CNIPA Issues New Examination Guidelines for AI Inventions (2026) | China Patent Eligibility

For innovators in artificial intelligence and machine learning, securing patent protection in China has historically been challenging due to ambiguous rules on subject matter eligibility. In 2026, the China National Intellectual Property Administration (CNIPA) has issued comprehensive new examination guidelines specifically addressing AI inventions. The guidelines clarify what constitutes a “technical solution” versus an abstract algorithm, establish disclosure requirements for AI models, and provide concrete examples of patentable AI inventions. This guide explains the key provisions, the concept of “technical character,” and practical strategies for drafting AI patent applications that satisfy CNIPA‘s new standards.

1. Background: Why New AI Patent Guidelines Were Needed

China has become a global leader in AI patent filings, ranking first in the number of AI-related patent applications worldwide. However, the rapid evolution of AI technologies outpaced existing examination rules, leading to inconsistent rejections. Many applications for AI-based inventions were rejected on the grounds that they claimed “mathematical methods” or “mental acts,” which are excluded from patent protection under Article 25 of China’s Patent Law. The new CNIPA guidelines (officially “Examination Guidelines for Inventions in the Field of Artificial Intelligence,” effective January 1, 2026) provide detailed criteria for assessing patent eligibility, technical character, and sufficiency of disclosure for AI inventions. These guidelines apply to all AI-related patent applications filed on or after the effective date.

2. Key Provision: The “Technical Character” Requirement for AI Inventions

Under Chinese patent law, an invention must be a “technical solution” to be patentable. This means it must use technical means, solve a technical problem, and produce a technical effect. The 2026 guidelines clarify how this applies to AI inventions. An AI-related invention is considered to have “technical character” if it meets any of the following conditions:

• The AI model is applied to a specific technical field (e.g., image recognition for autonomous driving, fault prediction for industrial equipment, or drug molecule screening). The application must be more than a general-purpose use; it must be tailored to a concrete technical domain.
• The AI model solves a technical problem that cannot be solved by purely manual or conventional algorithmic methods, such as improving processing speed, reducing computational resource consumption, or enhancing prediction accuracy in a noisy environment.
• The AI invention produces a technical effect beyond the mere output of data – e.g., controlling a physical system (robot, autonomous vehicle, manufacturing line), generating a technical parameter (compression ratio, signal-to-noise ratio), or improving hardware efficiency.
• The invention includes a technical feature in the data preprocessing or post-processing steps, such as sensor data acquisition, feature extraction using domain-specific knowledge, or hardware-specific optimization (e.g., for GPU or NPU).

By contrast, an AI invention that merely claims a generic algorithm (e.g., “a neural network for classifying images” without specifying the technical field or application) is likely to be rejected as lacking technical character. The guidelines provide the following examples:

• Patentable: “A method for detecting defects in semiconductor wafers using a convolutional neural network trained on near-infrared images, wherein the network architecture includes a custom attention layer that prioritizes edge features.”
• Not patentable: “A method for training a neural network using backpropagation” (pure mathematical method).

3. Sufficiency of Disclosure for AI Models – Training Data and Black Box Issues

A major challenge for AI patent applicants is meeting the “sufficiency of disclosure” requirement (Article 26.3 of the Patent Law), which mandates that the specification clearly and completely describe the invention so that a person skilled in the art can carry it out. For AI inventions, the guidelines introduce new rules:

• Training data must be described with sufficient detail. If the AI model relies on specific training data (e.g., labeled images, sensor logs, medical records), the specification must disclose the source, composition, and preprocessing steps of the data. If the data is publicly available, a reference is sufficient. If the data is proprietary, the applicant must describe its key characteristics (size, format, labeling method) to enable reproducibility.
• Model architecture and hyperparameters must be disclosed. The specification must include the network structure (number of layers, activation functions, connections), training algorithm (e.g., SGD, Adam), and hyperparameter ranges (learning rate, batch size, epochs). Generic descriptions like “using a deep neural network” are insufficient.
• For black box models or deep learning with millions of parameters, the applicant must provide a meaningful functional description of how the model achieves the claimed technical effect. This can be done through flowcharts, pseudo-code, or feature attribution explanations (e.g., attention maps).
• If the invention depends on specific hardware (e.g., an AI accelerator chip), the hardware architecture must be described to the extent necessary for reproduction.

The guidelines accept that fully re‑training an AI model may not be required for enablement, as long as the specification provides enough guidance for a skilled person to implement the invention without undue experimentation. However, applicants should avoid describing the AI model as a “black box” – CNIPA examiners will reject applications that merely state “the model is trained to achieve X” without any technical disclosure.

4. Patentable vs. Non‑Patentable AI Inventions – Examples from the Guidelines

The 2026 guidelines include several illustrative examples to help applicants distinguish between patentable and non‑patentable AI inventions.

• Patentable – AI for industrial process control: “A method for optimizing steel rolling temperature using a reinforcement learning model that adjusts roller speed and coolant flow in real‑time, with the model trained on historical sensor data from the rolling mill.” (Technical means, technical problem, technical effect).
• Patentable – AI for medical image analysis: “A system for detecting lung nodules in CT scans using a 3D convolutional neural network, wherein the network includes a custom attention module that highlights potential nodule regions, and the output includes a probability map overlaid on the CT image.”
• Not patentable – Pure mathematical algorithm: “A method for clustering data points using a k‑means algorithm optimized with a genetic algorithm.” (No technical application).
• Not patentable – Business method implemented by AI: “A system for recommending financial products using a user behavior prediction model.” (Solves an economic problem, not a technical one). However, if the recommendation is based on a specific technical constraint (e.g., power consumption or network bandwidth), it may become patentable.
• Patentable – AI with hardware co‑design: “A neural network inference accelerator implemented in an FPGA, wherein the accelerator includes a systolic array optimized for matrix multiplication, and the configuration data is generated by a training algorithm described in the specification.”

These examples show that the technical character requirement is the critical hurdle. Applicants should always frame the problem as technical and the solution as involving concrete technical means.

5. New Guidelines for AI‑Related Computer Programs (Software Patents)

AI inventions are often implemented as computer programs. China traditionally has a stricter approach to software patents than some other jurisdictions, but the 2026 guidelines create a more favorable environment for AI software. The key points are:

• If the AI program is part of a physical device or controls a physical process, it is patentable (e.g., software for an autonomous robot or a medical imaging device).
• If the AI program is purely a non‑technical software (e.g., a spreadsheet add‑in or a game AI), it may still be rejected. However, the guidelines introduce a “technical contribution” test: if the software solves a technical problem (e.g., reducing network latency or improving data compression), it may be patentable even without direct hardware interaction.
• Computer‑readable storage media claims for AI models are now explicitly permitted (e.g., “A non‑transitory computer‑readable medium storing instructions for executing the method of claim 1”). Previously, such claims were often rejected, but the 2026 guidelines accept them if the method itself is patentable.

6. Practical Drafting Strategies for AI Patent Applications in China

To maximize the chances of allowance under the 2026 guidelines, patent applicants (both domestic and foreign) should adopt the following strategies:

1. Start with a technical problem statement. Avoid framing the invention as improving an AI algorithm per se. Instead, describe the problem in technical terms (e.g., “low accuracy in detecting pedestrian crossings under low‑light conditions” rather than “improving neural network accuracy”).
2. Include specific hardware or data‑acquisition steps. Describe how the AI model receives data from sensors, cameras, or databases, and how the output controls an actuator, display, or other physical device.
3. Disclose model architecture and hyperparameters in detail. Even if you are filing for a broad invention, provide at least one concrete example with specific layer types, activation functions, training parameters, and data preprocessing steps.
4. Use flowcharts and pseudo‑code. Chinese examiners appreciate clear visual representations of the algorithm. Include a flowchart that maps the data flow from input to output, highlighting the technical transformations.
5. If the training data is critical, describe its characteristics. Mention the data source, size, labeling process, and any augmentation techniques. If the data is proprietary, file a separate “data appendix” (under confidentiality) if necessary.
6. Consider filing a Chinese translation of the AI‑specific terms carefully. Use the terminology from the CNIPA guidelines (e.g., “技术特征” for technical character, “充分公开” for sufficiency). Machine translations often introduce errors that can lead to rejection.

Foreign applicants are strongly advised to use a local patent agent with AI expertise. Many Western patent attorneys are unfamiliar with CNIPA‘s unique requirements for AI inventions, and a poorly drafted application may be rejected even if the invention is highly innovative.

7. Practical Compliance Roadmap for AI Patent Applicants

To successfully obtain a patent for an AI invention in China under the 2026 guidelines, follow this five‑step roadmap:

1. Conduct a prior art search focusing on both technical and AI‑specific databases (Month 1). Use CNIPA’s search system and commercial tools. Identify existing patents on similar AI methods, as well as technical literature.
2. Draft the specification with a clear technical problem and solution (Month 2). Work with a Chinese patent attorney to ensure compliance with the new guidelines. Include at least one detailed embodiment with hyperparameters and data processing steps.
3. Prepare claims in multiple independent categories (Month 2). Draft claims for a method, a system (apparatus), and optionally a computer‑readable medium. Use “technical character” language (e.g., “A method for controlling an autonomous vehicle…”).
4. File the application electronically via CNIPA’s portal (Month 3). Pay the filing fee (RMB 900 for an invention patent, plus translation fees if filing from abroad). Request expedited examination if the invention is in a high‑priority field (e.g., healthcare AI, autonomous driving).
5. Respond to office actions promptly (Months 6‑24). CNIPA’s examination for AI patents typically takes 12‑24 months. If an office action rejects claims on eligibility grounds, argue the technical character using examples from the guidelines. If necessary, amend claims to add technical features (e.g., a specific sensor or control step).

The total cost for an AI patent application in China (including attorney fees, translations, and official fees) typically ranges from USD 3,000 to USD 8,000, depending on complexity.

Summary: CNIPA‘s 2026 examination guidelines for AI inventions clarify that patent eligibility requires a clear “technical character” – the AI must be applied to a specific technical field, solve a technical problem, and produce a technical effect. Pure algorithms or business methods remain excluded. Sufficiency of disclosure demands detailed description of training data, model architecture, hyperparameters, and, where applicable, hardware implementation. The guidelines provide examples of patentable AI inventions (e.g., industrial control systems, medical image analysis, hardware accelerators) and non‑patentable ones (generic algorithms, business methods). To succeed, applicants should draft claims that tie the AI to concrete technical applications, disclose implementation details, and use local expertise. By following the practical roadmap, innovators can secure robust patent protection for their AI technologies in China‘s fast‑growing market.