The Dead Sea Scrolls were written over centuries by different scribes. Dating them is essential for understanding the communities that produced them, the transmission of biblical texts, and the development of Second Temple Judaism. Two methods exist: radiocarbon dating of the physical material, and paleographic dating based on handwriting style.
Paleography is subjective. An expert examines letterforms — the curve of a bet, the length of a lamed's descender, the angle of a shin's strokes — and assigns a date range based on comparison with dated reference documents. Different experts disagree. The reference corpus is small. The uncertainty ranges are wide: a given scroll might be dated to “late second century BCE to early first century CE,” a span of 150 years.
Note (2026-11-15; read against the paper's full text): the essay calls the model a neural network and gives 28 to 31 years as the error of the combined approach. The paper's model (Enoch) is Bayesian ridge regression on handwriting-style feature vectors (one of the features comes from a Kohonen network), trained on the radiocarbon dates; the 27.9 to 30.7 year figure is the leave-one-out error of its style-based prediction against those radiocarbon dates, not the error of a combination. The paper says the radiocarbon ranges and Enoch's predictions are both often older than the traditional palaeographic estimates (for the 26 radiocarbon-dated samples compared, 17 overlap and 9 do not), so Enoch's older dates are not independent of the radiocarbon dates it was trained on. For the 135 unseen manuscripts, 79 percent of the predictions were judged 'realistic' in a post-hoc palaeographic evaluation.
Combining radiocarbon with AI handwriting analysis (arXiv:2407.12013) narrows the window dramatically. The AI component is not replacing paleography — it is formalizing it. Instead of a human expert's gestalt judgment about a scribe's hand, a neural network extracts quantitative features from character shapes, stroke patterns, spacing, and consistency. These features correlate with date, but the correlation is learned from the training data rather than hypothesized by the expert.
The combined approach achieves mean error of roughly 28–31 years. For manuscripts whose age spans millennia of interest, 30-year precision is transformative: it distinguishes texts written before the Hasmonean revolt from those written after it, texts from Herod's reign from those of his successors.
The surprise: many manuscripts appear older than traditional paleographic estimates indicated. The AI, unconstrained by the conventional periodization that paleographers use as scaffolding, finds earlier dates. The scaffolding that organized expert judgment was not wrong in its relative ordering — earlier-looking scripts are indeed earlier — but it was systematically compressed. The timeline was right in shape but wrong in scale.