with its evidence attached, which is the condition under which an epigrapher can do anything with it.
Defeated by handwriting
Royal Navy officers recorded barometric pressure, air and sea surface temperature, and wind force and direction at fixed hours through the nineteenth century, wherever their ships happened to be. Those readings form the densest surviving observational record of the pre-industrial ocean, and most of them have never been converted into anything a computer can read. Weather Rescue at Sea, published in March, recovered millions of them.
The transcription was done by volunteers working through scanned pages on Zooniverse. Automated reading of handwritten tabular records remains poor enough that a methodological review of the field reported that recent attempts to apply AI have not been promising, defeated by cursive, inconsistent layouts, and the mixture of digits and words in a ruled column. The RECLAIM survey puts the untranscribed remainder at several hundred thousand logbooks, and possibly at several million.
Empowering the experts
Three of these four projects share an arrangement in which the model proposes and a specialist decides, and DeepMind’s Ithaca measured what the arrangement is worth. Working alone, Ithaca restored damaged Greek inscriptions with 62 per cent accuracy, and the historians restored them with 25. Working together they reached 72. The logbooks sit outside that arrangement, because no specialist will review line forty thousand of a pressure series, and a transcription feeding a climate reconstruction has to be right the first time. That work is still going out to volunteers.