Will an AI-Written Xiaohongshu Note Get You Flagged?

Somebody on your team drafts a Xiaohongshu note with a model, and the room divides. Half want to know whether the platform can tell. The other half want to know whether it matters if it can.

Asked as one question it has no answer. Split in two, both halves turn out to be answerable, and neither turns on whether a classifier catches you.

The short version: the first half is a declaration duty — the platform gives you a control for saying that content was AI-made, and you use it. The second half is distribution, and there the platform has no need to detect anything: it watches what readers do, and machine-drafted notes usually fail on thinness and sameness, which are visible in the numbers regardless of who wrote them. The real hazard is narrower and worth naming: a model will cheerfully write the absolute claim or the health sentence that gets a note pulled, and it has no idea it just did.

The declaration part, briefly

Since September 2025, anything published in China that a model generated or altered has to carry a label, and platforms give you a control for declaring it. We went through what the rule asks of an overseas operator, and what it does not, in China's AI-content labeling rule. The short of it: use the control, and understand that a Western tool leaves you carrying the duty alone.

Text drafted with a model sits in a greyer band than a wholly generated image, and the honest position is that the edges are not crisp. Declaring costs you nothing that matters. Not declaring, when you should have, is the kind of small thing that becomes a large thing once anybody is looking at your account for another reason.

Nobody has to detect you

The framing shifts here. Distribution on Xiaohongshu is decided by what readers do with a note — whether they stop, read to the end, save it, come back. The platform does not need a verdict on authorship to bury a note. It only needs the reading behaviour, and that arrives on its own.

So the question "can they tell" is close to irrelevant. What matters is whether the note has the properties that earn a read, and machine-drafted notes tend to fail on two:

Thinness. A model writes what is true of the category. It cannot know that the second stop has no shade at two in the afternoon, or that the coffee at the third one is the reason regulars go. First-hand specifics are the whole currency of the platform, and a draft produced without a person who was there has none.

Sameness. Give the same tool the same brief and it produces the same shape: an opening hook, three tidy points, a closing line that asks for a save. Readers recognise the pattern long before any system does, and a note that feels like the last eleven they scrolled past gets scrolled past too.

A human writing from a template produces the same result. That is the useful way to hold this: the platform is not punishing AI, it is punishing generic, and AI makes generic cheap to produce at volume.

The removal reasons have not moved

Suppression and removal are different events with different causes, and nothing about AI drafting added a new removal category. What gets travel notes pulled is what it always was: contact details, a paid collaboration nobody declared, images belonging to somebody else, absolute claims, and the stricter regime around health and medical wording. We set those out in why notes get removed.

The specific AI hazard lives inside that list rather than beside it. A model, asked for a persuasive note, will reach for the superlative. It will write "the best view in the region" or "guaranteed to see the northern lights" without hesitation, because those are normal marketing sentences in the English it learned from. In Chinese advertising terms they are a different matter, and a person who knows that would have caught it.

That is the concrete risk, and worth stating plainly: the danger is not that a model wrote your note. It is that a model wrote the one sentence in it that breaks a rule, and that sentence looks entirely reasonable to everybody on your side who cannot read Chinese well enough to flinch.

The account carries more risk than any note

Fifteen notes a week, all the same length, all the same structure, all posted at the same hour, is a pattern. Not because a system concluded they were machine-made, but because an account that produces interchangeable content at volume teaches the platform what to expect from the next one.

This compounds quietly, and undoing it takes far longer than fixing any single note. Volume is the part operators get wrong when a tool makes drafting cheap; the correct response to cheaper drafting is more variety, not more posts.

You will not be able to tell what happened

When a note goes quiet, the cause is usually invisible. Suppressed for feeling machine-made, suppressed for being thin, or simply an ordinary flop — all three look identical from your side, and the platform explains none of them.

A two-minute test separates removal from poor distribution, and running it comes before any rewriting — the removals post above sets it out.

What nobody can give you is the weighting. People selling AI-detection scores for Chinese platforms are selling a number they cannot validate, because the ground truth lives inside a company that publishes none of it.

What to actually do

• Use the declaration control when the content falls under it.
• Put one thing in every note that only somebody present could know. A time, a name, a number, a small inconvenience. One sentence is enough to change what the note is.
• Vary the shape across notes. Different lengths, different openings, different numbers of points. Sameness is the signal, and the easiest one to remove.
• Read for absolute claims and health wording before publishing. Every time, by somebody who reads Chinese. This is the check that has an actual downside if you skip it.
• Do not scale the volume just because the drafting got cheap.

The honest edge of this

How Xiaohongshu weights any of this is not published, the enforcement moves in waves, and the platform is under no obligation to be consistent. Everything above comes from how the failures present rather than from anything the platform has confirmed, and a year from now some of it will be wrong.

What holds regardless: a note with nothing specific in it has nothing to offer a reader, and that was true before anybody used a model to write one.

Where CN1X fits

Notes we publish for you are planned by somebody who has gone through your product, and the specific detail each one turns on gets chosen before a word is drafted. Whatever does the drafting after that, the Chinese gets read for the claim that would get the note pulled.

We do not fabricate reviews, testimonials or screenshots, and we will not put a number on how much reach any of this buys you. If notes of yours have gone quiet and you cannot work out why, send us three of them — whether it is a rules problem or a thinness problem is usually clear from the set.