Detectors flag 98.7% of human text

A 2019 book by science blogger Janelle Shane was flagged as 98.7% likely to be AI-written by one detection tool. Shane, who authored the text entirely by hand, described the result as a glaring false positive. Her experience underscores a growing problem in publishing: AI detectors frequently mistake human work for machine output, leaving writers to defend their own prose.

Ghostwriter Sally McGraw encountered a similar issue recently. After submitting a writing sample for a book proposal, she received an email stating that multiple tools had identified her work as machine-generated. McGraw, a nonfiction writer with more than 20 years of experience, said she was angry at the dismissal. She assured the client she had written the text herself but decided not to press the matter, viewing the job as minor. The client later confirmed via email that the detector results were not the sole reason for hiring a different writer, citing experience and fit instead. McGraw believes the flags cost her the opportunity to write the full proposal.

Editors are turning to these tools to manage a surge in submissions. Neil Clarke, publisher of the science-fiction magazine Clarkesworld, reports that the publication received roughly 1,100 submissions per month in 2021. By February 2023, that number had doubled following the wide availability of AI models like ChatGPT. Clarke said his own workload increased by about 30%, forcing the magazine to temporarily close its submission portal. He noted that the influx has introduced an element of mistrust into the editing process.

Research indicates that different detectors often reach conflicting conclusions about the same piece of writing. A study published in June in the International Journal for Educational Integrity tested four widely used detectors against 160 academic papers. The papers fell into four categories: entirely human, entirely AI, a mix of both, and human-rewritten AI. Pangram, a detection tool, was the most consistently accurate, correctly identifying AI usage in 37 of 40 papers in both the mixed and human-rewritten categories.

Chris Callison-Burch, a professor of computer and information science at the University of Pennsylvania, explains that many detectors work by asking a large language model how predictable a passage is. The catch is that these models often find text from their training data unusually predictable. This can cause previously published human writing to be flagged. For unpublished drafts, flags may trigger based on specific stylistic patterns. Callison-Burch noted that words like "delve," frequent use of em dashes, and lists of three ideas can trigger an AI-positive result. He added that detectors weigh patterns across a whole piece, so a single stylistic choice does not automatically indicate machine authorship.

The Authors Guild warned in May that the use of AI detectors creates an environment reminiscent of witch hunts. The organization stated that a false positive can cost an author a contract and reputation, while a false negative can erode reader trust. Clarke, who does not personally use detectors to screen manuscripts, believes certain tools can help overwhelmed editors manage volume if they verify the findings. He cautioned against turning suspicions into accusations, saying 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 structures and occasionally omits commas she used to include, hoping the inconsistency will prove a human wrote the text. She misses using em dashes, a tool she learned in journalism school. McGraw noted the irony that large language models learn from human prose, yet she feels pressure to make her writing look less like AI. She is bracing for a future assignment where the stakes are high and the outcome will make or break her career.