Asking Fable 5.1 to Rewrite 'LLM Slop' Docs Immediately Trips Its 'General Harm' Safeguard
StewartalsopIII · x · 2026-09-03
Matt Parlmer tried Fable 5.1, which people say sounds less obnoxious than typical LLM output, by asking it to pick five paragraphs of 'LLM slop' from his docs corpus and rewrite them to sound more human.
The request immediately tripped the model's 'general harm' safeguard and was blocked. The irony: a task explicitly about de-slopping AI text gets flagged as harmful, making it a textbook example of safeguard overtriggering.
More from Fun
- AI recreates a Call of Duty demo from a single screenshot in 30 hours — Dr_Singularity · 2026-09-03
- Joke prediction: by 2030 Nvidia macroducks will build full houses alone — RachelVT42 · 2026-09-03
- OpenAI's naming chaos: GPT-4.5 and o1 were each once slated to be GPT-5 — flowersslop · 2026-09-03
- "This meeting could have been a tool call": AI circles get a new catchphrase — GabGarrett · 2026-09-03
- Joking take: Musk and Jensen are running the century's compute stack — RachelVT42 · 2026-09-03
- Most movies ever made still can't be watched legally online, and BitTorrent lives on in 2026 — dbasch · 2026-09-03