A ghostwriter's own writing got flagged as AI. Here's what detectors can and can't tell you

On October 31, NPR told the story of ghostwriter Sally McGraw. She applied to develop a book proposal, sent a writing sample, and got an email back saying several AI detection tools had flagged it as machine-generated. "I was floored because I had not used AI to generate the text," she told NPR. "I'd written it all myself."

The client confirmed to NPR that she had screened the sample for AI, but said the results were not what decided it: she hired a different writer for experience and fit. McGraw was paid for her samples and believes the detector results cost her the full job.

What the research says. In June, researchers at Vrije Universiteit Brussel published a study in the International Journal for Educational Integrity that tested four popular detectors (GPTZero, Pangram, Copyleaks and Turnitin) on 160 academic papers with known answers: 40 fully human, 40 fully AI-written, 40 human papers with AI passages added, and 40 AI passages "humanised" with a prompt.

• Human papers: all four tools scored all 40 human papers as human. False positives were rare in this set.
• Fully AI papers: Turnitin, GPTZero and Copyleaks missed most or all of them. Pangram caught nearly all of them, with one case misclassified under the study's broader measure.
• Mixed and humanised papers: Pangram placed 37 of 40 in the expected range in each group. The other tools significantly underestimated the AI content.

The catch. Tools disagree with each other, a good result on one dataset does not cover every kind of writing, and a score does not explain itself. The study's authors conclude that detectors "can provide useful initial flags" but "should not be used as sole evidence in high-stakes decision-making."

If you hire writers or grade work: treat a flag as a reason to ask questions (drafts, notes, version history, a short conversation), not as a verdict. If you are flagged: keep your drafts and edit history, and ask which tool was used and what score it gave.

The ghost in our cartoon is a Halloween joke. The story and the study are real.

Sources: NPR (October 31, 2026); International Journal for Educational Integrity (June 2026). Not sponsored; no affiliate links.

MODEL Editors