Two Types of AI Slop: Bad Thought vs. Bad Register—And a Fix
A deep dive into why AI-generated content feels off and a step-by-step recipe to de-slop it.
Deep Dive
In a LessWrong post, the author argues that 'slop' conflates two distinct issues: superficial or incoherent writing (bad thought) and the recognizable AI cadence. They propose a 4-step recipe to de-slop: pick a capability, build an evaluation that fails, apply tricks like critique-and-revise loops, then optionally fold into training. A worked example shows how to fix a sloppy paragraph about Arnold Schwarzenegger.
Key Points
- Slop conflates two distinct issues: superficial/incoherent writing (bad thought) and the AI's recognizable stylistic register (cadences, lexical tells).
- Proposes a 4-step de-slopping recipe: choose a capability, build a failing evaluation, apply tricks (critique-revise, hillclimbing), and optionally fold into training.
- Demonstrates with a real example paragraph about 'tails come apart' in correlation extremes, highlighting weak phrasing.
Why It Matters
Understanding slop helps writers and AI users produce clearer, more trustworthy content from language models.