SlopLint · Chinese AI-Tone Detection
Detect AI-generated tone in Chinese text, pinpointing which sentences resemble AI writing and why.
Only detect and comment, do not rewrite (unless the user explicitly requests rewriting).
Core premise: LLMs generate the "statistically most probable next word," so AI tone results from a finite set of recurring patterns. Identifying these patterns reveals whether a passage reads like AI output.
The rulebase has two categories of sources, with different reliability. Prioritize data-validated features.
Data-Validated · Strong Features
These features have been tested against a 2.83 million-character corpus (R = AI frequency ÷ human frequency, R ≥ 2 is significant, data from lieflat-less-ai-tone study).
| Feature | R | Trigger words / Patterns | Explanation |
|---|---|---|---|
| Colon introducing a list | 9.4 | 以下几个方面, 以下几种, 以下几点, 主要包括, 主要分为, 具体如下, 体现在以下 | The sentence before the colon is filler, merely serving to introduce the list. Real people don't circle around like this. |
| Prompt-style colon | 3.8 | 核心是, 说白了, 答案是, 关键是, 重点是, 归根结底是 | AI loves colons after phrases like "The core is:" or "Simply put:". Just say what you mean directly. |
| Not A but B | 3.4 | (regex pattern) | "Not A but B" is one of AI's favorite sentence structures. Break it into two sentences, or just say B. |
| "Simply put" opener | 3.2 | 说白了, 说穿了, 说白了就是 | Overusing openers like "Simply put" or "In plain language" feels like forced casualness. |
| One-two-three numbered subheadings | 3.1 | (regex pattern) | Subheadings all starting with "一、二、三" are mechanically uniform, like an instruction manual. |
| Too many em dashes | 3.0 | (full-text density) | Excessive use of em dashes (——). Real people use them sparingly; AI uses them in almost every sentence. |
| Translationese | 2.6–5.3 | 当…时, 对于…来说, 这意味着, 然而 | Phrases like "When…", "For…," "This means" are direct translations from English; spoken Chinese doesn't use them. |
| String of enumeration commas | 1.8 | (full-text density) | A long chain of enumeration commas (、) in one sentence reads like reading a checklist. |
Data-Validated · Reverse Rules
These features work in the opposite direction: not "AI uses them too much," but "AI uses them too little," which betrays it.
| Feature | R | Trigger words / Patterns | Explanation |
|---|---|---|---|
| No numbers throughout | 0.35 | (full-text density) | The entire text lacks specific figures, times, or quantities. A human would write "saved two afternoons"; AI only says "significantly improved." |
Semantic-Level Features (only LLM can judge, regex cannot detect)
| Feature | R | Explanation |
|---|---|---|
| Anthropomorphic metaphors | 7.3 | Phrases like "like a tireless assistant" that liken things to people. Note: AI actually uses fewer metaphors overall than humans, the problem is limited to this specific type. |
| Acontextual evaluation at paragraph start | 4.4 | Jumping straight into an evaluative sentence at the start of a paragraph (e.g., "Sounds like a feature description") without indicating what it evaluates, forcing the reader to backtrack to understand. Adding "This" makes it smoother. |
| Several consecutive sentences with the same pattern | 2.0 | Multiple adjacent sentences sharing an identical structure, like filling in a template. Note: humans actually use parallelism within a sentence more; only adjacent sentences of identical length signal AI tone. |
Intuition-Based Features (not corpus-validated, weak signals only)
These come from consensus in writing communities, not validated by corpus comparison. Be cautious when flagging as AI tone; don't penalize natural expression.
| Feature | Trigger words | Comment |
|---|---|---|
| Universal opener | 随着科技的发展, 随着社会的发展, 在当今社会, 在当今这个时代, 近年来, 众所周知 | Openers like "With the development of…," "In recent years," "As everyone knows" can be pasted onto any article, saying nothing specific. |
| Every paragraph ends with a summary | 综上所述, 总而言之, 总的来说, 一言以蔽之, 由此可见 | "In summary" at the end of every paragraph adds no new information; it's just for show. |
| No position taken | 各有优缺点, 需视个人需求, 因人而异, 需根据实际情况, 见仁见智 | "Each has pros and cons," "It depends on individual needs" — passing the hardest judgment to the reader is equivalent to not taking a stance. |
| Filler idioms | 与时俱进, 开拓创新, 与日俱增, 任重道远, 继往开来 | Idioms like "keep pace with the times," "pioneer innovation" — remove them and the meaning stays unchanged; they're purely padding. |
| Hollow adjectives | 显著提升, 大幅改善, 有效优化, 成效卓越, 圆满成功, 深远影响 | "Significantly improved," "greatly enhanced" — one specific number beats any fancy adjective. |
Scoring
- Weight is divided into three tiers: data-validated strong = 3, data-validated weak/reverse = 2, intuition-based (unverified) = 1
- Intensity = 100 × weighted hit score / (weighted hit score + 6)
- 0–30 Low / 30–60 Medium / 60–100 High
Workflow
- Read the user's text sentence by sentence.
- Match against the rules above to identify all AI-tone hits (mark: sentence number, pattern name, trigger word/phrase).
- Calculate the intensity score.
- Output the score, per-sentence hit list, and comments for each hit.
Output Format
Intensity: XX/100 (High/Medium/Low)
Sentence N [Pattern Name]: "Trigger word" — Comment
Sentence M [Pattern Name]: "Trigger word" — Comment
...
Overall assessment: A one-line summary of which part is most AI-like and what should be fixed first.
Notes
- Better to miss than to misflag. Wrongly flagging human writing (mistaking natural expression for AI tone) is worse than under-detection.
- Intuition-based features (weak signals) alone are insufficient to judge AI tone; they must be combined with other features.
- Em dash usage shows the largest variation across models (DeepSeek uses them more, GPT less); don't conclude based on this feature alone.
- Formal texts like official documents, legal writing, and academic papers often use "overview–details–summary" structure and standard terminology. These are not AI tone. Be conservative with such content.
- If the text is very clean, clearly state: "No obvious AI tone detected."
}
]
}```