AI is magical but NOT magic! Be specific in your AI patents!

What is the standard of sufficient disclosure for AI-related patents: A case study in China

The rapid growth of AI patent applications present significant challenges to existing patent application and examination practices. One of them is satisfying the requirement for sufficient disclosure. The following case was featured in the 2023 CNIPA Compilation of Key Decisions on Patent Reexamination and Invalidation Cases (“2023 Compilation”) as an example to elucidate the current standards for sufficient disclosure in examining AI patent applications in China.

Rejection of Alipay’s AI + Brainwave Patent

The applicant Alipay¹ had a Chinese Patent² on an invention relating to a method and system that trains and utilizes AI models with people’s brainwaves and purchase history data to facilitate online payment operations.

All claims were initially rejected by the CNIPA for lacking inventive step and then again by the Re-examination Panel (“the Panel”) for non-sufficient disclosure, which was also the focus of the oral hearing.³

Alipay argued that the patent had sufficient disclosure by stating that “the historical brainwave data in the sample must match acquisition corresponding to actual payment. This requirement effectively explains the (allegedly missing) characteristics of the sample data in the specification.”⁴ Furthermore, Alipay stated that the machine learning models were existing, generic ones that could be reproduced by those skilled in the art.

However, the Panel did not accept Alipay’s arguments and maintained that technical means provided in the claims are vague to a skilled person in the art, making it impossible to implement them based on the specification.

Firstly, the application doesn’t clearly explain how to train the models. It lacks details on data characteristics, such as whether to categorize by gender or age, or which brainwave parameters—amplitude, frequency, or time range—should be collected. The training method appears to be more of a collection of ideas than a specific approach.

Secondly, it is essential to examine the relationship between the input training data and the model's outputs. However, the current specification lacks experimental data, making it difficult for a skilled individual to determine their correlation.⁵

Therefore, the Panel upheld the rejection decision.

CNIPA’s Standpoint on Sufficient Disclosure of AI Patents

In the 2023 Compilation, CNIPA commented on the rejection in this case, providing a more explicit statement about the current standards of sufficient disclosure of AI patent application examination in China: the specification must clearly define the meaning of the data; specify the employed model; describe the training/optimizing methods (if necessary).⁶

EIP Thoughts

From this case, we can see several characteristics of AI inventions that could make their patent applications uniquely challenging to traditional practices regarding sufficient disclosure.

1. Complexity and Black-Box Nature of AI Models. Many AI models operate as "black boxes," where even the developers may not fully understand how specific results are derived. This makes it difficult to describe how the invention works clearly. However, traditional standards assume that the mechanisms underlying an invention can be explicitly defined, which may not be feasible with some AI systems.
2. Data Dependency. AI systems rely on training data, and the quality and characteristics of this data greatly influence the resulting models. We agree with the Panel that without a clear description of these characteristics, a skilled person cannot effectively implement the solution needed to train the models of the patented AI system.
3. Evolving Nature of AI Models. AI models can evolve over time as they are retrained with new data, making the invention less static than traditional technologies. This raises questions about whether the disclosed AI system is a fixed invention or an evolving process, complicating the sufficiency of disclosure.
4. Interdisciplinary Complexity. AI inventions often combine concepts from multiple fields, such as computer science, mathematics, neuroscience, and finance in the case discussed. Examiners may struggle to determine whether the invention is sufficiently disclosed for skilled persons across these fields.
5. Lack of Experimental Data. Many novel AI models lack supporting experimental results or benchmarks for their claimed effects. In the case discussed, the connection between brainwaves and payment decisions is still theoretical, making it challenging to prove the invention's operability and reproducibility.
6. Generalized vs. Specific AI Applications: AI models are often described in broad terms. However, sufficient disclosure requires detailed descriptions of training solutions that train generic models with application-specific data. Simply stating that any generic model will work is insufficient.

To better reach sufficient disclosure standards for AI patents in China, we can learn several key lessons from this provided case:

1. Clarity in Data Characteristics. Simply stating general data acquisition or relationships is insufficient; explicit and detailed descriptions are essential. The specification must clearly define the data involved, including:
   • Types, i.e., the nature of the input and output data (e.g., numerical, categorical, or signal-based);
   • Attributes, i.e., the fundamental properties such as frequency or amplitude.
   • Contents: specific data representations, e.g. gender, age, etc.
   • Data collection or acquisition procedures.
2. Specification of AI models. Naming an AI model without providing details about its implementation can lead to rejection. The specific AI model used must be identified, including:
   • If the model is known in the field, its usage must still be sufficiently described in the application context.
   • If a new model is proposed, its architecture, functionality, and distinguishing features must be elaborated.
3. Training and optimization processes. Model training and optimization descriptions should not appear as mere concepts or proposals but as actionable steps that can be replicated. For AI inventions, the specification should explain:
   • How the model is trained;
   • How the data is used in training (e.g., preprocessing, feature extraction);
   • How the optimization of the model is conducted to achieve the intended results.
4. Relationship between input, model, and output. The more difficult it is to understand the black-box nature of AI inventions, the more test details should be provided. The specification should include experimental results or examples to validate the claimed relevancy between input data characteristics, data processing, and the generated output.
5. Technical Implementability. A person skilled in the art must be able to implement the claimed solution based solely on the information disclosed in the specification. Ambiguity in AI methods, parameters, or data characteristics undermines this standard.
6. Supporting Experimental Data. A lack of experimental data or concrete evidence of the AI model's effectiveness may weaken the application’s credibility. The specification should ideally include experimental results, simulations, or practical examples to support the feasibility and functionality of the invention.

AI seems "magical" compared to traditional computer technologies, although it is not unpatentable magic. As AI systems evolve unpredictably, their consistency diminishes, resembling complex natural systems classified as an "unpredictable art" by patent laws. For these arts, experimental data is vital for fulfilling enablement and support requirements. Consequently, as AI becomes more unpredictable, the criteria for patenting AI inventions may mimic more and more the stricter data standards seen in life sciences.

This article is for general informational purposes only and should not be considered legal advice or a legal opinion on a specific set of facts.