Starting from January 1, 2026, the revised “Guidelines for Patent Examination” of the State Intellectual Property Office will officially come into effect, adding exclusive examination chapters for patents related to artificial intelligence, big data, and algorithms for the first time, filling the gap in AI patent examination rules that have long existed.

This article combines the official laws and regulations of the State Intellectual Property Office with practical experience in patent agency, systematically disassembling the changes in new regulations, examination exclusion zones, and enterprise layout strategies, providing practical references for technology companies to apply for AI patents.

---

1、 New regulations come to fruition: AI patent examination rules officially standardized

On November 10, 2025, the China National Intellectual Property Administration issued a revised version of the Guidelines for Patent Examination, adding Section 6, Chapter 9, Part II, to develop independent examination specifications specifically for invention patent applications that include features of artificial intelligence, big data, algorithms, and business rules. The new regulations came into force on January 1, 2026. This is the first time in China that systematic review rules have been customized for the AI track, marking a new stage in China’s AI intellectual property supervision.

The new regulations cover two types of cases:

All AI, big data, and algorithm related invention patent applications submitted after January 1, 2026

Existing cases submitted before the effective date but not yet concluded as of January 1, 2026

---

2、 Three core review red lines

After the implementation of the new regulations, the State Intellectual Property Administration has clarified three major non crossing examination red lines, which belong to the statutory reasons for rejection and are also the most concentrated content of examination opinions in recent substantive examinations.

（1） Red line for ethics and compliance pre review: reject any non compliant data, algorithms, or applications directly

Ethical review is set as a pre examination item for AI patent preliminary review, covering three major areas: data source compliance, algorithm fairness, and landing application security, in accordance with Article 5 of the Patent Law.

Firstly, the source of the training dataset must comply with the relevant provisions of the current Data Security Law and Personal Information Protection Law. The algorithm training data corresponding to the patent shall not illegally capture the biometric information of natural persons. The collection and use of sensitive personal information such as facial features, fingerprints, gait, and health privacy must be separately authorized by the information subject; The collection of image data in public places is only for public safety and legal purposes. Unauthorized collection for commercial marketing, user profiling, and other scenarios is a typical violation.

Secondly, the algorithm logic must not embed discriminatory rules. Patent protected AI algorithms cannot rely on the applicant’s place of origin, gender, age, or physical characteristics to set differentiated judgment criteria. High frequency landing areas such as financial risk control, job screening, and autonomous driving are the focus of examination.

Thirdly, high-risk AI landing scenarios must be accompanied by comprehensive risk management and control plans. High risk applications such as autonomous driving, medical AI diagnosis, and credit risk control require patent documents to specify abnormal handling and risk mitigation mechanisms. If it cannot prove that the application risk is controllable, it will be directly judged as violating public interests and losing the authorization basis.

Example 1 (Privacy Infringement): A “big data mall mattress sales assistance system” uses cameras and facial recognition to collect facial features and identify identities without consumers’ knowledge, for preference analysis and precision marketing. The review conclusion is that this plan is not necessary for maintaining public safety and does not demonstrate the legality and compliance of data collection. It clearly violates the restrictions on facial recognition in public places under laws such as the Personal Information Protection Law, and therefore constitutes an illegal application of technology. According to Article 5.1 of the Patent Law, this application cannot be granted a patent right.

Example 2 (Ethical Bias): An invention of an “autonomous vehicle emergency decision-making model” that considers pedestrian gender and age to determine the protected and impacted objects in unavoidable collisions. This algorithm treats life differently based on gender and age in accidents, which has been found to violate the public’s ethical and moral concept of equal life, strengthen social prejudice, and raise public safety concerns. Therefore, the invention contains content that violates social morality, and according to Article 5 (1) of the Patent Law, a patent shall not be granted.

The above regulations and examples emphasize the necessity of introducing legal and ethical scrutiny in AI and data related inventions. Its legal basis is Article 5 (1) of the Patent Law, which prohibits inventions that are illegal or harmful to the public interest. The institutional intention is to guide innovation towards ethical and legal development, ensuring that patent protection does not endorse unethical technologies. The impact on application practice is significant: R&D personnel and applicants should self check whether there are any illegal or irregular aspects in data acquisition and algorithm decision-making of the invention scheme, and should also reflect compliance with legal and ethical requirements when writing patent applications. For example, inventions involving facial recognition and big data decision-making should specify the legality and compliance of data sources and usage in the specification, otherwise there is a risk of patent application rejection.

（2） Creative review red line: Simple scene transplantation without technical improvement, lacking authorized creativity

The new regulation completely abandons the previous loose judgment standard that “general algorithms have creativity when landing in new industries”, which is also the reason for the highest proportion of AI patent rejections at present.

When determining creativity, the examiner should base their judgment on the principle of overall evaluation of the complete technical solution, and break down the collaborative relationship between algorithm modules, hardware carriers, and practical applications. Only by making substantial technical improvements to existing technical pain points and bringing quantifiable technical effects can they serve as a creative support point.

Officially, there are four criteria for determining creativity:

It is not allowed to evaluate the algorithm features separately in isolation;

General algorithm optimization that can only solve common usage problems is not considered creative;

Directly calling open source pre trained large models and replacing business data with conventional neural network frameworks is not considered effective innovation;

Algorithm optimization must be closely integrated with hardware structure and landing scenarios to enhance efficiency.

（3） Object+Full Disclosure: Pure algorithms are not protected by patents, and all black box models are rejected

This is the basic threshold for AI patent applications, and the new regulations once again clarify the bottom line requirements, preventing confusion from both protecting the object and fully disclosing it, which is also the easiest link for beginners to fall into.

Object requirement: Simple mathematical algorithms and computational ideas without hardware carriers do not belong to the protected object of invention patents. The compliance plan must form a three in one structure of “algorithm improvement+hardware unit+implementation scenario”.

Full disclosure requirement: “Black box” writing is not allowed. The manual must disclose details such as model architecture, training parameters, input/output logic, etc., to ensure that technical personnel in this field can reproduce them.

---

3、 Enterprise practical response plan

In the new policy environment of strict review of AI patents, enterprises need to implement reforms from four dimensions: research and development, writing, inventory, and institutional aspects, covering the entire cycle from technology approval to patent certification.

（1） R&D end – pre control, conduct patent compliance pre-approval before project approval
Ethical verification: New project approval will be synchronized with data source compliance verification. Purchase agreement and user authorization documents for the retention of training data used in commercial AI; Privacy data related to facial and biometric features should be properly anonymized and archived, and unauthorized access is strictly prohibited. Risk control, recruitment, and autonomous driving algorithms are pre validated to eliminate discriminatory logic, high-risk AI supporting risk management plans are implemented, and validation records are retained.

Object and creativity screening: IP personnel intervene in R&D review in advance to screen out pure algorithm and commercial operation rule based solutions. Projects that only apply open source models without kernel parameter optimization are temporarily suspended from reporting inventions, and research and development are urged to implement hardware and model structure improvements, forming a complete technical solution of “algorithm optimization+hardware carrier+landing scenario”, and retaining comparative experimental data between new and old solutions.

（2） Writing end – standardized writing that conforms to new regulations
The instruction manual has added a separate chapter on ethics and compliance, actively disclosing data sources, desensitization methods, and risk control measures to reduce ethical questioning and review opinions from the source.
The entire article follows the logic of existing technology pain points, optimization solutions, and actual test results, with all creative improvements accompanied by comparative test results.
The model content should eliminate black box expressions, fully disclose key parameters such as network architecture, loss function, dataset partitioning, learning rate, and inference process, and meet the requirements of full disclosure.
The claims are based on systems, devices, and equipment, and algorithm improvements are limited to hardware operating logic and do not separately protect abstract algorithms.

（3） Inventory end – self inspection and remediation of patents under review in batches
For AI patents submitted before January 1, 2026 that have not yet been closed, check each item against the three major red lines:

For data compliance defects, supplement data authorization and desensitization certificates, and complete ethical explanations in the response; Patents that are only applicable in scenarios and lack substantive innovation should be appropriately divided or withdrawn to avoid wasting application fees due to rejection;
If there are black boxes or missing parameters in the instruction manual, the model experimental data should be supplemented in advance, and the review comments should be promptly responded to for correction. Qualified core patents will go through the pre examination channel of the local protection center and apply for priority examination to accelerate authorization.

（4） Institutional end – Establishing a long-term IP management mechanism
Customize AI exclusive technical disclosure templates with built-in modules for ethical information, model parameters, and experimental data filling, standardize R&D submission content, and eliminate paper-based disclosure; Establish a normalized linkage mechanism between R&D and IP, conduct regular training on new patent regulations for R&D, and have agents simultaneously deepen their knowledge of AI model technology, breaking down legal barriers between technology and patents;
Optimize patent layout ideas, cut down on low-quality and quantity based applications, and focus on core business technologies to create a gradient layout of core patents and peripheral patents; Establish a compliance ledger for enterprise data, archive all AI project data sources, test records, improvement and debugging documents, and respond to OA at any time to retrieve supporting materials.

---

4. Conclusion

The threshold for AI patent examination has been clearly raised. For enterprises, instead of passively waiting for rejection notices, it is better to proactively prepare for compliance from the four stages of project initiation, writing, self-examination, and system implementation. This is not a multiple-choice question, but a mandatory answer. Only enterprises that lay out in advance and operate in a standardized manner can truly use patents to build their own moat in the next round of competition.

As a professional intellectual property service organization deeply involved in the fields of biomedicine and mechatronics, Zhizhongde could have provided full process support for enterprises from patent layout to examination response, helping science and technology innovation enterprises build strong technical barriers under new regulations.

---

About Us

MeritsIP: Your Global Intellectual Property Partner

At MeritsIP, we provide comprehensive IP services in biomedicine, medical devices, manufacturing, semiconductors, and AI. We support 200+ top clients globally with strategic, high-quality IP services.

Stay Connected

• • Website: https://www.meritsipglobal.com

• • LinkedIn: Follow us on LinkedIn

• • Email: meritsip@meritsandtree.com

© 2025 MeritsIP. All Rights Reserved.