More than 63% of religious and spiritual books recently published on Amazon were likely written by artificial intelligence, according to a study released this week by AI detection firm Originality.ai. The research, which sampled 2,034 titles across 14 belief categories, found that machine-generated text has quietly become the majority in a category readers often turn to for guidance on health, healing, and personal meaning.
Originality.ai flagged 1,272 books as likely AI-written. The highest rate appeared in the witchcraft category at 78%, followed by Hinduism at 76% and Taoism at 74%. Mormonism came in at 42%, atheism at 40%, and Satanism was the lowest at 22%. Researchers cautioned that sample sizes varied between categories, which limits how directly the percentages can be compared.
How the study worked
Report author Michael Fraiman said his team analyzed book descriptions, author biographies, and text samples from each title, running them through a scoring model. Anything that scored 50 or higher was classified as "Likely AI." Fraiman was careful to frame the findings as probability rather than proof.
"We don't make claims to certainty about whether books are definitively AI-generated," he told Decrypt. "Our model determines the likelihood that something was AI-generated to a degree of certainty." He added that researchers spot-checked flagged titles manually, saying the team can "tell the difference between likely AI-written content and human-written content" by this point in their work.
The witchcraft category - pulled from Amazon's Wicca, Witchcraft & Paganism listings - also produced the most concerning overlap between AI authorship and factual accuracy. Researchers flagged 53% of fact-checkable claims in witchcraft titles as potentially false, covering books on healing crystals, herbal remedies, and cleansing negative energy. Fraiman suggested the appeal to publishers is obvious: these books can promise solutions for mental health or "detoxing" without needing to meet any scientific standard.
For writers working in spiritual publishing or adjacent nonfiction categories, the study raises a practical question about competition: AI-generated titles now outnumber human-written ones in a major commercial category. Understanding how detection works and where it fails is becoming part of the job for authors who want to distinguish their work. Resources on Generative AI and LLM tools can help writers understand what these systems produce and how to spot machine-generated text.
AI detection is not a verdict
Detection tools like Originality.ai's model are built to estimate likelihood, not deliver a verdict. That distinction matters, because AI-detection technology as a category has a track record of inconsistency.
In October 2024, a test of four separate tools analyzing the U.S. Declaration of Independence produced wildly different verdicts: ZeroGPT rated it 97.93% AI-generated, QuillBot called it entirely human-written, and GPTZero gave it an 89% probability of human authorship. Three tools, three contradictory answers on the same 18th-century document.
The stakes got more serious in March, when Colombia's Supreme Court rejected a legal filing after a detector flagged it as AI-generated. When a lawyer later ran the court's own ruling through the same detector, it came back rated 93% AI-generated. That episode underlines why treating detection scores as courtroom-ready evidence remains legally shaky.
Amazon's honor system and training data
Amazon sits at the center of two separate but related controversies: how it polices AI-written books sold on its marketplace, and how it sources raw text to train its own AI models. Neither practice currently comes with much public transparency.
According to Fraiman, "Amazon is aware of how common AI content is on their platform, and uses the honor system for authors to self-disclose their work." Originality.ai did not share its findings with Amazon before publishing, and Fraiman noted plainly: "We have no data on how many authors actually disclose that." Amazon did not immediately respond to Decrypt's request for comment.
Separately, an investigation tracked a shipment of rare books to an Amazon facility in Las Vegas, where workers cut bindings off books and scanned the pages into AI training datasets. Amazon confirmed that it "purchases books through commercial channels to help develop and improve the products and services our customers use," though the company's statement stopped short of confirming AI training specifically.
For writers, the training data question is not abstract: physical books - including rare ones - are being destroyed to build the systems that generate competing content. Authors have limited recourse under Amazon's current disclosure policies, and no public data exists on how many writers actually self-report AI use. Writers tracking these developments can find ongoing coverage and practical guidance under the AI for Writers tag.
Why this matters for writers
Two takeaways stand out for anyone making a living from written work. First, AI detection scores are not reliable enough to prove authorship in any formal setting - so don't rely on them to protect your work, and don't assume they offer certainty when reviewing someone else's. Second, Amazon's honor system for AI disclosure means the marketplace is effectively self-policing, which leaves human authors competing against machine-generated books that face no verification requirement. If you publish on Amazon, assume AI-generated titles are already in your category and plan your positioning accordingly - pricing, packaging, and platform presence will matter more than ever when the product itself is cheap to produce.
Your membership also unlocks:
