But what has a jar in a desert cave to do with machine learning? A Bedouin shepherd who sought one of the limestone hills close to the Dead Sea in 1947 tossed a stone into a cave and heard it hit pottery. Within were rolls of parchment which were later to be named the Dead Sea Scrolls- fragments and manuscripts that assisted in attaching the early history of the Hebrew Bible and the communities that reproduced, edited, and transmitted these works through centuries. Most of the scrolls have been dated to about 2,000 years old over decades, and the debate among scholars has pitted two imperfect datings radiocarbon (accurate but destructive) and palaeography (non-destructive but subjective to the expert).
The third way has now been suggested by a research team: retain the physics, retain the handwriting, and have software measure the middle ground. The model proposed by them and published in the journal PLOS One is based on an AI model nicknamed Enoch, which learns a correlation between letter-shape patterns and radiocarbon date ranges, and then makes probabilistic estimates of dates of manuscripts that are not yet carbon dated.
The engineering specifics are important as they resolve a silent issue of manuscript science: the “known-date” examples are few and few. The researchers had Enoch trained to 24 different radiocarbon-dated scrolls, and the experiment was conducted with the unseen images of the same series. The standard error of the model, as found in the background description of the study, is approximately 30 years, which is a bit of uncertainty to cause historical interpretations to shift when contests depend on generations, but not centuries.
Radiocarbon dating, however, is as good as the sample preparation, and the scrolls have led an ill-travelled modern existence. there were fragments upon which some substances, which were to cause the ink to be easier to read, were applied–substances that can give trouble to measurements unless removed. The paper records a pre-treatment stage, meant to clean samples prior to the dating stage and radiocarbon analyses were carried out on 30 manuscript samples in four locations in the Judean Desert with 27 valid dates being obtained.
Having that physical baseline, Enoch was involved in reading manuscript images and asked to use the geometry of handwriting to make inferences about the dates. Instead of working with opaque deep-learning models which are trained on unrelated modern images, feature extraction of the ink traces is performed and a statistical regression model is employed to produce the probability distributions instead of a single year. When their system was used to apply to 135 previously un-datable Dead Sea Scroll manuscripts, expert palaeographers considered some 79% of the AI estimates “realistic,” which the researchers take to be a practical but not definitive limit to broad application.
The implications of the headline are not of software the software itself, but of timelines in ancient Judaea put together. Among the discoveries made is that two great styles of scripts, Hasmonean and Herodian, seem in fact to have been overlapping longer than is commonly supposed, and that it is difficult to construct the neat concept of one style superseding another.
Another is that certain manuscripts can be older even than the dates estimated using traditional handwriting, a change which is significant since the Qumran caves are typically discussed as a single production site but not a library, but a storage location of manuscripts produced elsewhere. According to one academic, Joan Taylor, the implications of these findings are that the majority of the manuscripts discovered in the caves around the vicinity of Qumran would not have been done so in the location of Qumran, which was not inhabited until later, which makes a dating refocus a geographic refocus.
The argument is tangible with the help of specific scrolls. The model and radiocarbon scale brought 4Q114, containing portions of Daniel, in an earlier date than some previous estimates, 230-160 BCE, and the project also identifies a copy of Ecclesiastes as falling into a later period relative to the life of his supposed author. Meanwhile, it has been pointed out by critics that sometimes, the textual and historical constraints can provide textual and historical constraints an edge in fact-checking date ranges that a handwriting model may lack, and this underscores the message of the study itself: AI outputs should not be the answer, but rather be supplementary.
In the case of an ensemble of some 1,000 biblical manuscripts and fragments, the appeal is simple: fewer samples must be severed to be tested by carbon analysis, whereas additional manuscripts can be assigned a date estimate taking the same physical reference frame. The bigger movement is cultural rather than technical–palaeography, long influenced by tacit knowledge, is being pushed to co-exist with measured indefiniteness, and replicable pipes, even in instances where the documents themselves are defragmented, fragmentary, and doggedly human.
