The Tells of AI-Written Chinese

Most overseas brands worry about the wrong audience. The question we get is whether a platform will detect machine-written copy and bury it — a worry we took apart in will an AI-written Xiaohongshu note get you flagged. The audience that actually penalises you has no detection tooling at all. Your judge is the reader, holding a phone, giving your note somewhere between one and three seconds.

The short version: readers do not diagnose your copy, they disengage from it. The signal they pick up is register — formal written Chinese sitting where conversational Chinese belongs — followed by stacked four-character idioms, symmetrical clauses, superlatives with nothing specific under them, and an ending that summarises instead of stopping. None of this is exotic, and a native speaker with twenty minutes and permission to break the draft can fix all of it.

Nobody thinks "AI". They think "nobody's here"

This matters because it changes what you are looking for. A reader who suspected a machine might at least be curious. What happens instead is duller: a half-second of flatness, no reason to keep reading, thumb moves. The reaction never arrives in your analytics as a complaint. It arrives as notes that collect polite saves and produce no enquiries — easy to misread as a reach problem, and expensive to solve as one.

The underlying judgement is about presence. Chinese platform users are unusually good at reading whether a real operator stands behind an account, because they have little else to go on — the dynamic we covered in what Chinese consumers trust instead of reviews. Copy that reads institutional answers the presence question badly. It says: this was produced, not written.

The tells, roughly in the order a reader hits them

Register mismatch. The largest one by a distance. Chinese language models default to formal written register — the 书面语 of reports and notices. Connectives like 首先 / 其次 / 综上所述, framings like 值得注意的是, and the polite 您 in a lifestyle note all belong to a bank letter or a property-management announcement. Dropped into Xiaohongshu, where people write approximately how they speak, the effect is a stranger in a suit at a dinner party. Nothing is wrong with any single word. The whole thing is in the wrong room.

Four-character idioms, stacked. 美轮美奂, 流连忘返, 心旷神怡, 不虚此行. One of these in a paragraph passes without comment. Three in a row is the clearest single tell in travel copy, partly because models reach for them as filler and partly because a real traveller who felt 心旷神怡 would have described the moment rather than named the feeling.

Symmetry. 不仅……而且……, 既……又……, clause pairs of matching length and shape, paragraph after paragraph. Human writing is lumpy: a long sentence, then a short one, then a fragment. Machine writing is suspiciously even.

Superlatives resting on nothing. 超值, 性价比超高, 强烈推荐, 体验感拉满 — with no price, no duration, no name, no weather, and nothing that went wrong. The enthusiasm is unattached. Readers do not consciously audit for specifics, but they feel the absence, because every note they trust has them.

Missing 语气词. Conversational Chinese leans hard on sentence-final particles — 吧, 啊, 呀, 嘛, 哦. Model output runs noticeably particle-poor, part of why the text reads stiff even when the vocabulary is right.

Translation residue. If your Chinese started as English, a separate set of marks survives: chains of 的, pronouns spelled out where Chinese would simply drop them, 被 passives in places Chinese prefers an active verb, and a stray 一个 doing the work of an English article. We looked at the higher-stakes version of this in where AI translation stops being good enough; these ones are cosmetic by comparison, and they still make the text feel imported.

The ending. Machine text summarises. It restates the paragraphs above and closes the loop, the way essays do. Real notes stop — mid-thought, on a question, on a detail, sometimes on nothing at all.

The mismatch is the problem, not the features

Worth being precise here, because the list above can be over-applied. Formal register is correct in a formal place. An official announcement from a verified service account should read composed; four-character idioms are ordinary good Chinese; symmetry is a genuine feature of the language, not a defect. A human writer can produce every item on that list and often does.

What readers react to is the gap between the register of the text and the register of the room it appears in. The same paragraph that reads wrong in a Xiaohongshu note reads perfectly fine as a WeChat service notification. So the useful question is never "does this look AI-written" but "does this sound like a person who belongs on this platform".

What fixes it

Not a detector. The tools that claim to score Chinese text for machine authorship are unreliable in both directions, and optimising against a score produces text strange in new ways.

What works is duller and cheaper:

• Give the model the things it cannot invent. Your actual prices, durations, pickup points, vehicle, the fact that the trail floods in March. Specifics are the fastest cure for unattached enthusiasm, and they cannot be generated — they have to come from you.
• Have a native speaker rewrite the opening and the closing. Tells cluster at both ends: the intro reaches for formal register to establish itself, the ending reaches for a summary. Fixing two sentences and one paragraph removes most of the signal.
• Decide register before drafting, not after. 你 or 您, particles or none, idioms or plain description. Passing that decision to the model means accepting a formal default.
• Read the draft aloud. Stiffness becomes audible well before anyone can describe it, and this catches symmetric clauses immediately.
• Leave one rough edge. A short fragment, an aside, an admission that the queue was long. Polish is not what earns trust on these platforms; evidence of a person is.

What this post cannot tell you

No detection rate, and no threshold. Anyone offering you a percentage for how often Chinese readers spot machine text is guessing, and the honest answer varies by platform, category and how much the reader already trusts the account. A note from an account someone follows gets read generously; the same note from a cold account gets read for reasons to leave.

The other limit is measurement. Nobody can attribute an enquiry that never arrived, so there is no clean before-and-after here — the cost of flat copy shows up as an absence, and absences do not appear in dashboards. What you can do is have a native reader tell you, honestly, whether the last ten things you published sound like a person.

Where CN1X fits

We write and run the Chinese, which means this is our problem rather than yours. Drafts go through a native reviewer before anything is published, and the brief we work from is your specifics — prices, inclusions, the awkward details — because nothing else stops copy reading generic. The content and channel work sits under services, alongside the Mini Program and account setup it feeds.

If you have Chinese copy running now and you are unsure how it reads to the people you are trying to reach, send it over and we will tell you plainly.