63% of Amazon's religious titles look AI-written, and the only gate is a checkbox

Nothing about model access varies across a retailer's category tree, which makes the spread inside the sample more informative than its average: 56 points separate the Wicca, Witchcraft & Paganism shelf at 78% from Satanism at 22% ^[15][2][3]. What varies is whether a buyer can check the text against anything outside it. Report author Michael Fraiman described the ideal customer for the witchcraft titles as someone looking to heal themselves or detox from commonplace chemicals, and said such books can address those concerns "without being held to any scientific standard" ^[8]. The audit attaches a figure to that: 53% of fact-checkable claims in the witchcraft sample were flagged as potentially false ^[5]. A wrong herbal remedy is wrong whoever typed it, which makes authorship the second problem on that shelf.

The evidence underneath the headline is softer than the headline. Originality.ai scored book descriptions, author bios and samples, treated anything at 50 or above as "Likely AI", and Fraiman told Decrypt the model reports likelihood rather than proof ^[7]. He said his team spot-checked flagged books and can tell the difference by now ^[9]. Weigh that against how detectors behave on text known to be human: in an October 2024 test of four tools on the Declaration of Independence, ZeroGPT returned 97.93% AI-generated, QuillBot called it entirely human, and GPTZero put it at 89% human-written ^[10]. In March, Colombia's Supreme Court rejected a filing after detectors flagged it, and an attorney then fed the court's own ruling to the same detector, which rated it 93% AI-generated ^[11]. The researchers also noted that per-category sample sizes vary significantly ^[6], which is where the 22% and the 78% deserve the most caution. The arithmetic is tighter than the caveats: 1,272 of 2,034 is 62.5%, and 762 titles in the sample were not flagged ^[16][17].

None of that reverses the direction of travel. It does mean the number works as a read on supply economics and not as an enforcement artifact, and supply is where the platform's role begins. Fraiman said Amazon is aware how common AI-written content is, relies on the honour system for authors to self-identify, and that his team did not share its findings with the company, with no data on how many authors actually disclose ^[12]. Amazon did not immediately respond to Decrypt's request for comment ^[13].

Set that against the company's conduct on the input side. 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 it purchases books through commercial channels for that operation ^[14]. Acquiring human-written text to feed models gets a purchase order, a warehouse and a scanning line. Establishing whether the text sold back to customers was machine-written gets a self-report field, filled in by the seller.