Here is a humanized summary of the article, retold in six paragraphs.
The moment Sally McGraw opened that email, the floor seemed to drop out from under her. A professional ghostwriter with more than two decades of nonfiction experience, she had submitted a writing sample for a promising job developing a book proposal. The response should have been a discussion about next steps, a question about her process, or even a polite rejection. Instead, she found herself staring at an accusation she never saw coming: multiple AI detection tools had flagged her work as machine-generated. McGraw was stunned, then furious. She had not used artificial intelligence to write a single word. She had crafted that sample herself, as she had done for hundreds of clients over a long and respected career. The accusation felt not only wrong, but deeply personal—as if her very humanity had been called into question. She could feel the heat rising from her chin to her scalp, a cartoonish surge of indignation that made her want to defend herself, to prove that her words were her own. Yet, despite her anger, she chose not to press the issue. The job, she reasoned, was relatively minor, and she had already assured the client of her authorship. But the sting remained, and it left her with a troubling sense of vulnerability: if her decades of writing could be dismissed by a flawed algorithm, what did that mean for her future?
The client later confirmed that she had, indeed, screened McGraw’s work for AI, although she insisted that the detector results were not the sole reason she hired a different writer. She cited experience and fit as the deciding factors. Still, McGraw believed the algorithm’s verdict cost her the opportunity, and she was paid for her samples but not for the full project. Her story is not an isolated incident. Across the publishing world, editors are drowning in a rising tide of AI-generated submissions, and many are turning to detection tools as a way to manage the flood. Neil Clarke, publisher and editor of the science-fiction magazine Clarkesworld, described how the number of submissions he receives has more than doubled since AI tools like ChatGPT became widely available. In 2021, he was seeing roughly 1,100 submissions a month; by early 2023, that number had jumped to over 2,000, forcing him to temporarily close the submission portal just to catch his breath. His workload has grown by about thirty percent, and the atmosphere has shifted. What was once a collaborative, trusting relationship between editor and writer now carries an undercurrent of suspicion. Clarke lamented that AI has introduced an element of mistrust into the process, making every submission a potential test of authenticity rather than a simple act of creative sharing.
The uncomfortable truth is that AI detectors are not the objective, infallible judges they are sometimes imagined to be. They are statistical tools, and they make mistakes. A study published in the International Journal for Educational Integrity tested four widely used detectors on 160 academic papers, including entirely human-written essays, fully AI-written texts, hybrid works, and AI-generated passages that had been rewritten by humans. The results were sobering. The best-performing detector, Pangram, correctly identified AI use in 37 out of 40 papers with mixed text, but even that success rate leaves room for error. Pangram’s CEO acknowledged that the tool can stumble, especially when it has only a short passage to evaluate. More troubling, however, is the inherent flaw in how many detectors function. They often rely on large language models to measure how predictable a piece of writing is. AI tends to produce highly predictable language, but so do well-crafted, fluent human texts. As University of Pennsylvania professor Chris Callison-Burch explained, language models often find text from their own training data unusually predictable—which means that a perfectly human-written essay that happens to mirror common patterns can be falsely flagged. This is exactly what happened to science blogger Janelle Shane when she tested excerpts from her own 2019 book about artificial intelligence. One detector declared with 98.7 percent confidence that her text was machine-generated. Shane, who had written every word herself, called it a glaring, blinking-light false positive. It was a perfect example of how these tools can turn against the very human voices they were meant to protect.
So what exactly triggers these false alarms? According to Callison-Burch, detectors look for subtle clues in word choice, punctuation, and sentence structure. A conspicuous use of the word “delve,” frequent em dashes, ideas organized in tidy lists of three, opening sentences that begin with adverbs like “notably” or “crucially,” and the habit of posing a question and immediately answering it can all nudge a text toward the machine-generated category. The irony is that these are also familiar habits of human writers. Journalists love an em dash. Essayists often structure arguments in threes. Adverbs are natural parts of everyday speech. Callison-Burch was quick to point out that detectors weigh patterns across an entire piece, so a single em dash or one distinctive word does not automatically mean a bot was involved. Still, for a writer like McGraw, whose unpublished draft would not have been in any training database, the false flag was likely based on stylistic patterns she had developed over decades of professional writing. Her style had become so polished, so fluent, so well-paced that it looked almost too perfect. In trying to write clearly and effectively, she had inadvertently written in a way that mimicked what an AI might produce. The very skills that made her a good writer were now making her look suspicious. That painful paradox lies at the heart of the problem: the things that make human writing good often overlap with the things that make AI writing detectable.
The stakes for writers are enormous. The Authors Guild, a leading advocacy organization for writers, has warned that the use of AI detectors creates an environment reminiscent of witch hunts. Writers now fear false accusations that could cost them publishing contracts, damage their reputations, and undermine years of hard work. At the same time, there is the opposite danger: if AI-generated work slips through undetected, readers may lose trust in the industry as a whole. Authors could be blamed for content they never wrote, or they could be unfairly suspected of using AI when they did not. Editors are caught in the middle, trying to protect their publications from being flooded with low-quality submissions while also trying to avoid harming innocent writers. Neil Clarke, for his part, says he does not personally use AI detectors to screen manuscripts. He believes some tools could help overloaded editors manage the sheer volume of submissions, but only if the results are carefully verified and never treated as proof of wrongdoing. He offered a simple, humane principle: submissions should end in rejection, not an accusation. A writer might be wrong for a particular magazine or project, but that does not mean they should be publicly shamed or silently blacklisted because of a flawed algorithm. The distinction between rejection and accusation is crucial. Rejection is about fit; accusation is about character. Clarke’s warning serves as a reminder that behind every submission is a human being, and that human beings deserve a measure of grace, especially when the technology meant to judge them is far from perfect.
For Sally McGraw, the emotional fallout of her experience has been lasting and tangible. She says she has changed the way she writes. She catches herself avoiding certain words and sentence constructions, second-guessing her natural rhythms, and even leaving out the occasional comma she would normally have included. She hopes that a little carefully chosen imperfection will help prove to any future detector that a human hand was behind the text. But she misses her em dash. That punctuation mark was something she learned in journalism school and had used proudly for decades. It was a small but meaningful part of her voice. Now she feels she has to suppress it, to flatten her style just to avoid suspicion. The absurdity is not lost on her. Large language models learned to write by reading human prose, and now she feels pressure to make her own writing look less human, or at least less smooth, to avoid being mistaken for a machine. It is, she said, like trying to fool a machine into thinking she is not a machine. She cannot shake the worry that another accusation will come, perhaps when she is chasing an assignment that truly matters, a project she desperately wants and needs. She is bracing for that moment, hoping she will have the strength to defend herself, but knowing that the burden of proof may be on her. In the meantime, she keeps writing, keeps polishing her sentences, and keeps her em dash tucked away, waiting for a time when being a human author no longer means having to prove her own existence.