--- title: "AI Writing Detectors Are Still Flagging Human Writers" description: "Writers tell NPR that AI detectors flagged their own work as machine-made. Here's what triggers a false positive and what freelancers can do about it." url: https://publishmyblog.com/creators/ai-detectors-flag-human-writers/ section: "Creators" author: "Saeed Shahriar" published: 2026-11-02T06:20:00.000Z --- # AI Writing Detectors Are Still Flagging Human Writers A ghostwriter with 20 years behind her now leaves out commas so her work looks less like AI. Ghostwriter Sally McGraw sent a client a writing sample, and **several AI detection tools flagged it as machine-made**. She'd written it all herself, [she told NPR on October 31](https://www.npr.org/2026/10/31/nx-s1-5963520/ai-writing-detectors-pangram). Do you write for clients, pitch guest posts or submit to magazines? Then **a client can run your work through a detector too**, and you may never see the score. ## What happened to the ghostwriter McGraw has written nonfiction for more than 20 years. **She was trying to win a book proposal job** when the client emailed her the detector results. The client confirmed to NPR that she'd screened the work for AI, but said the score wasn't what decided it and that she hired another writer for experience and fit. McGraw told the client she'd written it. Then she let it go. She was paid for her samples, but **she still believes the detector result cost her the full proposal**. So she changed how she writes, and she described the changes to NPR: - She avoids certain words and sentence constructions. - She sometimes leaves out commas she used to include, hoping the inconsistency proves a human wrote the text. - She misses the em dash, which she learned in journalism school and used for decades. ## Why editors use detectors at all Volume. Neil Clarke runs the sci-fi magazine Clarkesworld, and he told NPR it got **roughly 1,100 submissions a month in 2021**. By February 2023, after ChatGPT, Claude and Gemini became widely available, that number had doubled, and he had to close the magazine's submission portal for a while. He doesn't screen manuscripts with detectors himself. Still, he thinks some tools could help overwhelmed editors, **as long as someone checks what the tool says**. So how good are the tools you're up against? Mixed. NPR points to a June study that ran four detectors over 160 academic papers. Pangram did best. **It was the most consistent of the four**, catching AI use in 37 of 40 mixed papers and 37 of 40 papers a human had rewritten. Its CEO, Max Spero, told NPR it can still make mistakes, particularly when you give it little text. ## The Authors Guild test [The Authors Guild ran its own test in May](https://authorsguild.org/news/can-ai-detectors-be-trusted/), feeding **ten of its articles from 2022 or earlier**, before generative AI was widely available, into five of the top detectors. - Pangram: 0% AI on all ten articles. - Originality.ai: 0% on eight, 1% on the other two. - Grammarly: 0% on eight, then 7% and 9%. - ZeroGPT: anywhere from 5% to 76%, including 66% for the Guild's own obituary of Joan Didion, which is about as human as writing gets. - Sidekicker.ai: every article flagged as mostly AI, from 71% to 100%. Two hit 100%. ## What sets off a false positive Chris Callison-Burch, a University of Pennsylvania professor, told NPR how many detectors work. **They ask a language model how predictable a passage is.** But models find their own training text unusually predictable, so your published work can get flagged as AI. Science blogger Janelle Shane saw that in late 2022. A detector gave a passage from her 2019 book **almost 99% confidence of being machine-made**. She'd written it herself. Your unpublished draft avoids that trap. But your style can still trip a detector. **Callison-Burch named the patterns that can push a score up:** - Certain words that people now link with chatbots, and frequent em dashes. - Ideas grouped in threes, and sentences that open on certain adverbs. - Asking a question and then answering it straight away. He added that detectors weigh patterns across your whole piece, so one em dash alone doesn't mean a bot wrote your draft. The Guild's reason is the one that'll sting. **The more refined your style, the more your writing can look like AI**, since models learned from polished human prose. ## Why it matters A client may run your work through a detector and never tell you the score. **My advice depends on which side of the desk you sit on:** - Freelancing? Keep your drafts, your notes and your version history, so you can show a client exactly how your piece came together if anyone ever questions a detector score. - When you hire writers, treat a score from one tool as a lead to check, and ask the writer before you accuse anyone. - Never paste a score from one tool into your rejection email. Run a guest post program, like we do? You're the editor here. I'd follow Clarke, who told NPR that "submissions should end in rejection, not an accusation." Of course, our own house style bans em dashes too. That's a style choice. **A 66% score on a Didion obituary shows why it's a bad reason to call you a cheat.** Writing for money is hard enough **before anyone runs your work through a detector**. Want to [earn from your writing without depending on one client](https://publishmyblog.com/guides/how-to-make-money-blogging/)? ## Common Questions **Can AI writing detectors flag human-written text as AI-generated?** Yes. A ghostwriter told NPR in October 2026 that several detectors flagged her original sample, and an Authors Guild test found Sidekicker.ai scored ten pre-2023 human articles between 71% and 100% AI, while Pangram scored the same ten at 0%. **Which AI detector was most accurate in the 2026 tests?** Pangram. In the Authors Guild's May 2026 test it scored ten human-written articles at 0% AI, and in a June study of 160 academic papers it was the most consistent of the four detectors tested. **What writing habits make AI detectors flag a piece?** According to University of Pennsylvania professor Chris Callison-Burch, frequent em dashes, ideas grouped in threes and a question answered straight away can raise a score. Detectors weigh patterns across the whole piece, so one em dash alone doesn't mean a bot wrote the draft you sent to a client or an editor. ## Sources - [Your writing was flagged as AI. What does that actually mean?](https://www.npr.org/2026/10/31/nx-s1-5963520/ai-writing-detectors-pangram), NPR - [Can AI Detectors Be Trusted? The Authors Guild Put Five of Them to the Test](https://authorsguild.org/news/can-ai-detectors-be-trusted/), The Authors Guild