Can anyone actually tell if a book was written by AI?

The claim is easy to state. A book written by AI gives itself away, because it reads like AI. No single originator is identified in the record. The claim met a live dispute in late July 2026, when a crime novel sold in a 14-way auction in the United States was withdrawn by its primary agent over suspected AI use, although the basis for that suspicion has not been reported.¹⁰

The controlled record points the other way. In peer-reviewed studies published between 2021 and 2026, readers sorting human writing from machine writing scored at chance or worse.¹,²,³,⁴ Detection tools tested in 2023 were found neither accurate nor reliable on short texts.⁵ None of those studies tested a full-length book.

What was confirmed about the withdrawn novel?

The Bookseller reported that the crime novel sold in the US for $2.4m and was withdrawn by its primary agent on Friday 31 July due to suspected AI use.¹⁰ Publishers Weekly identified the book as Call Me, I'll Hide the Body by Jerry Adedayo Falade.⁹ PW gives the withdrawal date as July 29 and describes offers reportedly topping $2 million.⁹ The two trade outlets give different dates, July 29 in PW and Friday 31 July in the Bookseller, and these may describe different steps in the withdrawal. Their money figures differ in form, $2.4m against offers "reportedly topping $2 million", and may not conflict.⁹,¹⁰ PW says it is unknown whether its figure includes foreign rights or film and TV adaptations. The Bookseller article was read only as far as its opening paragraph, so any breakdown it gives was not checked.

In an email from UK agent Sandy Hodgman that PW saw, she wrote that she and Europa 'could no longer substantiate' that the novel was entirely human-written. PW quotes only those words.⁹ PW wrote: "How exactly the book was judged to have been written by AI remains unknown."⁹ PW reported that some industry publications said they had run the manuscript through the detector Pangram themselves.⁹ Falade maintains he wrote the book himself.⁹

None of the sources reviewed gives a detector result, method or finding.⁹ The record shows a deal that collapsed and an allegation the author denies. It does not show how anyone concluded the prose was machine-made.

What did the newest study test?

Sydney Sears and Deena Skolnick Weisberg published "Bot or not: Can people tell the difference between stories written by a human or by an AI system?" in Judgment and Decision Making, a Cambridge University Press journal, on 5 August 2026.¹ The materials were three pairs of stories of about 1,000 words each. Each pair set a human-written story against a ChatGPT-4.0 counterpart.¹ Participants were lay US adults recruited on Prolific.¹

The authors report that "participants were no better than chance at differentiating between the human-written and AI-generated stories."¹ Participants also rated the AI stories higher for quality and found them more absorbing.¹ Self-reported expertise with AI correlated with correct identification. Expertise with literature did not.¹

The limits are plain. The studies were not preregistered.¹ The stories were short, and the readers were lay US adults, not professional editors.¹ This is a study of short fiction and lay readers. It is not a test of a manuscript.

Do other reader studies agree?

Clark and colleagues, at ACL in August 2021, asked non-expert evaluators to separate human writing from GPT-2 and GPT-3 output in stories, news articles and recipes.² Without training, evaluators distinguished GPT-3 text from human text "at random chance level."² Brief training raised accuracy to 55% at most. The gain was not significant across domains.²

Jakesch, Hancock and Naaman, in PNAS on 7 March 2023, tested Airbnb, dating and freelancer profiles written by people and by GPT-2 and GPT-3. Participants identified the source with 50 to 52% accuracy.³ This study offers the claim its only partial support. Crowdworkers rated AI text as repetitive more often than human text, 12.7% against 7.1%, and as nonsensical more often, 13.6% against 9.6%.³ A classifier built on those cues reached 58.8% accuracy.³ People's own judgments stayed near chance, and the authors say this was partly because they also relied on cues that did not track authorship. Those cues included grammatical errors, first-person pronouns and family topics.³ So a cue such as repetition can carry some signal in short GPT-2 and GPT-3 profiles. It is a weak one, and readers mix it with false cues.

Literary expertise did not close the gap, and AI expertise helped only somewhat.¹ Casal and Kessler asked 72 linguists to judge research abstracts written by people and by an early version of ChatGPT. The reviewers achieved an overall positive identification rate of 38.9%.⁴ The sample is small and comes from one field.⁴

Does detection software do better?

Weber-Wulff and colleagues tested 12 free tools plus Turnitin and PlagiarismCheck on about 54 short English documents. Their paper appeared on 25 December 2023.⁵ They concluded that the tools "are neither accurate nor reliable" and lean towards classifying text as human-written.⁵ The authors note that the two paid vendors knew the documents were part of a test.⁵

Liang and colleagues at Stanford tested seven detectors in Patterns on 10 July 2023. They used 91 TOEFL essays and 88 essays by US eighth-graders.⁶ On the TOEFL essays, all written by humans, the average false-positive rate was 61.3%.⁶ A simple prompt asking the model to elevate its essays with literary language cut detection of AI essays to near zero.⁶ The detectors wrongly flagged non-native English writers while classifying the US eighth-grade essays almost perfectly. A simple prompt also let ChatGPT-3.5 college application essays slip past them.⁶

OpenAI launched its own classifier on 31 January 2023. On its English challenge set, it reported a 26% true-positive rate and a 9% false-positive rate. It described the tool as very unreliable on text below 1,000 characters.⁷ On 20 July 2023 OpenAI withdrew the classifier "due to its low rate of accuracy."⁷

The strongest accuracy figures come from a vendor. Pangram, in a 15 September 2025 blog post, put its overall false-positive rate at about 1 in 10,000. It gave rates of 0.009% for short stories and 0.003% for books.⁸ These are self-reported internal benchmarks, not independent or peer-reviewed results. The post concedes weaker performance on short text and some domains.⁸ It reports no rate at which the tool catches AI-written books.⁸ A false-positive rate measures how rarely a tool accuses a human. It does not measure how often it catches a machine.

What does the industry rely on instead?

The question has legal weight. The US Copyright Office stated in January 2025 that copyright "does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements."¹¹ The report is policy guidance, not a court ruling, and it does not address detection.¹¹

Amazon's Kindle Direct Publishing requires authors to inform it of AI-generated text, images or translations when they publish a book.¹² It distinguishes AI-generated content, which must be disclosed, from AI-assisted content, which need not be. The page does not say how Amazon verifies either.¹² The Authors Guild licenses a certification mark, Human Authored, to indicate that a book was written by a human and not generated by AI.¹³ Its usage guidelines do not explain how claims are verified.¹³

Amazon's policy requires author disclosure of AI-generated content. The page also says Amazon enforces its content guidelines with machine learning, automation and human reviewers, but it does not say how AI-generated text is identified.¹² The Authors Guild's guidelines do not say how claims are verified, and the Copyright Office report does not address detection.¹³,¹¹ The pages reviewed do not say how any of these schemes check authorship. The Amazon page mentions human reviewers for content guidelines in general, and the Authors Guild FAQ was not opened.

What has not been tested?

Every study above used short texts: stories, profiles, abstracts, essays and recipes.¹,²,³,⁴,⁵,⁶ Most used models from 2023 or earlier.²,³,⁴,⁵,⁶ The newest study used ChatGPT-4.0 stories.¹ None tested a full novel. A novel may give a reader more to go on, or it may not. The record does not say.