WIRED investigates Pangram, the startup making and breaking publishing careers
The AI-detection arbiter claims a false-positive rate of 0.0041% and is integrated into Substack, but writers warn of neurodiverse and non-native bias.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Brooklyn-based AI detection startup Pangram has emerged as the publishing industry's primary arbiter of authenticity, according to a WIRED investigation on September 2.
- Major publishers like Penguin Random House and platforms like Substack use the tool, which has already led to the cancellation of multi-million dollar book deals.
- Critics warn the tool exhibits bias against neurodiverse and non-native English writers whose formal prose patterns are frequently flagged as robotic.
Pangram, a Brooklyn-based startup with 24 employees and $13 million in funding, has quietly become the most powerful arbiter of authenticity in the publishing world, according to an in-depth investigation by WIRED published on September 2. Founded by Max Spero (formerly of Google) and Bradley Emi (formerly of Tesla), the company's software is now used by major publishers like Penguin Random House and top literary agencies to vet manuscripts before they are acquired.
The tool's impact is already highly material, leading directly to the cancellation of lucrative deals. The novel *Shy Girl* by Mia Ballard was dropped by publisher Hachette after Pangram flagged it as 78% AI-generated, and a thriller that sold for $2.4 million, titled *Call Me, I'll Hide the Body*, was flagged by the software as 97% synthetic. Pangram has also been integrated directly into Substack's writing platform to detect automated content farms.
Pangram's founders claim their latest model, Pangram 4, has a remarkably low false-positive rate of 0.0041%, achieved by training models on human texts paired with closely mirrored AI outputs. However, the system's growing influence has drawn intense backlash. Critics and researchers warn that AI detectors are biased against non-native English speakers and neurodiverse writers whose structured, predictable prose patterns are easily misidentified as robotic. WIRED notes that the three biggest public scandals involving Pangram's flags all featured writers of color, fueling concerns about equity in publishing.
Furthermore, the tool faces technical limitations. While highly effective against raw generative text, research shows Pangram caught less than 4% of AI-generated text that had been run through a 'humanizer' editing tool. Spero offers cash rewards to flagged authors who can prove they wrote their work themselves, though he admitted he rarely pays the bounty, remaining confident in the tool's judgment.
When an unproven, proprietary algorithm with a documented bias against non-native speakers is given the power to terminate million-dollar book deals and ruin writing careers, 'authenticity' becomes a source of systemic risk. The publishing industry's wholesale adoption of Pangram reveals a desperate, defensive rush to police the boundaries of human creativity — even if it means sacrificing neurodiverse and diverse human voices to keep out the machine-generated churn.