Tools, research, and word-lists for catching and fixing AI slop: the tells that mark writing as machine-generated, in every language you publish in.
Most "AI detector" products are English-only black boxes. But AI slop is not only English. In Korean it surfaces as 번역투, in Russian as канцелярит, in Chinese as 公文腔, in Vietnamese as Hán-Việt overuse. This list is organized by language and grounded in the research on how AI text actually reads, so you can find the tells, the tools, and the fixes for the languages you work in.
Disclosure: the maintainer of this list also maintains slop-gate, which appears in several sections below.
- What is AI slop
- The research
- Catch it: linters
- Detect it: classifiers
- Fix it: humanizers
- By language
- Word and phrase lists
"Slop" is text that reads as machine-generated: a handful of over-used words, one punctuation mark, and stiff calques that show up far more often in LLM output than in anything a person writes unprompted. Readers have learned to spot it, and the moment they do, trust drops even when the substance is fine.
What the tells actually are, measured at scale.
- Quantifying LLM usage in scientific papers - Liang et al., Nature Human Behaviour (2025). Over a million papers; "realm," "intricate," "showcasing," and "pivotal" were flat for a decade, then surged from 2023.
- Delving into LLM-assisted writing through excess vocabulary - Science Advances (2025). 15 million biomedical abstracts; "delve" rose about 1,500% between 2022 and 2024.
- Measuring AI "Slop" in Text - Shaib et al. (2025). Builds a taxonomy of slop from expert interviews and span-level annotation, and finds that slop judgments track dimensions like coherence and relevance.
- EQ-Bench Slop Score - Public leaderboard scoring language models on slop words, "not x, but y" constructions, and over-represented trigrams.
Tools that flag specific tells in your own prose, locally or in CI.
- slop-gate - Zero-dependency CLI: the em-dash and ~40 English tells, plus opt-in translationese packs for Korean, Russian, Vietnamese, Chinese, and Filipino.
- Vale - Customizable prose linter with style rules; the standard for docs teams.
- alex - Catches insensitive, inconsiderate phrasing.
- proselint - Lints prose against a body of style-guide best practices.
- write-good - Flags weak, passive, and wordy phrasing.
- LanguageTool - Open-source grammar and style checker across many languages.
Tools that try to classify whether text is AI-generated. Treat the verdicts with caution: detectors produce frequent false positives, are easily bypassed, and should not drive high-stakes decisions about a person.
- GPTZero - AI-text detector aimed at education.
- Originality.ai - AI-detection and plagiarism checking for publishers.
- Pangram - Commercial AI-text classifier with a published technical report and independent academic evaluations of its error rates.
- Binoculars - Open-source, zero-shot detector that contrasts how two related language models score the same text (ICML 2024).
Tools that remove tells, by rewriting the text or by stopping the tells at generation time. Use with judgment: the goal is clear human prose, not evading a detector.
- im-not-ai / Humanize KR - Claude Code skill that rewrites AI-drafted Korean to strip 번역투 without changing the meaning.
- QuillBot - Paraphrasing and rewriting (English-focused).
- antislop-sampler - Sampler that backtracks during generation to suppress thousands of known slop phrases; part of the Antislop framework (ICLR 2026).
The fingerprint of AI text differs by language. Each section names the phenomenon and lists the tools for it.
The over-used LLM vocabulary (delve, leverage, robust, tapestry) plus the em-dash.
- slop-gate vocabulary pack - English LLM overuse words and phrases.
- slop-gate punctuation pack - Punctuation tells, including the em-dash.
- Wikipedia: Signs of AI writing - Field guide to the vocabulary, punctuation, and formatting tells of chatbot text, built from real examples in flagged Wikipedia articles.
- WikiProject AI Cleanup - The Wikipedia editor project that reviews suspected AI-generated content and maintains the signs guide.
Translationese: English calques (의 경우, 따라서) and stiff written-register forms. The naturalness lever in Korean editorial writing is stripping 번역투 and restoring native discourse particles.
- slop-gate Korean pack - Korean translationese patterns and suggested rewrites.
Bureaucratic register: the copula является, split-predicate verbs like осуществить. Kornei Chukovsky named it the central disease of the language; Maxim Ilyakhov's Пиши, сокращай is the canonical guide against it.
- slop-gate Russian pack - Russian bureaucratic-register tells.
Officialese: 进行+noun padding, 予以, and stock openers like 众所周知.
- slop-gate Chinese pack - Chinese officialese and padding patterns.
Sino-Vietnamese vocabulary and calque openers (thực hiện, trong trường hợp) where plain Vietnamese is simpler.
- slop-gate Vietnamese pack - Vietnamese calque and over-formality patterns.
Over-formal literary Tagalog and calque openers (upang, sa pamamagitan ng) where editorial Filipino is conversational Taglish.
- slop-gate Filipino pack - Filipino over-formality and calque patterns.
The raw lists of tells, editable and worth reading.
- slop-gate rule packs - JSON lists of tells per language, each with a suggested fix.
- The Ultimate AI Slop Word Blacklist - One-page blacklist of words and phrases over-represented in LLM output.