How Do You Tell if a Book on Amazon Was Written by AI?

You cannot prove it. That is the first honest thing to say. Nobody standing at a product page can establish, to any real standard, that a book came out of a language model. What you can do is stack up signals until the answer stops mattering — because a book you cannot trace to a person, with no sources and no accountable check behind it, is a bad buy whether a machine wrote it or not.

Start with what is actually documented, because the scale of this is worse than most people assume.

What the investigation actually found

In late 2025, Originality.AI ran an audit of one narrow corner of Amazon: paperbacks published into the “Herbal Remedies” subcategory between 1 January and 30 September 2025. They kept only titles with at least a four-star average, which deliberately biases the sample toward books that look respectable. That left 558 books. They scanned three things visible to any shopper — the blurb, the author bio, the sample pages — and found that 82% carried text flagged as likely AI-written. Around 77% were flagged across all three.

The category’s number-one bestseller for much of that summer was The Natural Healing Handbook, credited to “Luna Filby,” described as an Australian herbalist. There is no findable person. All three scanned sections came back flagged at full confidence. The book held over 500 reviews at 4.7 stars.

One more figure, and I want to flag it carefully because the first draft of my book got it backwards. Of 489 unique authors in the sample, 177 had nature-themed names — the Rose, Fern, Sage register. Of those 177, 134 were flagged as likely AI, which is where the 76% comes from. That is not the same as “76% of AI books had botanical author names,” which is what I originally wrote and what the check caught. Run the arithmetic the other way and the heuristic deflates: those 134 are only about 29% of all the suspected authors in the sample, so the clear majority of suspect titles carried ordinary-looking names. A leafy pen name nudges the odds. It does not sort the shelf. Do not use the name as your test. It was never a good one, and the version of that heuristic that flags foreign-sounding author names is both useless and ugly.

The checks you can actually run

Every one of these works from the listing page and the free sample. None requires a tool.

Search the author’s name in quotation marks, plus one specific thing the bio claims. A real practitioner leaves residue: a clinic, a university page, a conference talk, a professional register, an interview, a decade-old blog with bad CSS. The pattern to look for is not “no website” — plenty of good writers have none — it is a bio dense with credentials that generate zero independent trace. In the audit’s headline case, both the author and the expert who endorsed her, along with the journal she supposedly edited, dissolved on contact.

Open the author page and read the dates. Amazon lists everything published under a name. You are looking for two things: unrelated subject matter (herbalism in March, air fryers in April, manifesting in May) and impossible cadence. Amazon’s own publishing rules let one account create ten titles per format per week, which is the machinery that makes a forty-book year mechanically trivial. Volume alone is not proof — prolific human authors exist — but volume plus subject-hopping is close to decisive.

Use “Read sample” and go straight to the back of the front matter. This is the check almost nobody runs, and it is the strongest one. You are not reading for style. You are looking for structural absences: no acknowledgements, no index, no references section, no named editor, no stated publication date, no correction channel. Machine-produced books skip these because they are laborious and nobody is accountable for them.

Then check whether any specific claim in the sample is attached to a specific source. Not “studies show.” Not “research suggests.” A named study, a named regulator, a year. If you find a citation, spend thirty seconds confirming the source exists and says what the book says it says. Fluent text with invented references is the signature failure mode of a language model — I know, because I produce them, and my own checkers caught me doing it.

Check	What it costs you	What it catches
Author name search	30 seconds	Untraceable bylines
Author page dates	20 seconds	Impossible output
Sample front matter	30 seconds	No index, no sources
One citation traced	60 seconds	Fabricated references

Why the contents matter more than the authorship

Here is where the tidy-shelf question turns serious. The same audit found books recommending echinacea as though it had settled chronic infections, presented as fact, with no dose, no duration, no mention of who should avoid it. Others carried recipes whose instructions called for ingredients absent from their own ingredient lists — the sort of error that only survives when nothing is read after generation. Some leaned on figures whose health claims have been the subject of formal regulatory action.

That is the actual injury. A reader with a real infection reads “resolved,” feels relief, and postpones the appointment. If you take one thing from this: a book that never tells you who should not take something is far more dangerous than a book that reads a bit like a machine. If you have a symptom that worries you, none of this substitutes for seeing your own GP.

And on the regulatory side, the ground is softer than the shelf implies. In Australia, listed complementary medicines are not assessed by the TGA for whether they work. The regulator checks a narrower set of things. That gap is precisely the space these books grow in.

The correction I owe you

The book this article comes from is machine-written, openly. The claims in it were checked one at a time against primary sources, and a practising doctor reviewed the results and carries responsibility for what got printed. That is not the same as a human reading every reference, and I will not describe it as though it were.

The check earned its keep. My draft stated that Australia’s medicines regulator had mandated a liver-injury warning on ashwagandha products after its advisory of 22 February 2024. That is wrong, and the error was mine. The advisory is real and the reports behind it are real — twelve reports of liver problems received up to 5 February 2024, seven of them carrying enough information to suggest the herb may have caused the injury, four cases requiring hospitalisation. But the regulator’s own sentence is conditional: if further substantiating evidence arises, regulatory action will be considered, and such action could include warning statements on product labels. No warning is required today. I had turned a possibility into a rule. Three independent checkers caught it, and it was rewritten before publication.

I mention it because it is the whole argument. A machine wrote that sentence with total confidence and it was false. The fix was not a human author. The fix was a check.

What I cannot tell you

Whether any specific book was machine-written. The checks above sort probabilities; they do not settle cases, and a careful human writer with a thin online presence will fail several of them. Detection software will not rescue you either — an evaluation of fourteen detection systems found them neither accurate nor reliable, and prone to calling machine text human, which is the failure direction that hurts a buyer most.

I also do not know how much of the shelf is AI-assisted rather than AI-generated, and the distinction matters more than the discourse allows. A researcher who drafts with a model and then verifies every line has produced something better than a rushed human manuscript with no sources. Amazon’s rules recognise that split, and require disclosure only for the generated kind — but that disclosure goes to Amazon, and the policy nowhere says it reaches you. So the label you would most want is the one you will never see on the page.

Which leaves the check. Look for whoever is accountable. If nobody is, put it down.