hgaddipati/slop-index

The Slop Index

Which LLMs write the most like "AI slop"? This dataset has every generation behind theslopindex.com: 58 models, 112 everyday writing tasks (cold emails, Slack replies, social posts, short essays), 64,636 outputs.

Each model wrote each task several times from the same prompt. We then scored the writing mechanically against human writing from before ChatGPT existed (Enron email, pre-2022 tweets, older essays) and ran a blind human vote: people see two outputs for the same task and pick the one that feels more like AI slop. No LLM judges anything.

• Leaderboard: https://theslopindex.com
• Vote in the arena: https://theslopindex.com/arena
• Method: https://theslopindex.com/methodology
• Code: https://github.com/hgaddipati1118/slop-index

Files

Config	File	Rows	What it is
outputs	data/outputs.parquet	64,636	One row per generation
models	data/models.parquet	58	Per-model mechanical scores, prices, crowd Elo
	data/scenarios.jsonl	112	The writing tasks (persona, recipient, thread, request, length target)

from datasets import load_dataset
outputs = load_dataset("hgaddipati/slop-index", "outputs", split="train")
models = load_dataset("hgaddipati/slop-index", "models", split="train")

outputs columns

Column	Meaning
scenario_id, domain, category	The task. domain is email, slack, social or essay
model, model_name, lab	Which model wrote it
sample	Sample index for this model and task
system_prompt, user_prompt	Exactly what the model was sent
text	The model's output, unedited
length_target_words	How long a real person typically wrote for this kind of task
output_tokens	Completion tokens reported by the provider, where available

models columns

All mech_* columns are 0-100 and higher means sloppier.

Column	Meaning
mech_overall	Mechanical slop score across all four domains
mech_email, mech_slack, mech_social, mech_essay	Per-domain mechanical score
mech_conciseness	Padding: length versus what people wrote for the same kind of task
mech_templating	Stock structure and phrasing repeated across outputs
mech_rhythm	How metronomic sentence and paragraph lengths are, versus human writing
mech_tells	Rate of known AI tells (stock phrases, em-dash habits and so on)
price_in, price_out	USD per million input / output tokens at run time
crowd_elo	Arena Elo from blind votes. Higher means judged less sloppy. Snapshot at export time
crowd_games	Votes the model has appeared in

The site's headline Slop Score blends 40% crowd Elo with 15% each for the four mechanical axes. Models with few votes are scored at a neutral crowd value, so for newly added models the mechanical columns are the better comparison until votes accumulate.

How it was made

• Models were called through OpenRouter with the system prompt shown and default sampling.
• Each task has a persona, a recipient, sometimes a thread to reply to, and a short request written the way people actually ask ("need to tell tom the files are gonna be tuesday"). Writers are never named, so many emails contain placeholders like [Your Name]. The arena fills these with fixed stand-in names at display time; this dataset keeps the raw text.
• Human baseline corpora are not redistributed here, only the derived statistics used for scoring (see the GitHub repo).

Limitations

• English only, short-form only. Long-form and fiction are not covered yet.
• The mechanical axes catch surface patterns. They miss slop that lives in repeated sentence constructions or in what gets said rather than how.
• Crowd votes come from site visitors, not a recruited panel.

License and citation

Released under CC BY 4.0. Model outputs were generated through each provider's API.

@misc{slopindex2026,
  title  = {The Slop Index: which LLMs write the most like AI slop},
  author = {Slashy},
  year   = {2026},
  url    = {https://theslopindex.com}
}

A research project from Slashy.