What Makes Chinese AI Text Sound Like AI?
An open study compared 300 Chinese articles written by five AI models with 329 written by people. Eleven common "AI tone" signs held up; fifteen did not, and several popular rules point the wrong way. Here is what it found and how to use it when you edit Chinese text.
Chinese writers call it AI 味, "AI tone": the feel of text a model wrote. Checklists for spotting and removing it circulate widely, but few have been tested. One open project did test them, and its results are worth knowing before you edit Chinese text by hand or with a tool.
The study
The project is lieflat-less-ai-tone, published on GitHub under the MIT licence (repository). Its authors built a corpus of 629 Chinese articles, about 2.83 million characters: 300 generated by five models (Claude, DeepSeek, Gemini, GPT and Kimi, 60 each, topic only and no style instructions) and 329 written by people. For each of 26 commonly cited "AI tone" features, they compared how often it appears in generated and human text.
One limitation, which the authors state themselves: the corpus is not published, so the figures cannot be checked independently. The method and scripts are public.
Signs that held up
The authors kept a feature only if generated text used it clearly more often than human text, consistently across their human samples, and if the rule points to a specific sentence to change. Among the eleven that passed:
| Feature | Generated vs human, as reported |
|---|---|
| A paragraph opening with a comment that does not say what it refers back to | 4.4× |
| Personifying an abstract idea as the vehicle of a metaphor | 7.3× |
| A colon after a cue phrase such as 核心是：, or a line ending in a colon that leads into a list | 3.8× and 9.4× |
| The "not A but B" construction | 3.4× |
| Em dashes | 3.0×, with a hundredfold spread among human writers |
| Ordinals used as subheadings | 3.1× |
| The same sentence pattern repeated across adjacent sentences | about 2× |
| Five translationese constructions, including 当…时 and 这意味着 | 2.6–5.3× |
It also found that human text carries far more concrete figures, so one rule is simply not to replace an existing number with a summary phrase.
Popular rules that failed
Several checklist items did not survive measurement:
- "AI sentences are all the same length." No difference once the sentences were split correctly; an early figure of 51× came from counting Markdown tables as single sentences.
- "AI overuses metaphor and rhetorical questions." People used figurative language 2.4× as often, and questions in body text 17× as often.
- "AI avoids particles like 就, 很, 了." True, but "add particles" cannot be checked, so the authors dropped the rule.
- "值得注意的是 and 扮演角色 are AI giveaways." Both appeared too rarely in generated text to count.
Different models, different tells
The study's clearest warning is that AI tone is not one thing. In its data DeepSeek used about 5.2 em dashes per thousand characters and Claude 4.3, while GPT used 0.1. Ordinal subheadings peaked with Claude, interrogative subheadings with Gemini. Judging a text by "what model X sounds like" is unreliable; it is better to look at the specific sentence.
How to use this when you edit
- Fix the paragraph openings first. If a paragraph starts with a judgment, add a demonstrative or the noun it refers to.
- Cut cue colons and the line that only leads into a list. Say the point in the sentence instead.
- Rewrite "not A but B" when there was never an A. State B directly.
- Keep your numbers and examples. Do not trade a figure for a general phrase.
- Leave alone what people also write. Metaphors, questions and parallel phrases inside a sentence are normal human Chinese.
RewriteMax can rewrite text into Simplified Chinese. As with any automated rewrite, check the result against your draft before you use it.