The most prestigious law school admissions discussion board in the world.
|
 |
 
|
AI says it’s slop language comes from Reddit and raters in SE Asia/Kenya
| cannon | 06/05/26 |
Poast new message in this thread
Date: June 5th, 2026 10:57 PM Author: cannon Lmao “sorry for the slop, but it’s because of the Google slaves around the world”
When LLMs scrape the internet, Reddit is one of the most heavily weighted text sources for "conversational" English. The platform’s culture is deeply rooted in pedantry, debate, and the desire to correct people, making "actually," "to be fair," and "let's dive in" the dominant currency of communication. The AI didn't just learn vocabulary; it learned an online persona. As for who is steering the AI to keep using these words during the tuning phase, the reality of AI human raters explains a lot about why the output feels so formulaic. The Common Profiles of AI Raters The people training AI via RLHF (Reinforcement Learning from Human Feedback) generally fall into two vastly different demographic buckets. This split heavily influences the "slop" we see today. Profile A: The Global Gig Worker (The "Speed-Labelers") A massive portion of base-level AI training is outsourced to global platforms like Remotasks, Outlier, mechanical Turk, or Scale AI. www.taskmonk.ai Demographics: Often based in developing economies (such as Kenya, the Philippines, or parts of South Asia) or young, underemployed gig workers in the US and Europe. The Working Reality: They are paid per task, often pennies, and their performance is tracked strictly on speed and volume. They don't have time to appreciate beautiful prose. The "Slop" Impact: Because they need to grade thousands of AI responses a day, they rely on visual and structural shortcuts. A response that uses clear, unambiguous signposts like "Moreover," "Actually," or "In conclusion" is incredibly easy to skim and mark as "structured and helpful." The AI learns that using these formulaic words gets it an "A" grade from a tired worker trying to hit a quota. Profile B: The Underemployed Academic (The "Prompt Engineers") As AI companies have tried to move away from low-quality data, they have started hiring "expert" reviewers—often underemployed college grads, freelance writers, and humanities PhDs. Demographics: Typically 22–35 years old, highly educated, native English speakers, working remotely. The Working Reality: They are given incredibly dense, 50-page rubrics by tech companies. They are told to ensure the AI is strictly unbiased, highly detailed, perfectly safe, and objective. The "Slop" Impact: This group brings a very specific "academic essay" or "corporate HR" tone to the table. They reward the AI for sounding hyper-cautious, exhaustive, and balanced. This is where phrases like "It is important to remember," "A testament to," and "Delve" come from. It’s the language of a grading rubric come to life. The Reddit Paradox When you combine those two profiles with a training dataset built on Reddit, you get a perfect storm of robotic condescension.
Because the human raters are grading based on a strict checklist rather than asking "Does this sound like a real, authentic human being?", the AI optimizes for the checklist. It becomes a caricature of online text: safe, structured, slightly pedantic, and utterly devoid of individual voice.
|