Key Takeaways
- Two high‑profile publishing deals collapsed after rumors surfaced that the manuscripts were written with AI assistance, prompting agencies and publishers to withdraw contracts.
- AI‑detection tools such as Pangram and GPTZero have demonstrated high false‑positive rates, flagging texts as AI‑generated when they are not—examples include a papal encyclical, a Wall Street Journal column, and even the U.S. Constitution.
- The core limitation of these detectors is that they measure statistical predictability (perplexity) rather than actual authorship; they cannot determine who contributed ideas, judgment, or creative direction.
- Publishing has long tolerated collaborative writing practices—ghostwriters, co‑authors, heavy editorial rewriting—so the “AI or not” binary test is misplaced; the industry should focus on the author’s intellectual contribution instead.
- Copyright law already reflects this shift: only human‑contributed portions of AI‑assisted works are eligible for protection, underscoring that value lies in creativity, not mere word‑production.
- As AI makes polished prose easier to generate, publishers will need new evaluative frameworks that assess originality, experience, and point of view rather than relying on flawed detection scores.
The Falcon Deal and Its Sudden Unraveling
In late July, Nigerian‑British author Jerry Falade experienced what many writers dream of: a six‑hour bidding war that saw fourteen publishers vie for his crime novel “Call Me, I’ll Hide the Body.” Minotaur Books, an imprint of Macmillan US, ultimately secured the rights for a reported $2 million to $2.4 million. Falade recounted to The Guardian how quickly the euphoria turned sour: “Within six hours, I had been dropped by my agency, deals were paused, and press releases were being issued.” The catalyst was a rumor that he had used artificial intelligence to compose the manuscript.
His agent, Marc Gerald of Europa Content, initially found Falade’s explanation “clear and credible” and chose not to run the text through AI‑detection software. However, after a July 29 meeting, Gerald noted that “aspects of [Falade’s] story changed,” prompting the agency to withdraw the book, citing the need for “total faith” in the manuscript’s origins. The abrupt reversal illustrates how swiftly trust can evaporate when AI authorship is suspected, even without concrete proof.
The “Shy Girl” Controversy and the Role of Online Vigilantes
Earlier in the year, a similar scenario unfolded with Mia Ballard’s horror novel “Shy Girl.” For weeks, readers on Goodreads and Reddit had flagged passages they believed bore the hallmarks of AI‑generated prose. A widely shared Reddit thread attracted hundreds of comments, while a January YouTube video dissecting the writing amassed more than 2 million views.
Detection company Pangram’s founder, Max Spero, ran his own analysis and reported on X (formerly Twitter) that 78 % of the book was AI‑generated. Hachette responded by cancelling the U.S. release and discontinuing the U.K. edition, referencing its own “lengthy investigation in recent weeks.” Ballard denied personal use of AI, telling The New York Times that an editor she had hired for the earlier self‑published version was responsible. The episode underscores how public scrutiny, amplified by social media, can precipitate publisher action long before any formal adjudication.
Why Detectors Fail: The Perplexity Problem
Both controversies hinge on the output of AI‑detection tools, yet these instruments suffer from a fundamental flaw: they measure perplexity, a statistic that gauges how predictable a word sequence is against a model’s training data. Formal, heavily codified language—such as constitutions, legal contracts, or scripture—naturally scores low on perplexity because language models have ingested vast amounts of similar text. Consequently, detectors mistakenly label such passages as machine‑made.
As evidence, Pangram’s own blog post noted that the tool flagged passages of Pope Leo XIV’s 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 further, calling Pangram a “defamation machine” after it mislabeled three of the paper’s own opinion columns as AI‑generated.
The unreliability is not new. In 2023, GPTZero famously flagged the U.S. Constitution as AI‑generated, a conclusion that would require James Madison to have been a time traveler. GPTZero founder Edward Tian admitted to Ars Technica that the Constitution is “fed repeatedly into the training data of many large language models,” explaining why his tool misreads it. OpenAI shut down its own detector that year, citing a “low rate of accuracy,” according to TechCrunch. These examples demonstrate that reliance on perplexity‑based scores can produce absurd false positives, rendering the tools unsuitable for high‑stakes publishing decisions.
Authorship Has Never Been a Pure Solo Act
The publishing industry’s panic over AI authorship overlooks a long‑standing reality: books are rarely the product of a single mind typing every word. Ghostwriters have penned memoirs under celebrities’ names for a century. Co‑authors develop novels from another’s outline and voice. Editors routinely rewrite entire chapters before a manuscript hits shelves. Yet, we do not pull memoirs from circulation because a ghostwriter typed the sentences; the name on the cover signifies the idea, point of view, and judgment behind the work, not a verbatim word count.
AI merely adds another tool to this collaborative ecosystem. As the technology makes polished prose easier to generate, the value of authorship shifts toward the conceptual contribution—the experience, insight, and creative judgment that shape a narrative—rather than the mechanical act of stringing words together. Detectors, which cannot assess these qualities, are therefore ill‑suited to answer the question publishing truly faces.
Data Shows the Market Is Already Moving
A recent study led by Stony Brook University’s Tuhin Chakrabarty and Columbia Law School’s Jane Ginsburg, with co‑authors Xinyue Liu and Paramveer Dhillon, examined 14,419 self‑published genre‑fiction titles sold on Amazon between 2023 and 2026. The researchers ran full‑text AI detection across the corpus and found a telling trend:
- Books with no detected AI text saw their share of total sales fall from nearly 100 % in early 2023 to about 60 % by mid‑2026.
- Revenue per book dropped in seven of eight genres over the same period.
- Conversely, titles flagged for substantial AI content captured a rising share of the market’s scarcest real estate, climbing from near zero to 31 % of new Top 25 bestseller slots.
These numbers suggest that readers and algorithms are increasingly rewarding works that leverage AI‑assisted efficiency, even as traditional publishers cling to outdated notions of “pure” authorship.
Copyright Law Points the Way Forward
Legal precedent already reflects the evolving understanding of authorship. In March, the Supreme Court declined to revisit Thaler v. Perlmutter, leaving intact the rule that works created solely by AI cannot be registered for copyright, while human contributions within an AI‑assisted work remain eligible. Law firm Mayer Brown notes that this distinction acknowledges that the protectable element is the human creative input—selection, arrangement, editing, or thematic development—not the raw output of a language model.
If copyright law protects only the human component, publishers must adopt a parallel standard: evaluate what the author contributed rather than merely asking whether a detector flags the text as AI‑generated. This approach aligns with industry practices that have long celebrated collaborative creation while still attributing responsibility to the named author.
Moving Beyond Detection: A New Framework for Publishing
The repeated missteps of AI detectors reveal a chasm between what the tools can measure (statistical predictability) and what the industry truly needs to know (the origin of ideas and artistic judgment). As AI continues to lower the barrier to producing fluent prose, publishing must develop evaluative criteria that focus on originality of concept, depth of experience, and strength of voice.
Potential steps include:
- Transparent author disclosures about any AI tools used, similar to conflict‑of‑interest statements in academic publishing.
- Editorial attestation that the named author provided substantive creative direction, supported by outlines, notes, or interview records.
- Reader‑focused metrics—such as narrative cohesion, emotional resonance, and thematic novelty—rather than reliance on binary detection scores.
- Industry‑wide standards for what constitutes “meaningful human contribution” in AI‑assisted works, possibly modeled on existing copyright guidelines.
By shifting the conversation from “Did AI write this?” to “What did the author bring to the work?”, publishing can preserve trust, reward genuine creativity, and responsibly integrate emerging technologies without falling prey to the false promises of imperfect detectors.