Nearly two-thirds of religious books sampled on Amazon were likely written by artificial intelligence, according to a new report from AI detection firm Originality.ai.
The report, released Wednesday, examined 2,034 recently published books across 14 religious and belief categories on Amazon. Originality.ai flagged 1,272 titles as likely AI-written, equal to 63% of the sample.
Witchcraft titles posted the highest rate
The highest share appeared in Amazon’s Wicca, Witchcraft & Paganism category, where 78% of sampled books were classified as likely AI-written. The titles covered topics such as healing crystals, herbal remedies, and cleansing “negative energies.” Michael Fraiman, the report’s author, told Decrypt that many of the witchcraft books centered on those subjects.
“One can picture the ideal customer as someone looking for solutions to heal themselves, improve their mental health, or ‘detox' from commonplace chemicals and drugs,” Fraiman told Decrypt. “These Wiccan books can address these concerns without being held to any scientific standard.”
Hinduism ranked second at 76%, followed by Taoism at 74%. Sikhism, Buddhism, Islam, Catholicism, Protestantism, Orthodox Christianity, and Judaism fell between those levels, though the report excerpt did not list each percentage. Mormonism came in at 42%, atheism at 40%, and Satanism at 22%. Researchers said sample sizes varied significantly between categories.
How the study classified books
Fraiman said researchers analyzed book descriptions, author bios, and content samples. Text scoring 50 or higher was labeled “Likely AI.” He added that the findings point to likelihood, not proof.
“We don't make claims to certainty about whether books are definitively AI-written,” he said. “Our model determines the likelihood that something was AI-written to a degree of certainty.”
Fraiman also said the research team spot-checked books that had been flagged as likely AI-generated.
“By this point, our research team can tell the difference between likely AI-written content and human-written content,” he added.
Detector reliability remains contested
While Fraiman said he is confident in the findings, AI detectors have also produced false positives and conflicting results.
In October 2024, a test of four AI detectors reached different conclusions when analyzing the U.S. Declaration of Independence. ZeroGPT said the text was 97.93% AI-generated, QuillBot judged it entirely human-written, and GPTZero gave it an 89% probability of being human-written.
More recently, in March, Colombia’s Supreme Court rejected a legal filing after detectors flagged it as AI-generated. When an attorney submitted the court’s ruling to the same detector, it was rated 93% AI-generated.
Amazon did not immediately comment
Decrypt reported that Amazon did not immediately respond to a request for comment.
“Amazon is aware of how common AI-written content is on their platform, and uses the honour system for authors to self-identify their work as AI-written,” Fraiman said, adding that the company did not share its findings with Amazon. “We have no data on how many authors actually disclose that.”
Book scanning for AI training has also drawn scrutiny
The use of physical books for AI training has also come under scrutiny, as developers buy millions of titles, strip them apart, scan their pages into training datasets, and discard the originals.
In August, an investigation tracked a shipment of rare books to an Amazon facility in Las Vegas, where workers cut off the bindings and scanned the pages for AI training data. Amazon confirmed that it buys books through commercial channels for that operation.