Can Pangram’s AI Detector Be Trusted? False Accusations, Canceled Book Deals, Hidden Bias and the Startup Becoming Publishing’s Judge of Human Writing

An AI detector’s percentage score can now help decide whether a writer keeps a publishing deal, wins a prize, or faces a public accusation. In this episode of The Daily AI Chat, we unpack WIRED’s investigation into Pangram—the small Brooklyn startup rapidly becoming one of the most influential arbiters of whether writing is human or machine-generated.

Pangram has only 24 employees and has raised $13 million, yet its results are already reverberating across publishing, education, law, recruitment, and online media. The company analyzes text and returns a percentage estimating how much artificial intelligence contributed to it. Pangram’s growing reputation for accuracy has made those percentages extraordinarily powerful.

The most visible example involves Mia Ballard’s novel Shy Girl. After Pangram’s CEO publicly reported that the manuscript appeared 78 percent AI-generated, Hachette canceled its planned release. Ballard denied using AI. Other books, prizewinning stories, and newspaper articles have faced similar scrutiny, while some literary agents reportedly use detection results during private conversations with authors—and may quietly abandon projects without any public record.

Pangram says its newest model has a false-positive rate of just 0.0041 percent. It trains through techniques called synthetic mirroring and hard-negative mining, using mistakes to strengthen its detector. Independent testing helped establish Pangram as a leader, and Substack has integrated its technology so readers can evaluate possible AI use.

But no probabilistic detector is infallible. Critics warn that false positives can destroy reputations and careers. Research on AI detection has raised concerns about disproportionate effects on non-native English speakers and neurodiverse writers. A Notre Dame working paper found that an earlier Pangram model frequently classified lightly AI-edited academic abstracts as AI writing. Yet when fully AI-generated text was passed through a “humanizer,” Pangram detected it less than four percent of the time.

Context also changes results. The same passage may receive different scores when analyzed alone versus inside a longer manuscript. Pangram acknowledges weaker performance on short samples, especially below 100 words. These limitations matter because real decisions are often made from excerpts, proposals, essays, or online posts rather than complete books.

We examine the uncomfortable conflicts around AI detection: researchers receiving free Pangram credits, consultants making introductions to publishers, public callouts generating attention, and industry professionals becoming both advocates and business partners. None of these relationships automatically invalidate the technology, but they make transparency and independent validation essential.

The deeper question is whether society is asking an algorithm to answer something fundamentally ambiguous. Writing can be drafted by a person, lightly edited by AI, rewritten collaboratively, translated, or deliberately styled to resemble machine output. Reducing that complex history to a single percentage may offer confidence without certainty.

We discuss the safeguards publishers, schools, employers, and courts should adopt: never treat a detector score as proof; require independent review; preserve drafts and revision histories; give accused people a meaningful chance to respond; test for demographic bias; disclose conflicts of interest; and avoid irreversible decisions based on one proprietary tool.

Source: WIRED, published September 2, 2026. Written by Lexi Pandell. No individual editor was listed.

Follow The Daily AI Chat for clear, accessible analysis of the AI systems reshaping creativity, education, business, cybersecurity, policy, and society.