Slopsift

Slopsift is an NLP linter that identifies artificial writing patterns commonly found in AI-generated content. It works by analyzing grammatical structure rather than matching against word lists, which means it catches the actual mechanics of inflated prose instead of just flagging suspicious vocabulary. The tool runs entirely on-device, so nothing you write ever leaves your machine.

The problem it addresses is familiar to anyone who has tried to spot formulaic writing. AI-generated text often follows predictable structures, uses vague attribution, leans heavily on passive voice, and pads arguments with filler that sounds good but says little. Slopsift identifies these patterns by parsing the dependency relationships between words, creating a map of how each sentence is constructed and checking it against known problematic structures.

Under the hood, it uses a 16 MiB quantized ONNX model custom-trained through structured distillation. The model predicts parts of speech and dependency relationships, then a rule layer evaluates those predictions against a catalog of patterns. Findings come in three confidence levels: errors for high-confidence detections, warnings for probable issues, and notes for things worth reviewing. Each detection includes the exact text range and the reason behind it, so you can make your own call rather than trusting a black box.

It fits writers who want to keep their voice but clean up patterns that read as canned, developers building tools that generate text, and CI workflows that need to lint prose alongside code. The output format supports ESLint-style JSON, which means you can wire it into existing automation without adapting your tooling. If you run coding agents that produce documentation or comments, Slopsift can flag the output before it ships.

What sets it apart is the focus on explainability and local processing. This is not a classifier that returns a single AI probability score with no reasoning. It's a linter that tells you what it found and why, so you can accept, reject, or learn from each detection. The rules are deterministic, meaning the same input always produces the same output. You're also not sending your drafts to a server. Everything runs in the browser or CLI, powered by the bundled model weights.

Slopsift is free and open source, installable via npm. There's no paid tier in sight, and the project is MIT-licensed. The irony that it was partly built with AI assistance is acknowledged upfront, but the point stands: the tool catches patterns, not origins, and humans produce these patterns too.