On 13 November 2025, China’s National Intellectual Property Administration (CNIPA) issued Order No. 84, publishing amendments to the Guidelines for Patent Examination which are due to take effect on 1 January 2026. Among other changes, the amendments revise Part II, Chapter 9, Section 6 (“Provisions on the Examination of Invention Patent Applications that Contain Algorithmic Features or Business Rules and Methods”) of the Guidelines, adding two new examples (Examples 18 and 19) to clarify the inventive step assessment for AI-related inventions.
The 2025 amendments adopt a more structured and, in practice, often stricter approach to assess inventive step for AI inventions than the prior version. They emphasize that algorithmic features (e.g., neural networks or data-processing rules) contribute to inventiveness only when they “functionally support and interact” with technical features (e.g., hardware or software structures) to solve a technical problem and achieve a non-obvious technical effect. The two new examples clearly delineate cases where distinguishing features do, or do not, meet this threshold.
Example 18
This example describes a method for counting ships, which involves acquiring a dataset of ship images, pre-processing the images, marking the positions and boundaries of the ships, dividing the dataset into training and testing sets, training a detection model using deep learning, and determining the actual number of ships from the model output. This is also known as a typical approach to identifying objects using deep learning.
In the hypothetical prior art, a similar method is applied to count fruits on trees, with essentially the same image acquisition, annotation, model training and counting steps. The only difference is that the objects are fruit rather than ships.
Although differences in target appearance, size, and imaging environment may, in practice, necessitate some model adjustments, the claimed method introduces no modifications to the deep learning process, network structure, training method, or parameters. The distinguishing feature is thus merely the change in application object, without any functional interaction yielding a new technical effect.
Therefore, the claimed method in example 18 lacks an inventive step. It is clarified by the guidelines that simple substitution of the processed object, without substantive algorithmic adaptations, does not confer inventive step.
Example 19
Example 19 presents a positive case which involves establishing a convolutional neural network model for grading scrap steel during collection and storage, where pieces are often disordered and overlapping.
Compared to the closest prior art, which uses feature extraction and training on pre-categorised images of scrap steel, the claimed method differs in terms of the training data used and the features extracted, such as colour, edges and texture. It also differs in terms of the specific adjustments made to the number of convolutional and pooling layers, as well as their hierarchical configuration.
The technical problem solved by the claimed invention is how to improve grading accuracy when dealing with complex, overlapping images of irregular scrap steel. These adjustments are specifically tailored to the challenging imaging conditions, they support and interact with the technical features of the solution, resulting in a significant improvement in grading accuracy. As these distinguishing algorithmic contributions are not disclosed by the prior art and produce a non-obvious technical effect through synergy with the technical features, the method of example 19 is considered to be inventive in a holistic assessment.
The 2025 amendments thus raise the bar: Adaptations to AI models (e.g., for new fields or tasks) must demonstrate verifiable functional synergy and non-obvious technical advancements, such as demonstrable improvements in accuracy, resilience, or efficiency. Simple data swaps or minor tweaks without such effects risk rejection as obvious.
This holistic focus on technical contributions seems to align broadly with the European Patent Office (EPO) approach, which requires a "technical effect" under the problem-solution framework. However, China's guidelines place greater explicit emphasis on functional synergy between algorithmic and technical elements, providing dedicated examples to enforce substantive integration over superficial applications.