Revealing hidden language patterns in the Bible, with the help of AI

AI is transforming every industry, from medicine to film to finance. So, why not use it to study one of the world’s most revered ancient texts, the Bible? 

An international team of researchers, including Shira Faigenbaum-Golovin, assistant research professor of Mathematics at Duke University, combined artificial intelligence, statistical modeling and linguistic analysis to address one of the most enduring questions in biblical studies: the identification of its authors.  

By analyzing subtle variations in word usage across texts, the team was able to distinguish between three distinct scribal traditions (writing styles) spanning the first nine books of the Hebrew Bible, known as the Enneateuch. Using the same AI-based statistical model, the team was then able to determine the most likely authorship of other Bible chapters. Even better, the model also explained how it reached its conclusions. 

But how did the mathematician get here?  

In 2010, Faigenbaum-Golovin began collaborating with Israel Finkelstein, head of the School of Archaeology and Maritime Cultures at the University of Haifa, using mathematical and statistical tools to determine the authorship of lettering found on pottery fragments from 600 B.C. by comparing the style and shape of the letters inscribed on each fragment. Their discoveries were featured on the front page of The New York Times.  

“We concluded that the findings in those inscriptions could offer valuable clues for dating texts from the Old Testament,” Faigenbaum-Golovin said. “That’s when we started putting together our current team, who could help us analyze these biblical texts.” 

The multidisciplinary undertaking was made up of two parts. First, Faigenbaum-Golovin and Finkelstein’s team — Alon Kipnis (Reichman University), Axel Bühler (Protestant Faculty of Theology of Paris), Eli Piasetzky (Tel Aviv University) and Thomas Römer (Collège de France) — was made up of archaeologists, biblical scholars, physicists, mathematicians and computer scientists. The team used a novel AI-based statistical model to analyze language patterns in three major sections of the Bible. They studied the Bible’s first five books: Deuteronomy, the so-called Deuteronomistic History from Joshua to Kings, and the priestly writings in the Torah.  

Results showed Deuteronomy and the historical books were more similar to each other than to the priestly texts, which is already the consensus among biblical scholars.  

“We found that each group of authors has a different style — surprisingly, even regarding simple and common words such as ‘no,’ ‘which,’ or ‘king.’ Our method accurately identifies these differences,” said Römer.  

To test the model, the team selected 50 chapters from the first nine books of the Bible, each of which has already been allocated by biblical scholars to one of the writing styles mentioned above. “The model compared the chapters and proposed a quantitative formula for allocating each chapter to one of the three writing styles,” said Faigenbaum-Golovin. 

In the second part of the study, the team applied their model to chapters of the Bible whose authorship was more hotly debated. By comparing these chapters to each of the three writing styles, the model was able to determine which group of authors was more likely to have written them. Even better: the model also explained why it was making these calls.  

“One of the main advantages of the method is its ability to explain the results of the analysis — that is, to specify the words or phrases that led to the allocation of a given chapter to a particular writing style,” said Kipnis. 

Since the text in the Bible has been edited and re-edited many times, the team faced big challenges finding segments that retained their original wording and language. Once found, these biblical texts were often very short — sometimes just a few verses — which made most standard statistical methods and traditional machine learning unsuitable for their analysis. They had to develop a custom approach that could handle such limited data.  

Limited data often brings fears of inaccuracy. “We spent a lot of time convincing ourselves that the results we were getting weren’t just garbage,” said Faigenbaum-Golovin. “We had to be absolutely sure of the statistical significance.”  

To circumvent the issue, instead of using traditional machine learning, which requires lots of training data, the researchers used a simpler, more direct method. They compared sentence patterns and how often certain words or word roots (lemmas) appeared in different texts, to see if they were likely written by the same group of authors. 

A surprising find? The team discovered that although the two sections of the Ark Narrative in the Books of Samuel address the same theme and are sometimes regarded as parts of a single narrative, the text in 1 Samuel does not align with any of the three corpora, whereas the chapter in 2 Samuel shows affinity with the Deuteronomistic History (Joshua to Kings). 

Looking forward, Faigenbaum-Golovin said the same technique can be used for other historical documents. “If you’re looking at document fragments to find out if they were written by Abraham Lincoln, for example, this method can help determine if they are real or just a forgery.” 

“The study introduces a new paradigm for analyzing ancient texts,” summarized Finkelstein. 

Faigenbaum-Golovin and her team are now looking at using the same methodology to unearth new discoveries about other ancient texts, like the Dead Sea Scrolls. She emphasized how much she enjoyed the long-term cross-disciplinary partnership.  

“It’s such a unique collaboration between science and the humanities,” she said. “It’s a surprising symbiosis, and I’m lucky to work with people who use innovative research to push boundaries.” 

# 

REFERENCE: Faigenbaum-Golovin S, Kipnis A, Bühler A, Piasetzky E, Römer T, Finkelstein I (2025) Critical biblical studies via word frequency analysis: Unveiling text authorship. PLoS One 20(6): e0322905. https://doi.org/10.1371/journal.pone.0322905 

FUNDING: Alon Kipnis was supported in part by funding from the Koret Foundation and the BSF under Grant No. 2022124. Shira Faigenbaum-Golovin is grateful to the Eric and Wendy Schmidt Fund for Strategic Innovation, the Zuckerman-CHE STEM Program, the Simons Foundation under Grant Math+X 400837, and Duke University for supporting her research.