AI Detectors Flag 98.7% of Human Writing as Machine Generated

A passage from science blogger Janelle Shane’s 2019 book received a 98.7% confidence score from one AI detector in late 2022. Shane, who wrote the text herself, described the result as a glaring false positive. The incident illustrates a growing friction point in the publishing industry as editors attempt to manage a surge in submissions using automated software.

Ghostwriter Sally McGraw recently faced a similar situation when a prospective client flagged her writing sample as machine-generated. McGraw, a professional nonfiction writer with more than 20 years of experience, stated she was angry that her work was dismissed. She assured the client she had written the proposal herself but decided not to press the issue for the minor job. The client confirmed via email that she had screened the work for AI but noted that experience and fit were the deciding factors in hiring a different writer. McGraw was paid for her samples but believes the detector results cost her the full book proposal.

Neil Clarke, publisher of the science-fiction magazine Clarkesworld, reports that the influx of AI-generated content has strained his editing staff. The publication received roughly 1,100 submissions per month in 2021. By February 2023, after AI agents like ChatGPT and Claude became widely available, that number had doubled. Clarke temporarily closed the submission portal and estimates his own workload increased by about 30%. He noted that the tools have introduced an element of mistrust into the editing process.

Researchers in a study published in June in the International Journal for Educational Integrity tested four widely used detectors against 160 academic papers. The papers were categorized as entirely human-written, entirely AI-written, a combination of the two, or AI material rewritten by a human. Pangram emerged as the most consistently accurate tool, correctly identifying AI use in 37 of 40 combined-text papers and 37 of 40 human-rewritten AI papers. Max Spero, CEO of Pangram, acknowledged that the detector can make mistakes, particularly when evaluating short pieces of writing.

Chris Callison-Burch, a professor of computer and information science at the University of Pennsylvania, explained that many detectors function by asking a large language model how predictable a passage is. Because these models are trained on human data, they often find previously published text unusually predictable. This can trigger false positives for established authors. Detectors analyze patterns in word choice, punctuation, and sentence structure. Elements such as the word "delve," frequent em dashes, and lists of three items can trigger an AI-positive result, even when a human wrote the text.

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 their publishing contract and reputation, while a false negative can cause readers to lose trust in the industry. Clarke said he does not personally use AI detectors to screen manuscripts but believes certain tools could help manage volume if editors verify the findings. He cautioned against turning suspicions into accusations, noting that submissions should end in rejection, not an accusation.

McGraw says the fear of future accusations has changed her writing habits. She now avoids certain words and sentence constructions and occasionally omits commas to demonstrate human inconsistency. She misses the em dash, a tool she learned in journalism school and has used for decades. McGraw noted the irony of large language models learning from human prose while she feels pressure to make her writing look less like AI. She is bracing for the time when a major assignment is on the line and the outcome could make or break her career.