Xinzhiyuan Report
Editors: Ding Hui, Hao Kun
【Xinzhiyuan Guide】Scientists use AI to reconstruct the Dead Sea Scrolls timeline, shaking the community! New research indicates that parts of the Books of Daniel and Ecclesiastes scrolls were actually composed much earlier, even revealing clues about biblical authorship. The AI model Enoch, combining carbon-14 dating with handwriting analysis, pioneers an AI dating method that significantly surpasses traditional palaeography.
Just now, scientists used AI to redefine the historical timeline of the Dead Sea Scrolls!
Many scrolls are actually much older than previously thought, with two Dead Sea Scrolls being the earliest fragments of the Book of Daniel!
The study also shows that Ecclesiastes was likely completed by an anonymous author in the third century BCE, rather than by King Solomon as traditionally believed!
Even more explosive is the potential for AI to help clarify the authorship of the Bible.
Researchers from the University of Groningen, the University of Southern Denmark, the University of Pisa, and KU Leuven created a new AI program called Enoch, based on radiocarbon dating data from 24 ancient scroll samples and machine learning handwriting analysis techniques.
The result was astounding – according to research published in the journal PLoS ONE, many scrolls are older than previously believed!
Paper address: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0323185
“It’s like a time machine,” the paper’s authors enthusiastically stated. Enoch allows us for the first time to “see with our own eyes” the hand of those who wrote the Bible, and also provides new research avenues for other ancient documents.
Who would have thought that AI could help us reconstruct historical memories from two thousand years ago?
For years, scholars studying the Dead Sea Scrolls have employed various methods to analyze these ancient parchment scrolls: for example, X-rays, multispectral imaging, “virtual unfolding,” and palaeography (i.e., the study of writing style characteristics).
These scrolls are believed to date back to between the third century BCE and the first century CE, but these dates primarily relied on palaeographical analysis, as only a few scrolls explicitly recorded dates.
However, traditional palaeographical methods are inherently subjective and rely on the experience of specific scholars.
As previously reported, these ancient Hebrew texts – totaling approximately 900 complete or fragmentary scrolls, stored in clay jars – were initially discovered scattered across multiple caves north of the Dead Sea by Bedouin shepherds between 1946 and 1947.
Psalm Scroll (11Q5), one of the 972 texts of the Dead Sea Scrolls, partly transcribed in Hebrew
It is said that a shepherd, while searching for a lost sheep, casually threw a stone, accidentally breaking a clay jar containing scrolls, thus leading to this discovery.
Qumran was destroyed by the Romans around 73 CE, and historians believe that the Essenes hid these scrolls in the caves to protect them from damage.
The natural limestone environment and conditions within the caves helped preserve these scrolls for thousands of years.
This is not the first time AI tools have been applied to analyze handwriting on the Dead Sea Scrolls.
As early as 2020, it was reported that AI was used to assist in analyzing the handwriting on the Great Isaiah Scroll.
Because the writing style of the Isaiah Scroll appeared quite uniform, most scholars once believed it was completed by a single scribe. However, some suggested that it might have been the work of two scribes with similar styles, each responsible for copying different halves of the scroll.
The 2020 analysis results showed that the text indeed likely came from two scribes.
The analysis also indicated that the second scribe's handwriting was more variable than the first's, although the two styles were very similar, suggesting they might have received common training.
Enoch's Debut
Before this, the team had proposed a neural network for discovering ink handwriting patterns in digitized manuscripts, capable of geometric analysis at a microscopic level.
Enoch was developed from this.
Researchers believe that given the small size of the scroll dataset, the most sensible approach is not to rely on pre-trained models, but to “let the available data speak for itself.”
“Enoch emphasizes common features and similarity matching between training and test manuscripts, while traditional palaeography focuses on subtle differences considered indicative of style development. The combination of difference matching and adaptive reinforcement learning can reveal hidden patterns.”
From the model’s results, many Dead Sea Scrolls date earlier than previously estimated based solely on palaeography.
This point should be significant for answering questions such as when two ancient Jewish writing styles – “Hasmonean” and “Herodian” – developed.
The former was once thought to appear between 150 and 50 BCE, but researchers believe the Hasmonean style may have appeared earlier; the same applies to the Herodian script.
Therefore, these two scripts may have coexisted since the late second century BCE, challenging the previous common view that they appeared or coexisted only until the mid-first century BCE.
Enoch may even help clarify the authorship of the Bible.
For example, researchers concluded that two of the scrolls are the earliest known fragments of the Book of Daniel, which is believed to have been finalized by an anonymous author around 160 BCE.
And Ecclesiastes was most likely completed by an anonymous author in the third century BCE, rather than by King Solomon in the tenth century BCE as traditionally believed.
Regarding this, the researchers excitedly stated:
With the Enoch tool, we have opened a new door to the ancient world. It is like a time machine that allows us to study the handwriting of those who wrote the Bible, especially now that we have for the first time confirmed that two biblical scroll fragments indeed originate from the era of their presumed authors!
“It’s truly exciting to take a significant step forward in solving the dating problem of the Dead Sea Scrolls and to create a new tool that can be used to study other partially dated manuscript collections in history.”
Integrating Multiple Dating Methods
Radiocarbon Dating
Researchers performed carbon-14 dating on 30 manuscripts from 4 locations, estimated to span 5 centuries –
25 from Qumran caves, 1 from Masada, 2 from Murabba’at caves, and 2 from Nahal Hever caves.
This study is the first to use a specially designed chemical treatment method – solvent extraction – to remove fatty substances before carbon-14 dating the scrolls.
Before and after sample pretreatment, other specialized analytical chemistry methods were also applied to demonstrate that the total amount of lipid material was below a threshold that would not significantly distort the carbon-14 age.
Samples were measured by two accelerator mass spectrometers (AMS).
The AMS results yielded 27 valid carbon-14 age data, improving and expanding the existing series of carbon-14 dated Dead Sea Scrolls.
Overall, the carbon-14 results indicate that the dates, both for individual manuscripts and for the appearance of the so-called “Hasmonaean” and “Herodian” scripts, are earlier.
Figure 1 shows a comparison between the accepted calibrated age ranges and traditional palaeographical estimates (blue and red).
Figure 1: Overview of date estimates from three sources versus calendar dates
As can be seen, only two manuscripts' age ranges point to potentially later dates. Most carbon-14 results confirm the basic distinction between earlier Hasmonean manuscripts and later Herodian style manuscripts, as well as the distinction between so-called “Archaic” and Hasmonean manuscripts.
However, manuscripts traditionally considered to be from the Hasmonean and Herodian periods show vastly different distributions of carbon-14 age ranges on the timeline.
As shown in Figure 1 (blue section), Hasmonean manuscripts are concentrated in a narrower part of the timeline, while Herodian manuscripts are more scattered across the timeline, extending back from the second century CE to the second century BCE.
Sample 4Q114 is one of the most important findings among the carbon-14 results. This manuscript preserves Daniel chapters 8-11, which scholars have dated to the 160s BCE based on literary-historical evidence.
The accepted calibrated age range for 4Q114 is 230-160 BCE, which overlaps with the presumed period of composition for the final part of the biblical Book of Daniel.
AI Age Prediction Model
Due to the lack of large, general, representative, and labeled datasets applicable to the scroll period, researchers applied specialized pattern recognition and machine learning models, using only relevant scroll data to train the age prediction model.
Given the importance of this topic, using pre-trained deep transfer learning based on external materials might reasonably concern palaeographers about the correlation between the target data from these scrolls and training data from potentially very different sources and periods.
Researchers used 24 manuscripts with valid dates from carbon-14 samples as the labeled dataset, forming the primary training set for the Enoch model.
For data labeling, researchers used OxCal v4.4.2 to obtain the raw data points of probability distributions. This is because carbon-14 results are not single dates like dated documents, but rather represent age ranges with probability distributions. The carbon-14 data input used to train Enoch included the probability distributions of the accepted calibrated age ranges.
In cases where there was evidence of bimodality, researchers used a Bayesian approach outside of the OxCal program to constrain the age range produced by each calibration.
Researchers did this to prevent carbon-14 bimodality from propagating to the next processing stage of training a style-based predictor. Palaeographical domain knowledge was only used when evidence of bimodality was present, as it allowed for a binary split of the (bimodal) probability distribution given by OxCal using a Heaviside function, with the step position set at a low-probability point on the curve that had small and near-zero influence.
In addition to the primary training set, researchers also created different combinations of training data for comparative analysis and to further test the robustness of the model.
These combinations included experimental additions or removals of 4Q52, some previously tested carbon-14 samples, dated documents from the fifth to fourth centuries BCE and the second century CE, and the Maresha ostracon from 176 BCE.
Researchers performed cross-validation in three ways: using a training-validation set split, by splitting manuscript images into training-test sets, and by using leave-one-out for training data points.
Style-Based Dating of "Undated Manuscripts"
Before applying the Enoch model to other previously undated scrolls, researchers first successfully tested it on a known medieval charter benchmark dataset, MPS.
Then, researchers applied the trained Enoch model to 135 unseen manuscripts from approximately 1000 Dead Sea Scrolls to explore the feasibility of style-based dating at this stage.
Thus, the Enoch model produced an empirical assessment that modified a previously uniform expected age distribution into a curve distribution, making certain ages more probable for a sample and others less probable.
Similar to the OxCal program used for carbon-14 dating, the Enoch model provides age probabilities and corresponding error estimates. This is an advancement compared to older, more rudimentary methods that provide only a single age point as an answer.
When analyzing the researchers’ results, it is necessary to consider simultaneously (1) the probability of an age point and (2) the reliability of that point estimate. These are the first published age estimates for this manuscript collection. Future research, utilizing more data and improved images, aims to further validate and refine them.
Table 1: General workflow for Enoch model analysis of manuscript images
Expert palaeographers among the authors of this paper evaluated the style-based age predictions, categorizing them into two main categories: reasonable and unreasonable, with the latter further subdivided into too early and too late.
As shown in Table 2, according to the palaeographers’ evaluation, 107 (79.3%) of the undated manuscripts had reasonable dating results, while 28 unreasonable predictions (20.7%) were divided into too early (7.4%) and too late (13.3%).
Table 2: Palaeographical post-assessment of Enoch model age predictions for 135 undated manuscripts
Previously, researchers demonstrated that the Great Isaiah Scroll was completed by two scribes.
Now, the Enoch model shows that there is no time difference between the two parts of the manuscript, meaning one part was not written much later than the other. Instead, it is estimated that the two scribes copied the corresponding parts of the 1QIsaa scroll at the same period.
Figure 2 shows that the Enoch model consistently dates both parts to between 180-100 BCE.
Figure 2: Enoch’s date prediction chart for the two halves of the 1QIsa scroll
Dating Ancient Manuscripts with Enoch
To the researchers’ knowledge, Enoch is the first complete machine learning model that uses raw image input, leverages the probability distributions from carbon-14 outputs, provides probabilistic age predictions for handwritten manuscripts, and is supplemented by palaeographical input, while ensuring transparency and interpretability through its explainable design.
Furthermore, Enoch's integration of multiple dating methods significantly enhances the value of evidence sources and allows evidence from two sources – physical (material) and geometric (shape-based) – to corroborate each other.
As an illustrative example, Figure 3 shows that Enoch was able to provide an age prediction estimate for 4Q319 that is similar to the accepted 2σ calibrated range for 4Q259.
Figure 3: Enoch’s date prediction estimate for the 4Q319 scroll
The advantage of the Enoch model is that it provides quantified objectivity to palaeography by empirically offering a limited number of probability-based options – including physical (carbon-14) and geometric (shape-based analysis) evidence – reducing the subjectivity and implicit expert knowledge typical of the method. These options can help palaeographers confirm, refine, or modify their own estimates for individual manuscripts.
Additionally, the methods underpinning Enoch can be used for age prediction of other partially dated manuscript collections.
Finally, instead of adopting any off-the-shelf models, researchers developed a robust model capable of (1) predicting ages using only a very small amount of data (i.e., 24 samples or data points), (2) handling uncertainty, and (3) providing interpretability.
In manuscript analysis, Enoch differs from traditional palaeographical methods.
Enoch emphasizes common features and similarity matching between training and test manuscripts, whereas traditional palaeography focuses on subtle differences considered indicative of style development. The combination of difference matching and adaptive reinforcement learning can reveal hidden patterns.
This interdisciplinary fusion may enrich researchers’ understanding of text content, material properties, and historical context, thereby enhancing interpretations of the past.
New carbon-14 evidence, or as new discoveries are made, an entire batch of manuscripts with dating information can be added to Enoch’s training data to further refine and improve accuracy, continuously enhancing accuracy.
The impact of each new manuscript sample added to the Dead Sea Scrolls carbon-14 reference collection can now be easily calculated using the Enoch method.
Development of Ancient Aramaic/Hebrew
The results of this study offer four new insights into the development of Aramaic/Hebrew script during the studied period and the dating of individual manuscripts.
First, carbon-14 age ranges and Enoch’s style-based estimates are generally earlier than previous palaeographical estimates.
The earlier dating of these scrolls is realistic. The accepted 2σ calibrated range for Hasmonean manuscripts allows for earlier dates, traceable back to the first half of the second century BCE, sometimes even slightly earlier, and not just approximately 150-50 BCE.
There are no compelling palaeographical or historical reasons to exclude these earlier dates as reliable time markers for “Hasmonean” script. This also applies to the accepted 2σ calibrated range for 4Q70 and its “Archaic” script.
Second, “Herodian” script appeared earlier than previously thought.
This indicates that the transition from “Hasmonean” to “Herodian” script did not begin only from the mid-first century BCE, but rather they coexisted much earlier.
Third, this novel palaeographical method has yielded new scroll datings, which impact our understanding of ancient Judean history and the people behind the scrolls.
Hypotheses regarding whether the movement behind the scrolls originated in the second or first century BCE need to be re-evaluated based on Enoch’s dating predictions of the second century BCE for Hasmonean manuscripts (e.g., 1QS and 4Q163), which contain texts considered characteristic of this movement.
Scholars generally assume that the rise and expansion of the Hasmonean dynasty from the mid-second century BCE led to increased literacy and fostered a culture of copying and knowledge.
However, the results of this study demonstrate that multiple literary manuscripts were already copied before this period. An example is 4Q109, a copy of the biblical Book of Ecclesiastes, which scholars initially dated to the end of the third century BCE, while Enoch gives an age prediction of the third century BCE, close to archaic manuscripts such as 4Q52 and 4Q70 – copies of the biblical Books of Samuel and Jeremiah, respectively.
Fourth, this study’s carbon-14 results for 4Q114 and Enoch’s age prediction for 4Q109 now identify these fragments as the earliest known biblical scrolls from the era of their presumed authors.
Therefore, the results of this study break down unconfirmed historical assumptions and dating constraints, and challenge the validity of the default model's relative typology.
This relative typology can only be maintained under limited circumstances.
The distribution of Hasmonean manuscripts on the timeline did not significantly impact the default relative typology, but the earlier, second-century BCE age range of Herodian manuscripts challenges the relative typology.
References:
https://arstechnica.com/science/2025/06/are-dead-sea-scrolls-older-than-we-thought/?utm_source=flipboard&utm_content=topic%2Fartificialintelligence
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0323185&utm_source=pr&utm_medium=email&utm_campaign=plos006