slop-detect — Privacy-Preserving AI Text Detection

The arXiv slop crisis is real. A 2026 study estimates that 65% of CS papers on arXiv are now AI-written. And it’s not just academia — AI-generated content is flooding blogs, marketing copy, product reviews, and social media.

But here’s the thing: most AI text detectors today are cloud-based. You paste text into a web form, it gets sent to a server, analyzed, and returned. That’s a privacy nightmare if you’re analyzing sensitive documents, internal communications, or unpublished research.

So I built slop-detect — a CLI tool that runs entirely on your machine. No API calls. No data leaves your computer. Just five local heuristics that together give a surprisingly accurate picture.

How it works

Five detectors, each scoring text from 0 (human) to 1 (AI-generated):

1. Burstiness — Human writing has variable sentence lengths. AI text is more uniform. The coefficient of variation tells the story.
2. Perplexity Estimate — A character-level 4-gram model built from the text itself. Surprising sequences = more human-like.
3. Repetition — AI text repeats n-grams more frequently. Checks 3-gram, 5-gram, and 8-gram repetition ratios.
4. Hedging Phrases — Counts frequency of AI-typical phrases like “delve,” “foster,” “landscape,” “it’s worth noting,” “in conclusion.” There are 34 phrases in the lookup.
5. Readability Consistency — Uses the GINI coefficient of sentence lengths. Human text naturally mixes short and long sentences; AI text is more uniform.

Usage

# Analyze a file
slop-detect paper.txt

# Pipe text
cat paper.txt | slop-detect --stdin

# JSON output (great for CI pipelines)
slop-detect --json article.md

# Run only specific detectors
slop-detect --detectors burstiness,hedging sample.txt

# Adjust sensitivity
slop-detect --sensitivity high draft.txt

Example output

Running it on a known AI-generated article about “AI transforming industries” — the kind of slop that fills half the internet:

┏━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━┓
┃ Detector             ┃ Score ┃ Weight ┃
┡━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━┩
│ Burstiness           │  0.66 │   0.25 │
│ Perplexity Estimate  │  0.52 │   0.20 │
│ Repetition           │  0.51 │   0.20 │
│ Hedging Phrases      │  1.00 │   0.20 │
│ Readability Variance │  0.62 │   0.15 │
├──────────────────────┼───────┼────────┤
│ OVERALL CONFIDENCE   │  0.66 │        │
│ VERDICT              │ AI-GENERATED │
└──────────────────────┴───────┴────────┘

The hedging detector nailed it — 35 hits per 1000 words, including “delve,” “foster,” “in conclusion,” “landscape,” “tapestry,” and “transformative.”

Why fully local matters

Most AI detectors are SaaS products. You upload text, they analyze it, and somewhere a server stores your data. If you’re a journalist verifying sources, a researcher reviewing peer papers, or a lawyer checking client documents — sending that text to a cloud service is a non-starter.

slop-detect is pip install away. Runs on any machine with Python 3.10+. No internet required after install.

Check it out

• GitHub: github.com/vikasudasi/slop-detect
• Install: pip install slop-detect (coming soon to PyPI)
• License: MIT

The best way to fight AI slop is with tools that are transparent, local, and verifiable. slop-detect is my contribution to that effort.