feedhive/ai-slop-detector is a list repository that collects 120 expressions that look like they were written by AI, organized into 9 categories. In commit 95d5507 on October 12, 2026, all 5 files were newly added (2,369 lines added, 0 lines deleted): README.md at 1,163 lines, patterns.json at 909 lines, scripts/build.mjs at 111 lines, and skills/ai-slop-detector/SKILL.md at 165 lines arrived together. The license is MIT, and the copyright notice reads 2026 FeedHive.
The categories break down as Vocabulary 20, Stock phrases 15, Sentence patterns 22, Structure and rhythm 12, Imagery and metaphor 6, Substance 12, Formatting and punctuation 10, Chatbot residue 8, and Social media and LinkedIn 15. Each pattern comes with a short description, a bad example, and a version rewritten in plain prose. The targets include vocabulary such as 'delve into', 'game changer', and 'testament to', as well as introductions like 'In today's fast-paced world'.
Usage is also broken down into several approaches. The README explains how to install the skill for agents such as Claude Code and Codex, or copy the skills/ai-slop-detector directory into ~/.claude/skills/; how to paste the README into a chat window to check a draft; and how to put a draft into the browser tool (feedhive.com/tools/ai-slop-detector). The original list is patterns.json, and a build script is included so that running node scripts/build.mjs regenerates the README and the skill file.
That said, this repository is not an authorship-detection tool. The README makes clear that this list alone cannot tell whether something was written by a human or by AI, that a single pattern is also common in human writing, and that it takes several overlapping patterns within one paragraph for a passage to read like AI slop. There are no detection accuracy figures or false-positive examples in the newly added files. So this project is less a program for detecting AI slop than a reference chart for comparing a writing model against human work side by side.