Writers Cope With AI Detector Errors

Professional nonfiction writer Sally McGraw recently lost a book proposal job after an AI detection tool flagged her writing sample as machine-generated. McGraw, who has worked in the field for more than 20 years, wrote the text herself. The client confirmed they used the screens but stated that experience and fit were the primary reasons for hiring a different writer. McGraw says the detector results still cost her the opportunity to write the full proposal.

Publishers are using these tools to manage a sharp rise in submissions. Neil Clarke, publisher and editor of Clarkesworld magazine, reported that monthly submissions doubled from roughly 1,100 in 2021 to double that number by February 2023. This increase followed the wide availability of AI agents like ChatGPT, Claude, and Gemini. Clarke said the surge forced him to temporarily close his submission portal. He estimates his own workload has increased by about 30% since the shift began.

Detection tools often disagree with one another. A study published in June in the International Journal for Educational Integrity tested four widely used detectors against 160 academic papers. The papers included entirely human-written work, entirely AI-written work, and mixed samples. The study found that Pangram was the most consistently accurate detector. It correctly identified AI usage in 37 of 40 papers containing combined text and in 37 of 40 papers where humans rewrote AI-generated text. However, Pangram CEO Max Spero acknowledged in an emailed statement that the company’s detector can make mistakes, particularly when there is not much writing to evaluate.

False positives can occur for technical reasons. Chris Callison-Burch, a computer and information science professor at the University of Pennsylvania, explained that many detectors ask a large language model how predictable a passage is. He noted that these models often find text from their training data unusually predictable. This issue affected science blogger Janelle Shane in late 2022. One detector reported a 98.7% confidence level that a passage from her 2019 book, You Look Like a Thing and I Love You, was machine-generated. Shane called the result a glaring false positive because she wrote the text herself.

Specific writing habits can trigger AI-positive results. Callison-Burch said that words like delve, frequent em dashes, and lists of three ideas can flag a text. He added that opening sentences with adverbs such as notably or crucially can also cause a flag. These are common habits for human writers. Detectors weigh patterns across a whole piece, so a single em dash or specific word does not automatically indicate a bot was used.

The Authors Guild stated in May that the use of AI detectors creates an environment reminiscent of witch hunts. The organization warned that a false positive can cost an author a publishing contract and reputation. A false negative, where AI-generated work slips through undetected, can cause readers to lose trust in authors and the industry. Clarke said he does not personally use AI detectors to screen manuscripts. He believes certain tools could help overwhelmed editors manage volume if they check the findings, but he cautioned against turning suspicions into accusations.

McGraw says the fear of another accusation has changed how she writes. She now avoids certain words and sentence constructions. She occasionally leaves out commas she used to include, hoping the inconsistency will help prove that a human wrote the text. She noted that large language models learn from human prose, creating pressure to make her writing look less like AI. She described the feeling as trying to fool a machine into thinking she is not a machine.

McGraw said she is bracing for a future assignment where the outcome is critical. She wants to ensure she is prepared if the work is on the line.