Author: Dawn YU

In the rapidly evolving landscape of Artificial Intelligence, defining the demarcation between "mathematical algorithms" and "patentable technical solutions" has become a cornerstone of global IP strategy. By comparing recent invalidation decisions from the China National Intellectual Property Administration, specifically Cases No. 586065, 584244, and 588362, with the U.S. Federal Circuit’s ruling in Recentive Analytics v. Fox Corp., we can discern critical jurisdictional nuances in how the validity of AI patents is adjudicated.

I. Subject Matter Eligibility: The "Specific Technical Association" Requirement

Under the 2023 revision of the Chinese Patent Examination Guidelines, the eligibility of algorithm-related applications hinges on whether the algorithmic improvement maintains a "specific technical association" with the internal performance of the computer system.

1. Algorithmic Optimization of Internal System Performance. In a notable reexamination case involving neural network optimization, the applicant proposed a method to transform multiple individual matrix operations into batch matrix operations to utilize better underlying parallel computing libraries (e.g., CUDA, MKL). The Reexamination Panel held that this addressed the technical problem of excessive system call overhead and enhanced hardware execution efficiency, thereby constituting a "technical solution" under Article 2.2 of the Patent Law.

2. Mapping Algorithmic Outputs to Physical Control Parameters (Decision No. 586065) In a case concerning chip placement parameters, the claims utilized an AI model to calculate a "tray deviation degree", which directly dictated the physical movements of a rotation mechanism. The Panel found that by converting abstract recognition results into specific industrial control parameters, the solution achieved a technical character, thereby securing its eligibility.

II. Inventive Step: Model Fine-tuning vs. Generic Functional Implementation

Once the "eligibility" threshold is crossed, the survival of an AI patent depends on whether its logic provides a non-obvious contribution over the prior art.

1. The Independent Value of Model Fine-tuning (Decision No. 584244 – Upheld). This case involved a method for generating dynamic images from audio. The Panel noted that while generating images from audio was known, this patent used a "trained generative network" and "reference images" to create a "target generative network". This targeted fine-tuning solved the engineering problem of reducing the cost of generating high-precision, subject-specific images, and was thus found to possess an inventive step.

2. The Risk of "Black Box" Algorithmic Modules (Decision No. 588362 – Invalidated). In contrast, a patent for a "Sleep Induction Device" was declared invalid.

• The Invalidation Logic: The Panel      determined that, given AI's known ability to process physiological      signals, merely introducing a neural network for "data mining"      and "parameter correction" within a traditional monitoring      framework constitutes a routine technical choice.
  
• Practice Tip: If an AI      algorithm is presented merely as a generic tool to improve accuracy      without disclosing specific architectural improvements or targeted      adaptation strategies, the inventive step is highly vulnerable to being      characterized as a "simple aggregation" of common knowledge.
  

III. U.S. Perspective: the “Abstract Idea” Doctrine Under § 101

Unlike China’s engineering-centric focus on solving a specific technical problem, U.S. patent jurisprudence centers on whether the claimed algorithm is an “abstract idea” and, if so, whether the claims include an “inventive concept.”

1.     Field-of-use limits are not enough to clear the eligibility bar.
In Recentive Analytics, the Federal Circuit held that merely applying established machine-learning methods to a new data environment (e.g., scheduling optimization), without improving the machine-learning model itself, is not patent-eligible. The court reasoned that iterative training and dynamic adjustment are inherent characteristics of machine learning and, therefore, cannot be treated as a meaningful inventive feature.

2.     How U.S. courts define “functional improvement.”
U.S. courts often view improvements, such as higher prediction accuracy or greater efficiency, as still abstract if they do not amount to an improvement in computer functionality itself. This diverges sharply from CNIPA’s reasoning in Decision No. 584244, where “reducing generation cost through targeted fine-tuning” was credited as a contribution supporting inventiveness.

IV. Strategic Recommendations for AI Patent Portfolios

Based on these cross-border precedents, practitioners should adopt the following strategies:

1)Avoid "Generic AI" Descriptions: As demonstrated in Decision No. 588362, practitioners must avoid describing AI modules as "black boxes" for generic analysis. It is essential to detail how the algorithmic logic is optimized for a specific scenario (e.g., sleep staging or image fine-tuning).

2)Strengthen the integrated description of algorithms and underlying technical components.
In drafting, explain in detail how the algorithm is adapted to specific underlying software libraries or hardware architectures, and clearly articulate how that adaptation improves internal system performance (e.g., memory overhead, computational efficiency).

3)Highlight inventiveness in the training/fine-tuning stage.
As shown in Decision No. 584244, positioning the claims at the “model fine-tuning” stage and disclosing stage-specific labeling, reference data, or iteration logic can be a decisive entry point for proving inventiveness.