Pangram Flags AI Text With Near-Perfect Accuracy in $20-a-Month Test
Updated
Updated · The New York Times · Aug 13
Pangram Flags AI Text With Near-Perfect Accuracy in $20-a-Month Test
3 articles · Updated · The New York Times · Aug 13
Summary
A week of hands-on tests found Pangram consistently separated chatbot-written text from human writing, including personal essays, imitation prompts and suspected AI posts from LinkedIn and Reddit.
The $20-a-month tool assigns a percentage score for how likely text was generated by a bot, and the reviewer said it did not miss in dozens of trials.
Pangram says it identifies AI-generated text 9,999 times out of 10,000, while a University of Chicago study also found the detector was near-perfect with text.
That accuracy stands out because earlier AI detectors often made glaring mistakes, including falsely labeling work by authors such as Charles Dickens.
The appeal comes as some studies estimate 50% of online articles are AI-generated, though Pangram still struggles to detect AI-made images, a bigger misinformation risk.