Pangram has moved from a little-known 24-person startup to an influential judge of whether books, essays and articles were written with AI. Its scores have appeared in publishing controversies, academic research and platform integrations, even as critics question whether any detector should carry that authority.
The Brooklyn company has raised $13 million and offers a best-guess percentage for AI involvement in a piece of text. In July it raised $9 million, introduced Pangram 4 and advertised six openings that would expand its headcount by a quarter.
Substack began integrating the detector in late July so readers could check for possible AI use. Publishers and literary agents are also experimenting with it, making Pangram’s reliability consequential for writers whose work may be rejected or publicly questioned.
Literary controversies put Pangram in the spotlight
Pangram became widely known after speculation that author Mia Ballard used AI for the novel “Shy Girl.” Ballard denied the allegation, while CEO Max Spero publicly reported a 78% Pangram score. Hachette later canceled the book’s release, although Spero said the detector was only one part of a larger sequence of events.
Further accusations involved a New York Times Modern Love column, a Commonwealth Short Story Prize winner and several novels. Pangram returned scores ranging from 60% to 100% in the cases described by Wired. Those outputs are model estimates, not direct records of how the works were produced.
Penguin Random House told Wired that editors may use approved detection tools as one additional signal, never as a decisive result. Other major publishers declined to comment or did not respond. Literary agent Todd Shuster said Pangram had prompted difficult conversations with authors and requests that some rewrite material in their own voice.
Shuster also consults for Pangram and was introduced to the company by researcher Tuhin Chakrabarty, his partner. Chakrabarty receives API credits from Pangram, has become friends with Spero and publicly defends the company. These disclosed relationships do not invalidate their assessments, but they matter when weighing endorsements.
Training methods target mistakes as well as patterns
Pangram uses “synthetic mirroring,” asking language models to create close matches to human writing so its detector can learn AI patterns. It also searches datasets for false positives, then uses those errors to expand its training material. Spero says all of the company’s datasets are properly licensed.
The startup serves education, legal services and recruitment, although creative writing makes up the largest portion of its training text. Early independent tests helped it emerge among competitors including Originality.ai, GPTZero and Turnitin.
Pangram claims its latest model produces false positives only 0.0041% of the time, down from a previously reported 0.01%. It says borderline text is deliberately classified as human, which lowers false accusations but permits more AI-generated material to escape detection. These are company-reported performance figures.
The detector also works less reliably on passages under 100 words. Pangram acknowledges that the same excerpt can receive a different verdict when tested alone or inside a longer document, an important limitation when users paste short samples from manuscripts.
False positives and evasion limit certainty
Critics warn that automated detection can disproportionately flag non-native English speakers and neurodiverse writers. Research predating Pangram found such bias in other detectors, while Spero points to internal work that he says counters the concern. Wired notes that three major publishing scandals involving Pangram all concerned writers of color, though that pattern alone does not prove the model caused unequal treatment.
Evasion is another problem. A Notre Dame working paper found that Pangram 3.2 labeled lightly AI-edited academic abstracts as AI-written 64% to 80% of the time, yet detected fewer than 4% of AI-generated passages after they had been processed by a “humanizer.” Pangram’s newer model may perform differently, but the cited test shows how editing can confuse both sides of the classification.
Spero wants Pangram to provide more granular information, including detection of light AI editing, and says technical reports and transparency will build trust. The company’s scores can help focus an investigation, but the evidence in Wired’s account supports a cautious role: a probability to examine alongside drafts, editorial judgment and an author’s explanation, not a mechanical verdict on originality.
