Pangram False Flags AI Writing Claims as 1-in-10,000 Error Rate Faces Scrutiny
Updated
Updated · The Atlantic · May 30
Pangram False Flags AI Writing Claims as 1-in-10,000 Error Rate Faces Scrutiny
1 articles · Updated · The Atlantic · May 30
Summary
Taylor Lorenz was falsely accused this month of using AI for a Vanity Fair story after Pangram flagged her work; CEO Max Spero later said the detector had erred.
Pangram says it falsely labels human text as AI only about 1 in 10,000 times, but its misses are far more common—around 1 in 70 in one test—and AI “humanizer” tools repeatedly fooled it in the report’s experiments.
Those limits matter because Pangram is already used by publishers, universities and scientific groups, and can plug into Canvas to scan student submissions at a scale of more than 30 million U.S. high school and college students.
Recent disputes have widened beyond Lorenz: Wall Street Journal editor James Taranto called Pangram a “defamation machine,” and Pangram-driven claims also helped fuel allegations about Pope Leo XIV’s encyclical and other high-profile texts.
The broader risk is that institutions treat a black-box detector as proof even as chatbot writing keeps changing, making outside accuracy checks quickly outdated and false accusations harder to contain.
Is the fight against AI-generated text creating a new kind of digital witch hunt?
When AI can perfectly mimic humans, how can anyone prove their own words are original?
Pangram and the High-Stakes Future of AI Detection: Accuracy, Policy, and the Battle for Human Authorship
Overview
By 2026, Pangram has become a leading tool in AI detection, standing out for its broad and accurate approach compared to earlier, less reliable detectors. Unlike tools that focus narrowly on student essays, Pangram’s training covers a wide range of texts, allowing it to accurately evaluate everything from academic papers to online posts and journalism. This comprehensive capability puts Pangram at the center of high-stakes disputes over human authorship, where authenticity is crucial. Its rise follows the shortcomings of previous detectors, such as ZeroGPT’s errors and OpenAI’s abandoned tool, highlighting Pangram’s pivotal role in verifying content in critical domains.