“It’s like a time machine. So we can shake hands with these people from 2,000 years ago, and we can put them in time much better now,” said Professor Mladen Popović, University of Groningen, in reference to a discovery that is rewriting the chronology of the Dead Sea Scrolls. For centuries, the exact dating of these ancient manuscripts hitherto the domain of tedious palaeography and radiocarbon dating had been an issue of learned argument and technical constraint. Then a new paper combined artificial intelligence with improved radiocarbon dating to achieve a quantum leap in accuracy and challenge conventional wisdom concerning the origins and development of some of the most influential religious texts on the planet.
The Dead Sea Scrolls, which were found in the mid-20th century in Qumran caves, have served as a benchmark for learning about Jewish and Christian history. However, the overwhelming majority of the scrolls do not contain overt historical dates. Conventional dating has depended on two imperfect techniques: palaeography, which analyzes handwriting patterns to estimate age, and radiocarbon dating, which detects the deterioration of carbon-14 in organic matter. Both methods have their limitations. Palaeography is always subjective, and radiocarbon dating, while a wonder of postwar technology, can be biased by contamination and is bounded by the necessity to physically sample valuable artifacts.
The most recent breakthrough, published in PLOS One, presents a machine learning model called Enoch, after the biblical character famous for wisdom. Enoch was taught on a set of 24 Dead Sea Scrolls that had been through careful radiocarbon dating, with samples thoroughly cleaned to avoid residues like castor oil an artifact of the 1950s preservation methods that previously plagued results (initial attempts at dating were biased by application of the castor oil). The innovation of the AI model is its capacity to scan the geometry of ink marks at the level of characters and micro-curves, deriving mathematical characterizations of script form that are beyond human perceptual capabilities.
Enoch uses Bayesian ridge regression, a probabilistic method that combines the radiocarbon calibration curve uncertainty with the subtle variability of handwritten script. This combination allows Enoch to make date predictions with a mean absolute error of around 30 years a tolerance comparable to, and at times exceeding, the accuracy of radiocarbon dating alone (Enoch could predict 14C-based dates with mean absolute errors (MAEs) ranging from 27.9 to 30.7 years). In validation experiments, predictions by Enoch coincided with original radiocarbon probability distributions in more than 85% of instances.
This method is a break from past AI uses in manuscript research, which tended to use deep learning models that had been trained on irrelevant datasets. Rather, Enoch was specifically designed for the Dead Sea Scrolls, with only pictures and data expressly relevant to the time and scripts being studied (specific pattern recognition and machine-learning models, which only used the scrolls data pertinent to what they were being trained to do). The model’s feature extraction process uses both allographic mapping clustering characters by shape through the use of self-organizing neural networks and textural analysis, which captures the angular distribution of ink trace curvature, a technique that has been proven in writer identification and document dating research.
The consequences of Enoch’s predictions are far-reaching. It was used on 135 previously undated manuscripts, and 79% of its predictions were deemed realistic by expert palaeographers (Expert palaeographers tested the AI’s results, and found 79 per cent of them to be realistic predictions). Among the most striking findings: several scrolls, including fragments of Daniel (4Q114) and Ecclesiastes (4Q109), now appear to date from the very period of their presumed authorship, rather than being later copies as previously assumed. The Daniel fragment, for example, was radiocarbon dated between 230 and 160 BCE, dating it consistent with the events described in the text and assigning its composition back by as much as a century (Enoch’s prediction of a date range of 230–160 B.C.E., at the higher end of Rollston’s estimates).
More important to the writing’s history, however, is that Enoch’s analysis disrupts standard typology of ancient Jewish scripts. The investigation discovers that the Hasmonaean and Herodian script types, hitherto considered to have followed one another in a tidy chronological order, in fact existed simultaneously for much, much longer. Herodian script is found on manuscripts contemporary with the reign of King Herod by up to 50 years, whereas Hasmonaean script is now demonstrated to go as far back as the early second century BCE (the model postulates that there is evidence to suggest that both Hasmonaean and Herodian script styles existed as early as the late second century BCE). This discovery necessitates rethinking how literacy, scribal culture, and religious thought diffused in ancient Judea and the wider eastern Mediterranean.
The technical success of Enoch is also a testament to the improving standards in radiocarbon dating. Contemporary protocols now involve sophisticated chemical pretreatments to eliminate impurities and employ globally calibrated curves like IntCal20, which correct for local and seasonal fluctuations in atmospheric carbon-14 (Radiocarbon dating is essential for archaeology and environmental science to age everything from the ancient remains of the oldest modern human skeletons to ancient catastrophic volcanic eruptions). But as Professor Joan Taylor of King’s College London points out, radiocarbon dates the parchment, not the date of inscription, and AI programs are no stronger than data against which they are trained (radiocarbon dating merely illuminated how old the parchment was, not when it was inscribed).
However, the combination of AI and radiocarbon techniques is unveiling new horizons. Enoch’s non-destructive analysis allows hundreds of scrolls that are undated to be analyzed without destroying valuable material, providing a scalable way forward for both the Dead Sea Scrolls and other collections of ancient manuscripts (the AI process does not involve destructive sampling that traditional radiocarbon dating requires). As scientists increase the training data set and the model’s precision, it is hoped that the chronology of ancient documents will increasingly become more accurate, shedding light not only on the development of scripts, but on the intellectual and political history of the ancient world.
The convergence of artificial intelligence and radiocarbon science is revolutionizing manuscript research from an art of inference to a science based on empirical probability. As Popović noted, “What we have created is a very robust tool that is empirically based based on physics and on geometry.” For the Dead Sea Scrolls, and for scholars who analyze them, the past is becoming clearer one ink trace at a time.
