1,100 monthly submissions: AI tools strain local writers and publishers

Ghostwriter Sally McGraw received an email this month stating that multiple artificial intelligence detection tools had flagged her writing sample as machine-generated. She had not used AI to create the text. McGraw has worked as a professional nonfiction writer for more than 20 years. The client confirmed in an email to NPR that she screened the work for AI. However, the client said the results were not the deciding factor in hiring a different writer. The client cited experience and fit as the primary reasons for the decision. McGraw was paid for her samples but believes the detector results cost her the full book proposal job.

Publishers face a growing volume of AI-generated submissions. Neil Clarke, publisher of Clarkesworld magazine, reported that the publication received roughly 1,100 submissions per month in 2021. By February 2023, that number had doubled. The increase followed the widespread availability of AI agents like ChatGPT, Claude, and Gemini. Clarke temporarily closed the submission portal to manage the load. He estimates his own workload has increased by about 30 percent. Clarke said the influx has introduced an element of mistrust into the editing process.

Researchers published a study in June in the International Journal for Educational Integrity. The team compared the accuracy of four widely used AI detectors. They tested 160 academic papers across four categories: entirely human-written, entirely AI-written, a combination of both, and AI material rewritten by a human. Among the four tools, Pangram showed the most consistent accuracy. It correctly identified AI use in 37 of 40 papers with combined text. It also correctly identified AI use in 37 of 40 papers where a human rewrote AI-generated text.

Max Spero, CEO of Pangram, said in an emailed statement that the detector can make mistakes. He noted that errors are more likely when there is not much writing to evaluate. Chris Callison-Burch, a professor of computer and information science at the University of Pennsylvania, explained that many detectors work by asking a large language model how predictable a passage is. Language models often find text from their training data unusually predictable. This can cause false positives for previously published writing.

Science blogger Janelle Shane tested excerpts from her 2019 book, You Look Like a Thing and I Love You, in late 2022. One detector reported a 98.7 percent confidence that a passage was machine-generated. Shane called the result a clear false positive. She wrote the text herself. Callison-Burch said that unpublished drafts can still be falsely flagged based on writing style patterns. Detectors look for clues in word choice, punctuation, and sentence structure.

Specific patterns can trigger an AI-positive result. These include the word delve, frequent em dashes, and ideas grouped in lists of threes. Opening sentences with adverbs such as notably or crucially can also trigger flags. So can posing a question immediately followed by answering it. Callison-Burch noted that these are familiar habits for human writers. Detectors weigh patterns across a whole piece, so a single em dash or word does not necessarily indicate bot use.

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

McGraw said 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 a human wrote the text. She said she misses her em dash, a tool she learned in journalism school and used for decades. She sees the irony in trying to make her prose look less like the human writing that large language models learn from.