July 26, 2026
Caught by the bot that cried bot
I wouldn't say Pangram is broken, but I would say that it's brittle
AI detector drama explodes as readers say the tool can’t even agree with itself
TLDR: A writer says Pangram labeled the same text both human and AI depending on how it was sliced, raising big doubts about the detector. Commenters were brutal, joking that it never worked in the first place and warning that false accusations could hit ordinary people.
The real spectacle here isn’t just one writer testing an AI detector — it’s the comment-section pile-on that followed. Freddie deBoer says Pangram, a tool meant to sniff out machine-written text, tagged one chunk of his very human essay as 100% AI-written, then turned around and called the full essay 100% human. Even wilder, smaller pieces inside that “AI” chunk could be judged human again. That set off a mini riot among readers, who basically responded: so this thing can accuse you with total confidence and still contradict itself? Not great.
The hottest reaction was pure contempt. One commenter went straight for the jugular with, “Can something be broken that never actually worked?” Another called the company’s defense a cartoonish dodge, dropping the instant-classic “Holy strawman-batman.” And the fear underneath the jokes is serious: multiple people shared their own stories of being falsely flagged after writing ordinary, human text about their day or schoolwork. The mood was less “interesting bug” and more “this could get real people in trouble.”
Then came the dark humor. One reader joked that even if Pangram is nonsense, maybe it could still become a “Yelp-like racket,” which is about as subtle as throwing a chair across the room. That’s the vibe: sarcasm, distrust, and a growing sense that a tool marketed as a truth machine might actually be a drama generator. In this story, the comments didn’t just react — they delivered the verdict.
Key Points
- •Freddie deBoer says Pangram labeled a roughly 300-word section of his older essay as 100% AI-written with high confidence.
- •After purchasing a Pangram license and testing the full essay, deBoer says the same tool labeled the entire approximately 5,000-word piece as 100% human-written with high confidence.
- •DeBoer reports that splitting the flagged section into smaller pieces produced results labeling those subsections as 100% human-written.
- •The article states that Pangram's classifications can change depending on the larger textual context surrounding a passage.
- •DeBoer says he was able to generate false positives by writing text himself that Pangram identified as AI-written.