Some authors worry that AI detection software will falsely report their work, potentially ruining their reputation or job opportunities before they have a chance to prove the detectors wrong.
karetoria/Getty Images
hide caption
karetoria/Getty Images
Ghostwriter Sally McGraw was hoping to land a job developing a book proposal for a potential client. But after submitting a writing sample, she received an unexpected response: an email telling her that several AI detection tools had flagged her work as machine-generated.
“I was stunned because I hadn’t used AI to generate the text,” McGraw said. “I wrote everything myself.”
After spending more than 20 years as a professional nonfiction writer, McGraw said she was angry that her work was dismissed as being written by AI.
“You know where you see a cartoon character turn red from their chin to the top of their head? I felt my ears start to burn,” she said. “I thought, what’s going on?”
McGraw said she assured the client that she had written the proposal herself, but decided not to press the issue because she considered the work “minor.”
Sally McGraw, a ghostwriter, said she changed her writing style after a potential client suggested her work was AI-generated.
Sally McGraw
hide caption
Sally McGraw
In an email to NPR, the client confirmed that she had reviewed McGraw’s work for AI. But she said results were not the determining factor in choosing to hire another writer, instead citing experience and fit. McGraw says she was paid for her writing samples, but believes the detector results cost her the opportunity to write the full book proposal.
As publishers struggle to keep up with the growing influx of AI-generated content, some are turning to detection tools to help them sift through the growing number of submissions. But these tools aren’t perfect: They are known to falsely flag human-written work as AI, which can leave writers with little idea about what triggered the result and how to prove otherwise.
“Everyone knows these testing devices are not 100 percent accurate,” McGraw said.
Editors are overwhelmed by AI-written submissions
Neil Clarke, publisher and editor of Clarkes Worldclaims that a flood of AI submissions to the science fiction magazine put pressure on the editorial team.
In 2021, the publication received approximately 1,100 submissions per month. By February 2023, after AI agents like ChatGPT, Claude, and Gemini began becoming widely available, this number had doubled, forcing Clarke to temporarily close its submission portal. He estimates his own workload has increased by about 30%.
“It introduced an element of distrust into the process,” he said.
When AI write detectors disagree
Although AI detectors have become more sophisticated in recent years, the companies behind the detectors acknowledge that the tools are not foolproof. Researchers have found that different AI detection tools can reach radically different conclusions about the same text.
In a study published in June in the International Journal for Educational Integrity, researchers compared the accuracy of four widely used detectors. They tested 160 academic articles comprising four different categories: entirely written by humans, entirely written by AI, a combination of the two, and AI material rewritten by a human.
Of the four detectors, Pangram was the most accurate, correctly identifying AI use in 37 of 40 articles containing the combined text and in 37 of 40 articles in which a human had rewritten the AI-generated text.
Max Spero, CEO of Pangram, said in an emailed statement that his company’s detector can make mistakes, especially when there aren’t many scriptures to evaluate.
Previously published writings can pose an additional problem, says Chris Callison-Burch, a professor of computer and information sciences at the University of Pennsylvania.
“Many AI detectors work by asking a large language model how predictable a passage is,” Callison-Burch said. “The problem is that language models also tend to find the text in their training data unusually predictable.”
This could explain what science blogger Janelle Shane encountered in late 2022 when she tested excerpts from her 2019 book on how AI works, You look like a thing and I love you. One detector indicated a 98.7% confidence level that a passage was machine-generated.
“This is a clear, blatant, flashing false positive,” Shane said. “I wrote this text myself.”
An unpublished draft such as McGraw’s writing sample would not face this particular problem, Callison-Burch said, but could still be falsely flagged based on patterns in writing style.
Detectors look for clues in word choice, punctuation and the way sentences are composed. The word “frisk,” frequent em-dashes and ideas grouped into lists of three can trigger a positive outcome for AI, he said. The same goes for opening sentences with adverbs such as “especially” or “crucially,” or asking a question immediately followed by an answer.
At the same time, these are familiar habits for human writers. Callison-Burch says the detectors weigh the patterns on an entire piece, so an em-dash or a particular word doesn’t necessarily mean a robot was used.
What are the challenges for writers
The Authors Guild, an organization that advocates for writers’ rights, said in May that the use of AI detectors “creates an environment reminiscent of witch hunts,” with writers worried about false accusations and how to defend themselves.
“A false positive can cost an author their publishing contract and their reputation, and a false negative, when AI-generated work goes unnoticed in a reputable publication, can cause readers to lose their trust in authors and the industry,” the Guild wrote.
Clarke, the science fiction magazine’s editor, falls somewhere in the middle. He says that while he doesn’t personally use AI detectors to filter incoming manuscripts, he thinks some detection tools could help overwhelmed editors. manage the volume, provided you check the detector results. But he cautions against turning suspicions into accusations that could damage a writer’s reputation.
“Submissions should result in rejection, not accusation,” he said.
McGraw says the fear of another accusation changed the way he wrote. She says she now avoids using certain words and sentence constructions, and sometimes forgets commas she included, hoping that this inconsistency will help prove that a human wrote the text.
“I miss my ca dash. It’s something I learned in journalism school and I’ve used it for decades,” she said.
The Negro also sees the irony. Large language models learn from human prose, but she feels pressure to make hers less AI-like.
“It’s like trying to trick a machine into thinking I’m not a machine,” McGraw said.
And she fears she’ll face another charge when a bigger opportunity presents itself.
“I’m preparing for the moment when it’s a mission that I’m really aiming for and the result makes or breaks me,” she…
Gn entert
