Strategic Shift in China’s AI Patent Examination: Navigating the 2026 Revised Guidelines
Date: 23 January 2026
【Volume 164】
Effective January 1, 2026, the China National Intellectual Property Administration (CNIPA) will implement the revised Patent Examination Guidelines (the "Guidelines"). This revision introduces systematic updates to examination standards for emerging technologies, specifically Artificial Intelligence (AI) and Big Data. These changes directly address long-standing ambiguities in AI patenting and provide a clearer regulatory roadmap for applicants.
I. Beyond Technology: Ethics and Legality as New Patentability Thresholds
Under the revised Guidelines, technical innovation alone is no longer sufficient. Article 5(1) of the Chinese Patent Law is now a rigorous gatekeeping mechanism for AI inventions.
1. Data Compliance: The "Separate Consent" Requirement:
The Case: Precision Marketing System for Shopping Mall Mattresses—Personal Data Compliance as a Threshold for Patent Grant1
Based on Article 5(1) of the Patent Law of the People’s Republic of China, the revised Guidelines explicitly stipulate that where an AI invention involves data mining, analysis or decision-making mechanisms that violate laws, social morality, or harm public interests, such invention shall not be granted a patent.
A newly added example concerning a “big data-based sales assistance system for shopping mall mattresses” serves as a typical case. The invention collected facial images through cameras and facial recognition modules installed in shopping malls and conducted identity recognition and preference analysis without customers’ awareness, thereby achieving precision marketing.
The Guidelines point out that, under the Personal Information Protection Law of the People’s Republic of China, the installation of facial recognition equipment in public places shall, in principle, be limited to situations necessary for maintaining public security. In this case, however, the invention was used for commercial marketing purposes, and the application documents failed to disclose the acquisition of separate consent from customers or the existence of other legal bases. As such, the application was considered an illegal invention and was categorically ineligible for patent protection.
Practical Implications
The example demonstrates that for AI inventions involving personal data processing, compliance is no longer merely a legal risk during the later implementation stage; rather, it has become a potential ground for rejecting patent grant at the examination stage. In light of this, when drafting relevant applications, it is advisable that the applicant carefully consider whether it is necessary to disclose the legality of data acquisition in the specification.
2. Algorithmic Ethics: Value Judgments as Grounds for Rejection
The Case: Emergency Decision-Making Models for Autonomous Vehicles—Value Judgments May Also Negate Patentability2
Another newly added case involves an emergency decision-making model for autonomous vehicles. During the model training phase, the invention incorporated pedestrians’ gender and age as parameters for decision-making and, when accidents were unavoidable, determined the driving direction of vehicles based on such parameters.
The Guidelines indicate that this technical solution, in essence, took gender and age as bases for assigning different values to human life, embedding highly controversial value judgments into technical decisions. It violated the fundamental social morality that all lives are of equal value and therefore constituted an invention that contravened social morality as referred to in Article 5(1) of the Patent Law.
Practical Implications
Data compliance has shifted from a post-grant operational risk to a pre-grant examination hurdle. Applicants should consider disclosing the legality of data acquisition in the specification for AI inventions involving sensitive personal data.
II. Assessing Inventive Step: Substantive Improvement vs. Simple Migration
The Guidelines raise the bar for Inventive Step (Non-obviousness), targeting the practice of merely "re-skinning" existing AI models for new industries.
The "Fruit-to-Ship" Rule:
In the case of a “method for identifying the number of ships,” although the applicant applied a deep learning model to ship-image recognition, cited references had already disclosed the use of the same process to identify the number of fruits.3 The Guidelines note that, by merely changing the subject of recognition, without making substantive adjustments to the model structure, training method, or feature processing, the claimed invention involves only an obvious substitution that would have been readily made by a person ordinarily skilled in the art, and thus lacked an inventive step.
Technical Contribution:
Conversely, a "Scrap Steel Grading Model" was granted a patent because it featured specific adjustments to convolutional and pooling layers to address the unique technical challenge of "chaotic stacking" in steel piles.
Practical Implications
The focus of the "Inventive Step" assessment has shifted from where AI is applied to how the algorithm was substantively modified to solve a specific technical problem.
III. Opening the “Black Box”: Heightened Disclosure Standards
To address the inherent opacity of AI, the CNIPA has elevated the requirements for Sufficient Disclosure.
General Knowledge vs. Specific Correlation:
While standard architectures (like Spatial Transformer Networks) may rely on general knowledge, high-stakes claims (like "Cancer Prediction") require more.
In a case concerning a method for generating facial features, even though the specific position of a spatial transformer network within the model was not restricted, the disclosure was still considered sufficient as such configuration constituted the general knowledge commonly known in the art.4
On the contrary, in the case concerning the “prediction of cancer based on biological information,” the applicant failed to explain which blood indicators or facial features were substantively related to diagnosis of cancer, as well as to provide any verification data, therefore making the alleged technical effects remain speculative.5 As a result, the application was considered insufficiently disclosed.
Practical Implications
Technical effects must be substantiated, not merely alleged. Applicants must provide sufficient correlation data to ensure the invention is reproducible and the claimed effects are not speculative.
Wisdom Analysis: The High-Standard Era of AI Patenting
The 2026 Guidelines signal China’s transition toward a High-Standard Era for AI intellectual property, characterized by three pillars: Compliance, Substance, and Transparency.
Recommended Strategy:
1. Audit the Data Source: Ensure biometrics or personal data usage aligns with PIPL (separate consent).
2. Verify the Logic: Scrub algorithms of any discriminatory or ethically controversial parameters.
3. Document the "How": Focus the specification on algorithmic innovations and provide empirical data to support technical claims.
[1] See “China Patent Examination Guidelines,” Example 1.
[2] See “China Patent Examination Guidelines,” Example 2.
[3] See “China Patent Examination Guidelines,” Example 18.
[4] See “China Patent Examination Guidelines,” Example 20.
[5] See “China Patent Examination Guidelines,” Example 21.
