AI Detection Startup Pangram Gains Power Over Publishing Careers

A small AI detection company based above a Popeyes restaurant in Brooklyn has rapidly become one of publishing's most influential gatekeepers, with the power to make or break book deals worth millions of dollars.

Pangram, founded in 2023 by Stanford graduates Max Spero and Bradley Emi, analyzes text to determine what percentage was AI-generated. The 24-employee startup has raised $13 million and operates with a fraction of the resources of major AI companies, yet its verdicts have led to canceled book contracts and public accusations across the literary world.

The company gained prominence in January when CEO Spero posted that Mia Ballard's novel "Shy Girl" was 78 percent AI-generated after receiving a pirated manuscript. Publisher Hachette subsequently canceled the book's release, though Ballard denied using AI. Similar high-profile cases followed, including a New York Times Modern Love column flagged at 100 percent AI and a thriller that sold for $2.4 million scored at 97 percent AI-generated.

Why it matters

As generative AI tools become ubiquitous in writing workflows, the question of detection has moved from academic integrity offices to mainstream publishing. Pangram's rapid adoption by agents, publishers, and platforms like Substack demonstrates the industry's urgent need for solutions—but also raises critical questions about who controls these judgments and whether the technology is ready for such consequential decisions. When a single percentage score can end a career or void a seven-figure contract, the accuracy and fairness of detection systems become matters of professional survival.

Technical approach and limitations

Pangram uses "synthetic mirroring" to train its models, having large language models generate text similar to human writing to learn AI patterns. The company also employs "hard negative mining" to identify and learn from false positives. Spero emphasizes that all training datasets are properly licensed, with creative writing forming the largest segment.

The company claims its newest model produces false positives only 0.0041 percent of the time. However, independent research reveals significant limitations. A Notre Dame working paper found that Pangram's earlier model flagged lightly AI-edited academic abstracts as AI-generated 64 to 80 percent of the time, while AI text processed through "humanizer" tools evaded detection more than 96 percent of the time.

Pangram acknowledges its tool performs worse on text under 100 words and that identical passages can receive different scores depending on surrounding context—a significant issue given that the free service limits users to about 2,000 words daily, with a median input of just 350 words.

Concerns about bias and process

Critics have raised concerns about potential bias in AI detection systems. Research predating Pangram showed detectors were more likely to flag work by non-native English speakers as AI-generated, and neurodiverse writers report disproportionate false positives. The three biggest publishing scandals involving Pangram—"Shy Girl," "Daggermouth," and "Call Me, I'll Hide the Body"—all involved writers of color, though the selection of which manuscripts get scanned remains a human decision vulnerable to existing industry biases.

Regina Brooks, president of the Association of American Literary Agents, noted the need to examine inequities in how detection tools are applied. Sam Illingworth, a professor of critical AI literacy at Edinburgh Napier University, stated flatly: "I don't think [AI detectors] work. I think that detectors are prejudiced against certain people."

Research by Tuhin Chakrabarty at Stony Brook University, who analyzed over 14,000 self-published novels using Pangram, found nearly 20 percent had substantial AI-detection scores. Chakrabarty has become close friends with Spero and receives free API credits from Pangram to support his research, while also publicly defending the company.

Major publishers remain cautious. Penguin Random House confirmed its editors may use approved AI-detection tools but emphasized they are "not determinative and are only one component of a broader editorial process." Other Big Five publishers either declined comment or did not respond to inquiries.

These details were first reported by Lexi Pandell in WIRED.