Take a hypothetical. A patent attorney in Shenzhen files a claim set for a retail analytics model. The invention works. The drawings are clean. The examiner's objection, when it comes, is not about novelty or inventive step — it is about whether the cameras collecting the training images were legal.
Since 1 January 2026, that objection is written into the rulebook.
What changed, and when
The revision is the Decision on Amending the Patent Examination Guidelines (《国家知识产权局关于修改〈专利审查指南〉的决定》), issued on 10 November 2025 as Order No. 84 of the China National Intellectual Property Administration (CNIPA), effective 1 January 2026. The draft had been out for comment from 30 April to 15 June 2025, and the final text was published in amended form.
This is the stage to be precise about. The Guidelines are not a statute. They are the examination manual CNIPA's own examiners apply, which makes them binding inside the patent office even though they do not create new rights for applicants. A separate document often confused with this one — the Guidelines for Patent Applications Relating to AI Inventions (trial) (《人工智能相关发明专利申请指引(试行)》), issued in December 2024 — is a policy interpretation for applicants, not an examination rule. One tells examiners what to do; the other tells applicants what examiners are likely to think.
The November revision touched a long list of areas, from sequence-listing fees to invalidation procedure. Two of them matter to anyone filing AI: a new ethics screen at the entry gate, and a tighter disclosure standard for model internals.
The ethics screen, in the rulebook's own words
The new section 6.1.1 of Part II, Chapter 9 reads that for an invention containing algorithm features or business-method features, if its data collection, label management, rule setting or recommendation decisions contain content that violates law, social morality (社会公德) or harms the public interest (妨害公共利益), the invention cannot be granted under Article 5(1) of the Patent Law.
Two things about that sentence are worth slowing down for.
First, the trigger is not the model. It is the pipeline — how data is collected, how labels are assigned, how rules are configured, how recommendations are made. A transformer with impeccable maths and an illegal camera is unpatentable.
Second, the examiner may now look past the claims. The revision adds that examination should, where necessary, also cover the content of the description (必要时应当针对说明书的内容进行审查), not the claims alone. An applicant who keeps the data-collection method vague in the specification no longer escapes scrutiny by keeping the claims narrow.
The two examples that explain the whole thing
CNIPA's official interpretation pairs the new rule with two worked examples, and they are unusually concrete.
Example 1 is a mattress retail assistance system for shopping malls. It installs cameras in public areas, captures customers' facial features without their noticing, and uses the identity-linked result to target sales pitches. The interpretation says the collection is not necessary for public security, the application shows no lawful basis such as separate consent, and the scheme therefore conflicts with the Personal Information Protection Law (个人信息保护法).
Example 2 is a method for building an emergency-decision model for autonomous vehicles. It trains on pedestrians' sex and age and, at the moment of an unavoidable crash, uses those attributes to decide whom to protect and whom to hit. The interpretation calls this a codification of bias that violates the basic value of equal respect for life, and therefore social morality.
Neither example is exotic. The first mirrors a category of Chinese retail AI that was common before the 2021 privacy law tightened consent rules. The second is the trolley problem, filed as a patent.
Disclosure: the black box has to open
The revision also adds drafting requirements for the description in section 6.3.1. Where an application involves building or training a model, the specification should clearly record the necessary modules, layers or connection relationships, plus the specific training steps and parameters. Where a model is applied in a field or scenario, it should record how the model combines with that field and how input and output data are configured to show the internal relationship.
Two new sufficiency examples show how strictly this bites:
- Example 20 allows a face-feature generation claim to stand even though the specification omits where a spatial transformer network sits inside the convolutional network — because a skilled person knows it can go anywhere.
- Example 21 refuses a cancer-prediction claim that feeds routine blood panels and facial image features into a screening model without disclosing which indicators and which facial features correlate with accuracy. Not enabled, therefore not sufficiently disclosed (充分公开).
The pattern: architectural omissions that a skilled person can fill are tolerated. Correlations that only the applicant's private data could reveal are not.
Inventive step: a new scene is not an invention
Two further examples, 18 and 19, address the most common failure mode in Chinese AI filings — repackaging a known model for a new industry.
- Example 18 refuses a method of counting ships using conventional deep learning, where the only difference from the prior art (which counted fruit) is the object recognised. Different objects, same pipeline, no adjustment to the model: no inventive step (创造性).
- Example 19 allows a neural network for grading scrap steel, where the applicants changed feature extraction to colour, edge and texture, adjusted convolution and pooling layers accordingly, and solved a different technical problem with a measurable effect.
The distinction is effort at the model level, not novelty of the sales pitch.
Honest limitations
I have read CNIPA's official interpretation of the revision and the reported text of the new sections, not the full consolidated Guidelines as a paginated document. The examples are examination guidance, not adjudicated cases; how an examiner applies Example 1 to a scheme that is only partly non-compliant is not something the text resolves. I have no data on rejection rates under the new section since January 2026, and I would not trust any number circulating without an official release. I also cannot assess how the Patent Reexamination and Invalidation Department will treat appeals, because the revision is too recent for a body of decisions to exist.
What readers can do now
- Audit the data story before the claim story. If your specification describes a capture method that would fail a privacy assessment, no amount of claim drafting fixes it. Ask what a reviewer would see on page 12.
- Write the training recipe into the description. Layers, parameters, steps. The revision converts "proprietary detail" from a trade-secret habit into a disclosure risk.
- Stop filing scene swaps. If the only difference from your own prior patent is the industry, the examiner has an example with your name on it.
- Separate the two CNIPA documents in your compliance file. The December 2024 AI guidelines are persuasive; the amended Guidelines are operative. They get cited differently in an office action.
The amendment's logic is easy to state and hard to practise: China will still grant patents on AI, but only on AI whose inputs a regulator could defend in public.
