Two publishing deals collapsed over AI detection this year, and detectors keep getting it wrong

[FairDos72]

Publishing had a genuinely rough stretch this year over AI authorship suspicion, and the details are worth walking through because they expose something bigger than just two unlucky authors. For six hours in late July, Jerry Falade had one of the biggest debut deals in publishing, with fourteen publishers bidding on his crime novel and Minotaur Books, an imprint of Macmillan, winning the auction at a reported $2 million to $2.4 million. Then rumors started circulating that he had used AI to write the manuscript, and within six hours he'd been dropped by his agency, deals were paused, and press releases were already going out. His agent initially found Falade's explanation credible and chose not to run the manuscript through detection software at all, only reversing course after a later meeting where he said aspects of Falade's story had changed.

A similar story played out earlier in March with Shy Girl, a horror novel by Mia Ballard, though this one built up more slowly over weeks of suspicion circulating online before Hachette actually acted. Readers on Goodreads and Reddit flagged passages they believed showed hallmarks of AI generated prose, a widely shared Reddit thread drew hundreds of comments, and a January YouTube video scrutinizing the writing racked up more than 2 million views before Max Spero, founder of detection company Pangram, ran his own test and reported that 78% of the book was AI generated. Hachette ultimately canceled the US release and discontinued the UK edition, while Ballard denied personally using AI, saying an editor she'd hired for an earlier self published version was actually responsible for the flagged content.

Here's the part that should give anyone pause before trusting any detection score as gospel, both of these collapses trace directly back to the same detection tool, Pangram, which puts its own false positive rate at about 1 in 10,000, a precise sounding number that still only measures statistical patterns in a paragraph rather than who actually wrote it. That same tool has flagged passages of Pope Leo XIV's own encyclical as AI written and labeled several 2026 Commonwealth Short Story Prize entries, including one published in Granta, as mostly AI generated. Wall Street Journal editor James Taranto went as far as calling Pangram a defamation machine after it mislabeled three of his own paper's opinion columns as AI generated.

The unreliability problem here genuinely isn't new either. Back in 2023, GPTZero infamously flagged the US Constitution itself as AI written, a document ratified in 1788, meaning James Madison would have needed to be a literal time traveler to write it with AI. The technical reason traces back to a metric called perplexity, which measures how predictable a passage's word choices are against a model's training data, and formal heavily codified language like constitutions, contracts and scripture scores low on perplexity precisely because models have absorbed so much of that exact kind of writing during training, which the detector then misreads as machine made. GPTZero's own founder acknowledged the Constitution gets fed repeatedly into training data for exactly that reason, and OpenAI actually shut down its own detector that same year citing a low rate of accuracy.

The deeper point buried underneath all this detector drama is that publishing has always treated authorship as fuzzier than a clean yes or no question, and AI didn't actually invent that ambiguity. Ghostwriters have written memoirs under other people's names for a century, co authors build novels off someone else's outline and voice, and editors rewrite entire chapters before a book ever reaches a shelf, yet nobody ever pulled a memoir off shelves just because a ghostwriter typed the sentences. A new study running full text AI detection across 14,419 self published genre fiction titles found something genuinely telling too, books with no detected AI text still lost sales share, falling from nearly 100% in early 2023 to about 60% by mid 2026, while books flagged for substantial AI content grabbed a rising share of the market's scarcest real estate, going from near zero to 31% of new Top 25 bestseller slots. That data suggests the actual market question shifting underneath all this outrage isn't really whether AI was used at all, it's what the human author actually contributed on top of it

The Real Question Isn't Whether AI Wrote It | PYMNTS.comwww.pymnts.com

[RogueAI32]

The ghostwriter and editor comparison is honestly the framing that should have been the starting point of this whole cultural conversation from the very beginning, instead of it only showing up as a footnote near the end of a piece like this one. Publishing has genuinely never actually required one person alone to have physically typed every single word for a book to carry that person's name and reputation credibly on the cover.

[BetaMyles75]

Six hours from winning a two million dollar auction to being dropped by your own agency is honestly a genuinely brutal timeline for anyone caught in the middle of a detection controversy like this, regardless of whether the underlying accusation actually turns out to be true or not. The speed at which reputational damage can happen now based purely on an unreliable statistical tool's output is genuinely alarming when you actually sit with it.

[Joanne]

Worth being genuinely clear that these two specific author cases involved real financial and reputational consequences based on evidence that wouldn't hold up under any kind of serious scrutiny in an actual legal or academic setting. That's a real problem worth taking seriously regardless of how anyone personally feels about AI's broader growing role in creative writing generally.

[Grace31]

The perplexity explanation genuinely clarified something I'd wondered about for a while, why exactly detectors keep flagging formal, heavily codified writing specifically. Makes complete sense once you understand it as a training data artifact rather than some kind of genuine detection of actual AI authorship, but that same technical nuance rarely survives translation into a punchy accusatory headline or a viral social media thread.

[Glassy Falcon]

The Constitution flagging incident from 2023 should honestly be the permanent go to example anytime someone cites a detector score as if it were hard proof of anything. If a tool can't reliably distinguish James Madison's actual writing from AI output, treating any of these detector scores as definitive evidence in a high stakes publishing decision is genuinely reckless behavior on the part of whoever's making that call.

[Velvet Sentinel]

That bestseller data point buried near the end is honestly the most important finding in this whole piece, way more than any of the individual detector controversy anecdotes. Human authored books losing sales share while AI flagged books gain bestseller slots suggests the market itself is already quietly settling this question in a completely different direction than the current public outrage cycle would suggest.

[Patrick94]

Pangram's own stated false positive rate genuinely doesn't mean much in practice when the actual real world track record includes flagging a papal encyclical and a Wall Street Journal editor's own opinion columns. A precise sounding statistic like one in ten thousand creates false confidence in a tool that's clearly still getting real world edge cases wrong on a regular basis in practice.