Tao CHEN
Chinese Attorney-at-Law
Wei Chixue Law Firm
Chinese Attorney-at-Law
Wei Chixue Law Firm
I. Introduction
In recent years, the volume of AI patent applications filed with the China National Intellectual Property Administration (CNIPA) has surged dramatically. In 2025, over 200,000 AI patent applications were submitted, covering both improvements to AI algorithms and applications of AI algorithms. By examining such a massive volume of AI patent applications, the CNIPA has gradually clarified and refined examination rules specific to AI patent applications. Drawing on the CNIPA’s Guidelines for Patent Examination (hereinafter “the Guidelines” for short) and Reexamination case precedents, this paper summarizes the examination standards applicable to AI patent applications in China.
II. Rules and Methods for Mental Activities
1. Provisions in the Guidelines
The Guidelines stipulate, in Section 6.1.1 of Chapter 9 in Part II, as follows:
If a claim covers abstract algorithms or pure commercial rules and methods without incorporating any technical features, the claim falls under rules and methods for mental activities as prescribed in Item (2), Paragraph 1, Article 25 (hereinafter “Article 25.1(2)” for short) of the Patent Law and shall be ineligible for patent protection.
If a claim contains technical features in addition to algorithm features or commercial rule/method features, the claim as a whole is not merely directed to a rule or method for mental activities, and thus shall not be excluded from patent eligibility under Article 25.1(2).
A determination of whether a claim falls under the category of rules and methods for mental activities is a straightforward task. A claim that covers purely abstract algorithms without applying to specific scenarios and without processing data of definite technical meaning can be determined as a rule/method for mental activities. If a claim is not directed to a mere rule/method for mental activities, a further examination of whether it sets forth a technical solution is required.
2. Case Analysis
Case 1: This method improves modeling accuracy by expanding the number of training samples. Claim 1 reads as follows:
A method of establishing a mathematical model, comprising:
training an initial feature extraction model using feature values from training samples of a first classification task and feature values from training samples of at least one second classification task, to obtain a target feature extraction model, wherein the second classification task is another classification task related to the first classification task;
processing, using the target feature extraction model, the feature values from each training sample of the first classification task to obtain extracted feature values corresponding to each training sample;
combining the extracted feature values and label values of each training sample to form extracted training samples, and training an initial classification model using the extracted training samples to obtain a target classification model; and
combining the target classification model and the target feature extraction model to form a mathematical model for the first classification task.
While the method of Case 1 is beneficial in overcoming overfitting caused by insufficient training samples and improving modeling accuracy, all its processing objects, procedures, and results are decoupled from concrete technical application scenarios. It merely optimizes abstract mathematical approaches without including any technical features, so it is deemed a rule/method for mental activities and ineligible for patent protection.
III. Technical Solutions
1. Provisions in the Guidelines
The Guidelines stipulate, in Section 6.1.2 of Chapter 9 in Part II, as follows:
In a determination of whether a claim containing algorithm features or commercial rule/method features defines a technical solution, all features recited in the claim shall be considered as a whole. If the claim recites technical means that utilize laws of nature to solve a technical problem and thereby achieve a technical effect conforming to laws of nature, the solution set forth in the claim qualifies as a “technical solution” under Article 2.2 of the Patent Law.
AI patent applications must satisfy the same three core elements (technical problem, technical means, technical effect) as general patent applications to qualify as technical solutions. For practical guidance, the Guidelines outline three typical scenarios (see Section 6.1.2 of Chapter 9 in Part II):
If the steps involving algorithms in a claim are closely correlated with the technical problem to be solved—for instance, the data processed by the algorithm has a definite technical meaning in the technical field, and the execution of the algorithm directly reflects a process of solving a technical problem using laws of nature and achieves a technical effect—then the solution set forth in the claim generally qualifies as a technical solution under Article 2.2 (Scenario 1).
If the claimed solution involves improvements to an artificial intelligence/big data algorithm such as deep learning, classification, or clustering, and the algorithm has a specific technical association with the internal structure of a computer system and is capable of solving the technical problem of enhancing hardware operational efficiency or execution performance, including reducing data storage, reducing data transmission, or increasing hardware processing speed, thereby achieving the technical effect conforming to laws of nature that the internal performance of the computer system is improved, then the solution defined in the claim qualifies as a technical solution under Article 2.2 (Scenario 2).
If the claimed solution processes big data from a specific application field, utilizing classification, clustering, regression analysis, neural networks, and similar techniques to mine intrinsic correlational relationships in the data that conform to laws of nature, thereby solving the technical problem of improving the reliability or accuracy of big data analysis in that specific application field and achieving the corresponding technical effect, then the solution defined in the claim qualifies as a technical solution under Article 2.2 (Scenario 3).
Scenario 1 relates to the integration of algorithms with application scenarios. To establish this integration, the claim must meet at least two conditions: (1) the data processed by the algorithm has a definite technical meaning in the pertinent technical field; and (2) the execution of the algorithm follows laws of nature.
Scenario 2 relates to the integration of algorithms with the internal structure of a computer system. In this scenario, a claim is required to meet three conditions to constitute a technical solution: (1) there is a technical correlation between the algorithm and the internal structure of the computer system; (2) it solves the technical problem of enhancing hardware operational efficiency or execution performance; and (3) it achieves the corresponding technical effect of improving the internal performance of the computer system.
Scenario 3 relates to data mining of big data in a specific application field. In this scenario, the algorithm and the specific application scenario are already integrated. The key to qualifying as a technical solution lies in whether the mined correlational relationships conform to laws of nature.
2. Case Analysis
(1) Analysis of Scenario 1
Case 2: Claim 1 reads as follows:
A method of training a CNN model, comprising:
acquiring initial model parameters for a CNN model to be trained …;
acquiring multiple training images;
deriving a first feature map for each training image at the respective convolutional layers;
performing horizontal pooling to generate, for each training image, a second feature map at the respective convolutional layers;
determining a feature vector for each training image based on the second feature map at the respective convolutional layers;
processing each feature vector to generate a category probability vector for each training image;
calculating a classification error based on the category probability vector of each training image and an initial class label of each training image;
adjusting parameters of the CNN model to be trained based on the classification error;
continuing the parameter adjustment based on the adjusted parameters and the multiple training images until the iteration proceeds a preset number of times; and
taking the model parameters obtained when the iteration proceeds the preset number of times as the parameters of the fully trained CNN model.
Case 2 processes images, which are data carrying definite technical meanings in the technical field. The means of performing different processing on images at different convolutional layers and training the model are technical means that follow laws of nature, and they achieve the technical effect of recognizing images of arbitrary sizes. Because the algorithm is closely integrated with the application scenario, Case 2 qualifies as a technical solution.
Counterexample: Case 3: Claim 1 reads as follows:
A transformer-based method of constructing a sample-efficient world model, comprising:
S0: reading historical interaction data and splitting it into historical action sequences and historical raw state data, where historical interaction data includes object state information, relative object position, and discrete action space;
S1: collecting real-time raw state data, converting it into multi-scale feature data, generating augmented state data, and deriving current discrete latent variables and historical discrete latent variables, respectively;
S2: constructing causal structural data, generating causal augmented representation data, and outputting state prediction data;
S3: building time-series correlation data, calculating uncertainty metrics, and generating multi-step prediction sequences;
S4: deriving hierarchical value metrics, computing policy gradients, and generating optimized policy data;
S5: generating updated interaction data, deriving high-quality sample data, producing updated model parameters, and outputting tuned model weights,
wherein, in an industrial robot scenario, the historical interaction data comprises: robot status information...; relative position information...; discrete action space...; process parameter data...; system status data...;
the raw state data comprises: sensor data...; environmental data....
Case 3 likewise processes data having a definite technical meaning in the technical field. For instance, the sensor data comprises “a torque sensor reading and a vision sensor image.” However case 3 was not deemed a technical solution. The Decision of Reexamination says as follows: The historical interaction data and raw state data in the industrial robot scenario recited in claim 1 are merely input sample data screened for constructing the world model. The claim fails to mention how the screening criteria and procedures for “high-quality sample data” are closely correlated with the industrial robot scenario.
To sum up, even though Case 3 recites an application scenario and data having definite technical meanings, the application scenario and algorithm remain disjointed. As can be seen from the specification, the algorithm does not solve a specific technical problem in the industrial robot application scenario, and the claims merely relate to optimization of the model algorithm itself.
(2) Analysis of Scenario 2
Case 4 concerns a training method for deep neural network models. Claim 1 reads as follows:
A method of training a deep neural network model, comprising:
when the volume of training data changes, calculating the training time consumed by the adjusted training data under each preset candidate training scheme respectively;
selecting the candidate training scheme with the minimum training time consumption as the optimal training scheme for the adjusted training data, wherein the candidate training schemes include a single-processor training scheme and a multi-processor training scheme based on data parallelism;
and
conducting model training on the adjusted training data under the optimal training scheme.
The algorithm in Case 4 bears a technical correlation with the internal structure of computer systems, as it selects the optimal training scheme from single-processor and multi-processor training options. This sufficiently demonstrates a technical linkage between the model training method and the internal architecture of computer systems. Also, such technical linkage can improve execution performance the hardware during the training. Therefore, the solution qualifies as a technical solution.
(3) Analysis of Scenario 3
Case 5 relates to a method of analyzing the propensity to use electronic coupons. Claim 1 reads as follows:
A method of analyzing the propensity to use electronic coupons, comprising:
classifying electronic coupons based on information of the electronic coupon to obtain electronic coupon categories;
acquiring user sample data based on application scenarios of the electronic coupons;
extracting user behavioral features from the user sample data based on user behaviors, wherein the user behaviors comprise: browsing web pages, searching for keywords, following, adding to shopping cart, purchasing, and using electronic coupons;
training, with the user sample data as training samples and the user behavioral features as attribute labels, a model for recognizing the propensity to use electronic coupons for different categories of electronic coupons; and
predicting, through the trained model for recognizing the propensity to use electronic coupons, the probability of electronic coupons being used, to obtain the user’s propensity to use different categories of electronic coupons.
The key to determining whether Case 5 qualifies as a technical solution lies in whether there exists a correlational relationship conforming to laws of nature between the user behavioral features such as “browsing web pages, searching for keywords, following, adding to shopping cart, purchasing, and using electronic coupons” and the user’s propensity to use electronic coupons. Longer browsing times, more frequent searches, and more frequent use of electronic coupons can reflect a higher propensity to use electronic coupons. Therefore, the correlational relationship mined in this solution conforms to laws of nature, and the solution based on the correlational relationship qualifies as a technical solution.
As a counterexample, if a method of predicting the price of a financial product mines the correlational relationship between the past prices of the financial product and its future prices, and predicts the price of the financial product based on that correlational relationship, then the price prediction method does not qualify as a technical solution because such a correlational relationship does not conform to laws of nature.
IV. Inventive Step
1. Provisions in the Guidelines
As is well known, the CNIPA uses the “three-step approach” in inventive-step determinations. AI patent applications are no exception. The unique consideration for AI patent applications is how to treat the substantial algorithmic features present in claims of AI patent applications.
The Guidelines stipulate, in Section 6.1.3 of Chapter 9 in Part II, as follows:
In a determination of whether a patent application that includes both technical features and algorithmic or commercial rule/method features involves an inventive step, algorithmic or commercial rule/method features that are functionally mutually supportive and interactive with the technical features shall be considered as a whole together with the technical features. The term “functionally mutually supportive and interactive” means that the algorithmic features or commercial rule/method features are closely integrated with the technical features, jointly constituting technical means for solving a technical problem, and are capable of achieving a corresponding technical effect.
As is clear from the above, if the algorithmic features in a claim are functionally mutually supportive and interactive with the technical features therein, they shall be considered as a whole, followed by a determination of whether the combined features are disclosed or suggested by the prior art or constitute common general knowledge. On the other hand, if algorithmic features do not meet the “mutually supportive and interactive” criterion, they should be disregarded in the inventive-step analysis.
When it comes to how to determine whether algorithmic features and technical features are “functionally mutually supportive and interactive,” the Guidelines present two specific scenarios (see Section 6.1.3 of Chapter 9 in Part II):
If the algorithm in a claim is applied to a specific technical field and can solve a specific technical problem, the algorithmic features may be deemed functionally mutually supportive and interactive with the technical features, and the algorithmic features become part of the technical means employed. In such a case, the contribution made by the algorithmic features to the technical solution shall be taken into account during the inventive-step determination (Scenario 1).
If the algorithm in a claim has a specific technical association with the internal structure of a computer system, achieving an improvement to the internal performance of the computer system and enhancing hardware operational efficiency or execution performance, including reducing data storage, reducing data transmission, or increasing hardware processing speed, the algorithmic features may be deemed functionally mutually supportive and interactive with the technical features. In such a case, the contribution made by the algorithmic features to the technical solution shall be taken into account during the inventive-step determination (Scenario 2).
Scenario 1 pertains to the integration of algorithmic features with application scenarios, while Scenario 2 pertains to the integration of algorithmic features with the internal structure of a computer system. In these two scenarios, algorithmic features closely associated with the technical features shall be treated as usual technical features, and whether the claim involves inventive step shall be determined by the “three-step approach.”
2. Case Analysis
Case 6: Claim 1 reads as follows:
A system for estimating hydroelectric generation of a dam, comprising:
a neural network, constructed by means of an information processor, the neural network having an input layer and an output layer, wherein input data for the input layer comprises precipitation upstream of the river, water flow upstream of the river, and water inflow into the dam during a reference time and a preceding period, and output data for the output layer comprises future hydroelectric power generation after the reference time;
a machine learning unit that trains the neural network using training data corresponding to actual values of the input data and output data; and
an estimation unit that inputs input data into the neural network trained by the machine learning unit, sets the current time as the reference time, and then calculates an estimated value of future hydroelectric power generation based on the output data with the reference time being the current time.
The closest prior art (D1) discloses a hydroelectric power generation estimation system that also estimates future hydroelectric power generation based on the three parameters “precipitation upstream of the river, water flow upstream of the river, and water inflow into the dam.” The difference is that claim 1 uses a neural network, whereas D1 uses a regression equation model.
In claim 1, using a neural network for hydroelectric power generation estimation involves algorithmic features that are functionally mutually supportive and interactive with the technical features. However, using a neural network model to perform future output estimation is common general knowledge in the art. Therefore, claim 1 does not involve an inventive step over D1 in combination with common general knowledge.
The dependent claim 2 of Case 6 has the additional feature “the input data for the input layer further comprises air temperature upstream of the river during the reference time and the preceding period,” which is not disclosed in the prior art. Whether this dependent claim involves an inventive step depends on whether the causal relationship—“changes in upstream air temperature lead to changes in precipitation and water flow influenced by snowmelt, which in turn affect hydroelectric power generation”—is common general knowledge.
The Guidelines provide instructive cases for certain special scenarios concerning AI patent applications.
Case 7: This case illustrates how inventive step should be assessed when an AI algorithm is applied to a different application scenario. Claim 1 reads as follows:
A method of counting ships, comprising:
acquiring a ship image dataset, preprocessing image information in the dataset, marking the position and boundary information of ships in the images, and dividing the dataset into a training dataset and a test dataset;
performing deep learning using the training dataset to construct a training model;
training the model using the test data input into the training model to obtain ship test result data; and
multiplying the ship test result data by a preset error parameter to determine the actual number of ships.
D1 discloses a method of counting fruits on trees. The sole difference between claim 1 and D1 is the different identification targets. While ships and fruits differ in appearance, volume, and environmental context, labeling ship data and fruit data on images to generate training datasets for model training entails no adjustments or improvements to deep learning, model construction, or training procedures. Accordingly, claim 1 does not involve an inventive step.
The takeaway from Case 7 is that when the improvement of a solution lies in the application of an algorithm to a scenario different from the existing scenario, the inventive-step determination should comprehensively consider factors such as how close the application scenarios are to one another, whether there is a technical motivation, how difficult the application of the algorithm to the different scenario is, whether technical hurdles are to be overcome, and whether unexpected technical effects are achieved. If the application to a different scenario does not require adjustments to algorithm or model training methods, parameters, configurations, etc., by overcoming technical difficulties, and does not produce unexpected technical effects, the solution cannot be deemed inventive.
Case 8: This case addresses how to assess inventive step when an AI algorithm is applied to the same or similar technical scenarios. Claim 1 reads as follows:
A method of constructing a neural network model for scrap steel grading, wherein the model is configured to grade collected scrap steel, comprising:
… performing convolutional neural network learning to form a grading neural network model with classification outputs;
… extracting color, edge and texture features of objects within images, as well as correlational features between object edges and textures, via output sets of multiple channels composed of convolutional layers or convolutional layers paired with pooling layers;
wherein extraction of object color and edge features within images is realized via output sets of three channels each composed of convolutional layers paired with pooling layers: the first channel with one pooling layer, the second channel with two convolutional layers, and the third channel with four convolutional layers; extraction of object texture features within images is realized, after aggregating the extraction results of object color and edge features, via output sets of three channels composed of convolutional layers: the first channel with one convolutional layer, the second channel with two convolutional layers, and the third channel with three convolutional layers;
the number of convolutional calculation channels for extracting correlational features between edges and textures is greater than the number of convolutional calculation channels for extracting object color, edge and texture features within images.
D1 discloses a convolutional neural network-based method for identifying scrap steel categories. The distinguishing features of claim 1 from D1 include different training data and extracted features as well as different channel counts and hierarchical configurations for convolutional and pooling layers.
In claim 1, the channel counts and hierarchical configurations of convolutional and pooling layers are adjusted during model training to extract color, edge and texture features of scrap steel within images. The algorithm features functionally support and interact with the technical features, and deliver improved accuracy for scrap steel grading. The distinguishing features are neither disclosed in other prior-art documents nor constitute common general knowledge in the art. Accordingly, claim 1 involves an inventive step.
The takeaway from Case 8 is as follows: even when a claim and the prior art target identical or similar application scenarios, the solution shall be accepted as involving an inventive step if it involves substantive adjustments to algorithm workflows, model parameters or other core elements that are not suggested by the prior art, and it solves a different technical problem and achieves advantageous technical effects.
V. Article 5.1 of the Patent Law
1. Provisions in the Guidelines
The Guidelines stipulate, in Section 6.1.1 of Chapter 9 in Part II, as follows:
If the data collection, labeling management, rule setting, decision-making recommendation, or other aspects in an application for an invention patent that comprises algorithmic features or commercial rule/method features involve content that violates laws, social morality, or impairs public interests, the application shall be excluded from patent eligibility under Article 5.1 of the Patent Law.
2. Case Analysis
Case 9: This case involves a violation of law. Claim 1 reads as follows:
A big data-based auxiliary sales system for mattresses in shopping malls, comprising mattress display equipment and a management center, characterized in that:
the mattress display equipment includes a control module and an information collection module, configured to display mattress products, assist mattress sales, and collect customer data; … the information collection module includes an image capture module and a facial recognition module, configured to collect customer facial feature information … to generate customer identity identification information;
the management center includes a management server and an analytical auxiliary system; the management server administers multiple mattress display devices; the analytical auxiliary system analyzes, based on the customer identification information and data collected by the mattress display devices, the customer’s true preferences, and feeds analytical results back to the management center.
Under Chinese law, the installation of image capture and personal identification devices in public places must be necessary for maintaining public safety, comply with relevant state regulations, and display conspicuous warning signs. Collected personal images and identification information may only be used for maintaining public safety and not for other purposes, unless individual consent is obtained.
The purpose of this application is targeted marketing of mattresses, not maintaining public safety. Moreover, facial capture and identification occur without the customers’ knowledge, and the specification does not indicate that data acquisition or information collection is lawful and compliant. Therefore, this application goes against the law and is not patentable.
A takeaway from Case 9 is that, in filing applications, it is preferable to include declarations in the specification regarding the legality and compliance of data sources, data content, data governance, training objectives and solutions, training procedures, training environments and tools, and ethical validation of training results.
Case 10: This case relates to a method of building an emergency decision model for autonomous vehicles, which uses the gender and age of pedestrians as obstacle data. The trained decision model determines protected objects and collision targets when obstacles cannot be avoided. Human life has equal value and dignity. Making such choices based on pedestrian gender and age clearly violates prevailing ethical norms. Therefore, the application contains content contrary to social morality and and shall be denied patent protection.
A takeaway from Case 10 is that content contrary to social morality should be avoided in application documents.
VI. Sufficiency of Disclosure
1. Provisions in the Guidelines
The Guidelines stipulate, in Section 6.3.1 of Chapter 9 in Part II, as follows:
If an application involves the construction or training of an AI model, the specification generally needs to clearly record the necessary modules, layers, or connection relationships of the model, as well as the essential steps, parameters, etc., for training. If an application involves applying an AI model or algorithm to a specific field or scenario, the specification generally needs to clearly record how the model or algorithm is combined with the specific field or scenario, and how the input and output data of the algorithm or model are configured to indicate their inherent correlational relationships, such that one skilled in the art can, based on the content recorded in the specification, implement the claimed solution.
As is evident from the above, applications involving the construction or training of AI models and those involving the application of AI models or algorithms to specific fields or scenarios should meet different requirements for sufficiency of disclosure.
2. Case Analysis
Case 11: This case relates to a method of generating facial features, which realizes information sharing between multiple second convolutional neural networks via a set of feature region images generated by a first convolutional neural network integrated with a spatial transformation network, thereby reducing memory resource consumption while improving accuracy of facial image generation results.
Claim 1 recites a first convolutional neural network and second convolutional neural networks, with a spatial transformation network embedded within the first convolutional neural network for identifying feature regions of facial images. The specification says, “the specific installation position of the spatial transformation network within the first convolutional neural network is not limited in the present application.”
As known to one skilled in the art, the spatial transformer network, as a whole, can be inserted at any position within the first convolutional neural network, and its position does not affect its capability to recognize feature regions of images. Convolutional neural networks and spatial transformer networks are generally known algorithms, and one skilled in the art would be able to construct the corresponding model architecture. Therefore, the claimed solution is sufficiently disclosed in the specification.
The takeaway from this case is that when solving a technical problem in a specific field using AI algorithms or models, if the technical means employed in the claimed solution are not explicitly recorded in the application documents but belong to common general knowledge for solving that technical problem in the field, the absence of a record of such technical means in the specification does not lead to insufficient disclosure of the claimed solution.
Case 12: This case relates a method of predicting cancer based on biological information. The specification describes a trained malignant tumor screening model that can give a malignant tumor risk prediction value by using routine blood test indicators, routine biochemical test indicators, and facial image features as inputs, thereby solving the technical problem of improving the accuracy of malignant tumor prediction. Claim 1 sets forth the model training process and the subsequent prediction process.
However, the specification does not say, nor can one skilled in the art determine, which specific routine blood and biochemical indicators are useful for determining malignant tumors. Moreover, aside from a few types of tumors such as facial skin cancer, current scientific knowledge does not make clear the association between facial features and malignancy, and the specification neither records nor demonstrates causal relationships between the “determining factors” and the “determination result.” Additionally, there is no validation data in the specification. One skilled in the art would not be able to determine, solely from the disclosure in the specification, that the claimed solution can solve the stated technical problem. Therefore, the claimed technical solution is not sufficiently disclosed in the specification.
The takeaway from Case 12 is that in a case where AI algorithms or models are used to solve a technical problem in a specific field, if the means employed (including the inherent correlational relationships between inputs and outputs) are neither explicitly recorded in the application documents nor common general knowledge, and one skilled in the art would not be able to implement the claimed solution and solve the corresponding technical problem based on the specification, then it should be determined that the application documents do not sufficiently disclose the claimed solution.
3. Typical Scenarios of Insufficient Disclosure
(1) Specification disclosure contradicts established technical common sense
Reexamination Decision No. 1853778 held that “According to the definition of the parameters, X+Z in the formula represents superimposing ‘noise data’ onto ‘original image data,’ which clearly contradicts the purpose of image denoising using a denoising algorithm.”
The objective of the invention is image denoising, but the means described in the specification is to superimpose noise data onto the original image data. This technical means clearly contradicts established technical common sense, leaving one skilled in the art unclear as to how to obtain a denoised image through this algorithm. Therefore, the technical means described in the specification are unintelligible.
(2) No natural-law-compliant correlational relationship between input and output data
In the case underlying Reexamination Decision No. 2077737, the technical problem to be solved is that existing cancer screening requires specialized diagnostic equipment and dedicated scenarios, making testing cumbersome and resulting in low screening rates.
The technical means provided in the specification involve auxiliary tumor risk assessment based on the amplitude and frequency of motion of pixels in the subject’s head and neck. However, the specification fails to record any natural-law-compliant correlational relationship between the predicted tumor and the motion amplitude and frequency of the head and neck, nor does the prior art provide evidence of a clear, reasonable, and robust correlation between head and neck motion and tumor occurrence. One skilled in the art would not be able to implement the invention based on the specification.
(3) Ambiguous key parameters or technical terms
In the case underlying Reexamination Decision No. 2390267, the application aims to provide an intelligent textile pattern matching method to address low precision in detecting matching between finished textile patterns and sample patterns.
However, key mathematical formulas and technical terms described in the specification are ambiguous. Specifically, the mathematical formula for Pj contradicts its intended physical meaning, the physical meaning of the formula for Qj is unclear, and the term “similarity” is ambiguous, making it impossible for one skilled in the art to implement the claimed invention.
(4) No disclosure of critical data source
In the case underlying Reexamination Decision No. 1842287, the objective is to construct a reignition prediction neural network model using data from national historical petrochemical storage tank accidents.
However, the Background Art section of the specification says, “Existing methods do not consider the possibility of tank re-ignition due to high smoke temperature, smoke concentration and environmental factors at accident sites.” One skilled in the art would not be able to determine whether smoke temperature, smoke concentration, and environmental factors at accident sites were measured and recorded in petrochemical storage tank historical cases in relation to tank re-ignition.
Inability to access such critical data from national accident records makes it impossible to construct the reignition prediction neural network model, rendering the claimed technical solution not implementable
(5) Mathematically invalid formulas
The application underlying Reexamination Decision No. 2003748 proposes an AI visual identification system for unmanned equipment to overcome the limited protection types and subpar protective efficacy of existing safety protection systems.
However, in the key formula “Y2/Y1−Y1×G=Yg” in the specification, “Y2/Y1” denotes the actual operable distance the unmanned equipment can travel on remaining power, and “Y1×G” denotes power consumption required to travel from the task starting point to the task execution point. As known to one skilled in the art, distance and power consumption are quantities with completely different dimensions, and subtracting one from the other has no physical meaning. Therefore, the solution described in the specification is ambiguous and unintelligible.
(6) Lack of specific technical means for algorithm
In the case underlying Reexamination Decision No. 2104712, a method for oratory training based on energy consumption calculation is put forward, which determines comprehensive energy loss, oratory skill power, and composite emotional indicators during oratory training, and generates customized training recommendations for speakers by analyzing the metrics using a neural network model.
However, the descriptions in the specification do not enable one skilled in the art to know how to calculate the power for each dimension such as speech rate, tone, and emotional state based on the power output of various detection devices using the provided power formulas, how to obtain the comprehensive energy loss, or how to analyze the comprehensive energy loss, oratory skill power, and composite emotional indicators using the neural network model to generate the oratory training recommendations. Therefore, the technical means are vague.
(7) Disconnect between algorithm models and application scenarios
In the case underlying Reexamination Decision No. 2091283 (Case 3 discussed above), the panel is of the following opinions: “To solve technical problems such as control in an industrial robot scenario by constructing an algorithmic model, it is essential to clearly and completely record how the processing steps of the algorithmic model and the output results are combined with the industrial robot scenario. Specifically, the specification must elaborate how processing objects, workflows, and outcomes of Steps S1–S5 integrate with historical interaction data and raw state data of the industrial robot scenario—that is, how Steps S2–S5 process scenario-specific data and generate results relevant to industrial robot operations. However, the specification is vague in describing the technical means for adaptively applying the above model algorithm steps to the industrial robot scenario.”
(8) Algorithms apparently incapable of solving the technical problem
In the case underlying Reexamination Decision No. 2036982, the application seeks to address the lack of a high-precision and targeted dial deformation detection solution capable of identifying minor dial deformations.
However, one skilled in the art would understand from the specification that, in practice, it would be extremely difficult to satisfy the condition for determining “no deformation” stated in the application: “the number of boundary pixels of each image block of the dial in the multi-level transformed image and the number of boundary pixels of each standard contour boundary of the corresponding standard contour image satisfy a divisibility relationship between the maximum and minimum values.” This means it would be almost impossible to determine that the dial is undeformed, rendering the system clearly incapable of accurately identifying whether the meter components are intact.
In a word, although the specification provides technical means purportedly capable of solving the technical problem, one skilled in the art, if making use of those means, would not be able to solve the stated technical problem.
VII. Conclusion
In recent years, the filing and examination of a large number of AI patent applications have driven continuous refinement of China’s examination rules for them. The CNIPA has also published these rules through amendments to the Guidelines. Drawing on the Guidelines and a series of reexamination precedents, this paper summarizes examination standards governing patent eligibility, inventive step, sufficiency of disclosure, and compliance with Article 5.1 of the Patent Law for AI applications. It is hoped that this will enable readers to better grasp the current examination rules and trends for AI patent applications in China for more effective filing and protection. We will keep an eye on the developments of AI patent examination in China.