AI detectors flag 98.7% of human-written text in study

A study published in June found that different artificial intelligence detection tools often reach conflicting conclusions when analyzing the same piece of writing.

Researchers tested 160 academic papers across four categories: entirely human-written, entirely AI-written, a mix of both, and AI text rewritten by humans. Among the four detectors evaluated, Pangram showed the highest consistency. It correctly identified AI usage in 37 of 40 papers containing combined text and in 37 of 40 papers where humans rewrote AI content. Max Spero, CEO of Pangram, acknowledged in an emailed statement that the tool can make errors, especially when the text sample is short.

The rise of these tools follows a surge in submissions to publishers. Neil Clarke, publisher of Clarkesworld, noted that the magazine received roughly 1,100 submissions per month in 2021. By February 2023, that number had doubled to approximately 2,200 after AI agents like ChatGPT became widely available. Clarke temporarily closed the submission portal and estimates his personal workload increased by about 30%. He stated that this influx introduced an element of mistrust into the editing process.

Sally McGraw, a professional nonfiction writer with more than 20 years of experience, recently faced a false positive. After submitting a writing sample for a book proposal job, she received an email stating that multiple detection tools flagged her work as machine-generated. McGraw said she had written the text herself and was angry at the dismissal. Although the client confirmed the screening, the client said experience and fit were the primary reasons for hiring a different writer. McGraw was paid for the samples but believes the detector results cost her the full proposal opportunity.

Chris Callison-Burch, a professor of computer and information science at the University of Pennsylvania, explained that detectors analyze how predictable a passage is. He noted that language models often find text from their training data unusually predictable, which can cause false positives for previously published work. In late 2022, science blogger Janelle Shane tested excerpts from her 2019 book, You Look Like a Thing and I Love You. One detector reported a 98.7% confidence level that the passage was machine-generated. Shane called the result a clear and glaring false positive.

Detectors look for specific patterns in word choice, punctuation, and sentence structure. Callison-Burch said that common human writing habits, such as using the word delve, frequent em dashes, or listing ideas in threes, can trigger an AI-positive result. He added that opening sentences with adverbs like notably or crucially, or immediately answering a posed question, also contribute to these flags. The tools weigh these patterns across the entire piece, meaning a single instance does not automatically indicate machine generation.

The Authors Guild stated in May that the widespread use of AI detectors creates an environment reminiscent of witch hunts. The organization warned that a false positive can cost an author their contract and reputation, while a false negative can erode reader trust in the industry. Clarke said he does not personally use detectors to screen manuscripts but believes they can help overwhelmed editors if the findings are verified. He cautioned against turning suspicions into accusations, noting that submissions should end in rejection, not an accusation.

McGraw said the fear of future accusations has altered her writing style. She now avoids certain words and sentence constructions and occasionally omits commas to prove human authorship. She expressed regret at losing her use of the em dash, a style choice she learned in journalism school and used for decades. She described the pressure as ironic, noting that large language models learn from human prose, yet writers feel compelled to make their work look less like AI. McGraw said she is bracing for the moment when a high-stakes assignment depends on her ability to pass a detection screen.