A Chinese AI detector without a score

The output is a written analysis of the supplied Chinese text: which passages carry formulaic phrasing, where sentence rhythm is unusually regular, which transitions add nothing, and what each observation does and does not support. It is deliberately not a probability, and the reason is worth stating plainly rather than burying in a disclaimer.

Mechanically, this route sends the text to our server, where a language model performs the reading and returns findings. That makes it a server tool with the constraints that implies: the text is transmitted, the input must be between 1 and 12,000 characters, and you get 5 free runs per rolling 24 hours without an account. The local tools on this site are a different category, and this is not one of them.

Every commercial Chinese AI detector we looked at returns a number. The number has three problems and none of them is fixable by making the number more precise.

It is unfalsifiable at the document level. A single figure covering a mixed document cannot be checked against anything, because there is no passage it is about. It also invites a threshold, and a threshold invites a decision, which is how a soft signal turns into an accusation.

It carries no reasoning. A reader given 87% learns nothing about what the text did to earn it, so the only available response is to trust or distrust the vendor. Findings can be argued with, which is the property you want in anything used near a grade or a contract.

And it implies calibration that has not been demonstrated. Reported accuracy figures in this market describe a vendor's own evaluation on a dataset of their choosing. We publish none, because we have measured none, and borrowing one would describe somebody else's system.

Each run covers one document, so segment long material before you begin.

Step 1

Paste Chinese text, simplified or traditional, between 1 and 12,000 characters. Anything longer is rejected before the run rather than truncated silently.

Step 2

Run the analysis. The text goes to our server, the model reads it, and the findings return as prose with the limitations attached.

Step 3

Read each finding against the passage it names. Where you disagree, the disagreement is the useful part, because it tells you the signal was structural rather than decisive.

Research on this problem is unusually visible in the search results for the term, and it converges on the same three difficulties. Written Chinese carries no orthographic word boundaries, so a method that tokenises into words has to segment first, and the segmentation choice is itself a modelling decision that affects the answer. Characters are also semantically dense and frequently ambiguous outside context, which compresses the variation that surprise-based measures depend on. And the register conventions of formal written Chinese are tighter than in English, so prose that would read as unnaturally uniform in English is closer to normal here.

The consequence is that the signals worth reporting are structural rather than statistical: sentence-length variance across a whole piece, reliance on four-character set phrases in positions where a writer would particularise, paragraph openers drawn from a narrow set of formal connectives, and a register that does not move when the subject does.

Every one of those can be produced by a disciplined human writer following a template. That is not a weakness in the description, it is the accurate state of the problem, and it is the reason the output is a description.

It cannot establish authorship. Nothing observable in prose distinguishes a person imitating a register from a model producing it, and the classifier behind this route is instructed not to claim otherwise.

It cannot attribute a model. Claims that a given text came from one specific system, rather than another, are not supportable from the writing, and this tool does not make them.

It cannot anticipate a third-party result. We have no affiliation with Zhuque, GPTZero, Turnitin, Copyleaks or Originality.ai, no access to their models and no basis for predicting what they will report on the same text.

It does not check originality against sources, so it says nothing about plagiarism. It does not run offline, it does not accept images, and it does not process batches. Each of those is a real constraint of a server-side, single-document route rather than a roadmap item.

Treat a finding as a hypothesis about the prose, testable by reading the passage it names. If the passage really is interchangeable with a paragraph on another subject, the finding is describing something true about the writing, whatever produced it. If it is specific and grounded and the finding still fires, the signal was structural and you can discount it.

Mixed documents are the common case and the reason prose output matters. Part drafted, part written, part edited is what most real text looks like, and a description can locate the boundary where a score cannot.

None of this should carry a decision about a person on its own. The defensible use is to identify passages worth a conversation, and to conduct the conversation about the writing rather than about the result.

Publishers commissioning Chinese copy, where the operative question is whether the prose earns its length. Educators triaging a stack of essays toward the ones worth discussing. Localisation teams checking that a translated page has not flattened into the same shape on every section.

The most defensible use is on your own drafts, before anyone else reads them. The findings name weaknesses that are worth fixing regardless of provenance, and fixing them is editing rather than evasion. A separate Chinese humanizer on this site rewrites toward that goal, with the same refusal to promise anyone else's verdict.