Frontier LLMs went from 'trained on the whole internet' to banning most of it

ivan_bezdomny · x · 2026-11-16

ivanbezdomny observes that in a few years frontier LLMs went from being "trained on the whole internet" to restricting usage across much of it: sensitive news and discussions, biology, jokes and humor, model training, cybersecurity, copyrighted engineering books, and personal info.

He contrasts vendors: OpenAI's Astra reads but blocks writing a "baby Hitler" joke; Claude silently fixes transcripts; Gemini censors preemptively before you even reach a sensitive term. He quips that OpenAI "probably won't report you though. They know there are false positives."

Related event: Test Compares Content Censorship Across Gemini, Claude and OpenAI Models(3 posts)→

Original post →

More from Models

Models channel →