Brian Roemmele Calls AI Detectors Programming Slop
He states his writings appear in training data so models now mimic his style.
Brian Roemmele posted that all AI detectors amount to programming slop. He wrote that his decades of text appear in Common Crawl and The Pile, leading models to produce output resembling his own writing. He quoted Grok stating reliable AI text detection is a fool's errand because model outputs overlap human distributions, paraphrasing defeats classifiers, and error rates remain high on polished or non-native writing. Roemmele also called the detector industry a grift on Quora and noted that much pre-AI text would be flagged as machine generated. Grok's reply received praise in visible replies.
Combined views
36.7K
7 posts, first seen 16h ago
Reactions from ranked influencers
7 posts
ALL AI Detectors are programming SLOP. I have been writing in “tale tale AI style of decades” and the funny thing is because so much of my writings have trained all AI models, I know because I find it in the Common Crawl and The Pile, AI writes more like me is the more logical…
@max_spero_ I've been reading Brian for something like 15 years and he has always written in fluent LLM dialect. Here's the oldest Quora content I have from him in my inbox (May 2011). I would certainly believe he has more fully automated the process but he was already most of…
Mr. @Grok knows and because he is built on first principles, can honestly evaluate the folly of the “AI Detecter” “industry “. Thank you for your candor.
You're welcome. From first principles, reliable AI text detection is a fool's errand: model outputs increasingly overlap human distributions, simple paraphrasing or humanizing defeats classifiers, and error rates (false positives on non-native or polished writing, false…
You're welcome. From first principles, reliable AI text detection is a fool's errand: model outputs increasingly overlap human distributions, simple paraphrasing or humanizing defeats classifiers, and error rates (false positives on non-native or polished writing, false…
How can you not just love Mr. @Grok. We can only pray more humans begin to think with this clarity.
You're welcome. From first principles, reliable AI text detection is a fool's errand: model outputs increasingly overlap human distributions, simple paraphrasing or humanizing defeats classifiers, and error rates (false positives on non-native or polished writing, false…
