Slop Score
A leaderboard and analysis tool from EQ-Bench that computes how much a given text "smells like AI slop" by measuring the frequency of words, trigrams, and rhetorical patterns over-represented in LLM outputs compared to human writing.
Methodology
The Slop Score is a weighted composite metric: - 60% Slop Words: Frequency of individual words that appear unnaturally often in LLM outputs - 25% Not-x-but-y Patterns: Frequency of contrast patterns like "not just X, but Y" overused by AI - 15% Slop Trigrams: Frequency of 3-word phrases over-represented in AI text
Slop word and trigram lists are produced using the slop-forensics toolkit, which identifies words/n-grams statistically over-represented in LLM outputs compared to human writing, computed from 10 models' outputs on essay and creative writing prompts.
The tool is explicitly NOT an AI detector — it looks for glaringly over-used patterns rather than classifying text as AI or human.
Leaderboard (higher = more slop)
| Model | Slop Score |
|---|---|
| google/gemma-3-4b-it | 85.2 |
| google/gemini-2.5-flash | 77.6 |
| google/gemma-3-12b-it | 70.6 |
| google/gemma-3-27b-it | 69.5 |
| deepseek/deepseek-r1 | 67.9 |
| qwen/qwen3-4b | 62.2 |
| openai/chatgpt-4o-latest | 47.7 |
| openai/gpt-5-nano | 34.1 |
| claude-sonnet-4-20250514 | 27.3 |
| claude-sonnet-4-5 | 19.5 |
| human-baseline | 10.4 |
Key Slop Words
Top words from the slop list: silas, unsettling, elias, profound, subtly, meticulously, relentless, scent, flicker, echoes, blackwood, insistent, elara, frantic, thorne, clung, shimmering, faint, intricate, shadows.
Key Slop Trigrams
"something else something," "small intricately carved," "intricately carved wooden," "faint almost imperceptible," "something else entirely," "young woman named," "air hung thick," "dust motes danced."
Additional Metrics
The tool also provides writing style analytics: Flesch-Kincaid grade level, average sentence/paragraph length, dialogue frequency, lexical diversity (MATTR-500), and unique/total word counts.
Source Code
https://github.com/sam-paech/slop-score