July 23, 2026
Pride, Pangram, and Prejudice
Scanning for Pangram Errors
AI writing snitch gets stress-tested and the comments are already at war
TLDR: A book-community moderator tested Pangram by scanning rare old books that weren’t online, trying to see if the AI detector was just faking it through memory. Commenters are split between "finally, a useful AI snitch" and "slow down, this thing could still be confidently wrong," which matters because readers and writers are already using tools like this to judge what counts as real writing.
A romance-book community moderator decided to do what the internet always demands: bring receipts. After hearing big promises from Pangram, the buzzy tool that claims it can spot AI-written text with almost laughably low mistake rates, they went full detective mode and scanned old, obscure books that were never even digitized online. The idea was simple: if Pangram only looks smart because it memorized famous old writing, these hidden paper-only books should trip it up. Instead, the test came back looking surprisingly strong — and that immediately lit up the crowd.
The comments split into two camps fast. Team Believer was basically ready to install Pangram into their veins. One fan said they’d happily pay for a browser add-on that scans text live, while another argued AI writing has become easier to spot because some chatbots now sound so stiff and overpolished that they practically glow in the dark. Others took the calmer middle road, saying no detector is perfect, but if it’s used as a warning sign instead of a final verdict, it’s still useful.
But the skeptics were not about to log off quietly. One commenter waved around a Freddie deBoer critique, warning that Pangram may still produce misleading results. And the funniest side quest? A reader got tripped up by the sentence about "my community, mine included" and briefly thought the author was confessing to personally flooding book spaces with unlabeled AI slop. In other words: serious test, serious stakes, and of course, classic comment-section chaos.
Key Points
- •The article evaluates Pangram, an AI-writing detector that claims a false positive rate of 1 in 10,000.
- •The author uses Pangram as a moderation aid in a romance book community facing unlabeled AI-generated submissions.
- •Preliminary testing described in the article found no AI detections on RoyalRoad stories published before ChatGPT's release.
- •The post focuses on whether Pangram may be memorizing training data rather than learning general features of AI-generated text.
- •To test for memorization, the author scanned old books that were not available in digital form online and submitted them to Pangram.