Introduction
The rapid advancement of artificial intelligence (AI) technologies continues to create significant challenges for legal systems worldwide, especially regarding ownership of rights in AI-generated contents, transparency of algorithms, as well as eligibility and examination standards for AI-related inventions. As Taiwan is a key participant in the global technology supply chain, the Taiwan Intellectual Property Office (TIPO) has published a collection titled “AI-Related Invention Cases” in 2025, aiming to provide clear examination directions to the public and seeking greater consistency with the examination practices of the five IP Offices (IP5).

Based on the Patent Examination Guidelines for Computer Software-Related Inventions and the recently released AI-Related Invention Cases, the following sections provide a brief introduction to Taiwan’s examination approach for AI-related inventions and offer practical insights into patenting strategies and utilization of AI technologies. 

Issues and Regulations
The most common issues encountered in AI-related inventions generally fall into the following three categories:
  (1) whether the invention meets the definition of an “invention” under Article 21 of the Patent Act (eligibility);
  (2) whether the specification meets the enablement requirement, particularly whether the specification adequately discloses the logical relationship between the inputs, outputs, and the algorithm; and
  (3) whether the invention involves an inventive step, namely whether it merely incorporates AI techniques or demonstrates a substantive technological innovation.

Regarding eligibility, the determination of whether an AI-related invention meets the definition of an invention fully follows the evaluation framework set forth in the Patent Examination Guidelines for Computer Software-Related Inventions (see the below chart).
Based on the framework, if the inventions cannot be determined to be evidently meet or fail to meet the definition of an invention in the first step, the second step provides further guidelines by analyzing whether the information processing via computer software is concretely realized through the utilization of hardware resources. That is, for cases that are unable to determine whether they meet the definition of an invention in the first step, the collaboration between software and hardware should be clearly recited.

Regarding the enablement requirement, the specification should sufficiently disclose the algorithm to an extent that enables a person of ordinary skill in the art to clearly understand how to design the program. In other words, it should at least disclose the necessary steps or flow of the algorithm such that a person of ordinary skill in the art can carry out the algorithm and achieve the claimed function.

Regarding the inventive step, attention should be paid to whether there are factors that may negate the inventive step, including motivations for combining multiple prior art references, such as commonality of the technical fields, problems to be solved, functions or effects, or presence of teachings or suggestions for combination. In addition, lack of inventive step may also arise from mere simple variations, such as conversion between technical fields, systematize the methods of operation performed by humans, software achieving the functions of prior hardware technology, recreating a common knowledge at the time of filing in a computer-generated virtual space, application or modification of common general knowledge at the time of filing, or features that pose no contributions to technical effect.

Cases and Analysis
In the AI-Related Invention Cases released last year, an exemplified claim set that relates to a method for calculating mattress matching scores is given (discussed in details below) to demonstrate the examination basis in view of the eligibility, enablement requirement and inventive step. 
 
Claim 1 merely discloses a computational model for the calculation without elucidating the mechanism or details of the model, such that the model would likely be regarded as a general mathematical model. Even if the model were deemed as a software model, the entire claim recitation remains nothing more than a method of manmade arrangement and does not utilize the laws of nature. Therefore, it does not meet the definition of an invention.
 
Claim 2 specifically discloses inputting the user’s sleeping position preferences and body characteristic information into an AI matching module to calculate and display matching scores between the user and the mattresses, which satisfies the requirement that information processing implemented by computer software is concretely realized through the utilization of hardware resources. The communication between modules and the collaboration between software and hardware are indeed recited. Therefore, the claimed subject matter meets the definition of an invention.

However, the specification fails to clearly disclose the user’s body characteristic information, making it difficult for a person skilled in the art to understand the disclosure and carry out the invention. Furthermore, during the training phase, only ‘human’ features are inputted, while ‘mattress’ features are omitted; thus, such that the model cannot clearly identify the matching correlation between them. That is, the specification does not elucidate how the input/output data is constructed or whether a reasonable correlation exists, based on the training methods of the selected AI model. Consequently, the specification does not meet the enablement requirement.
 
Regarding claim 3, the specification clearly discloses that the information about the user’s body characteristics includes height, weight, shoulder width, and hip width, and thus this information significantly correlates with determining the mattress types suitable for a user. Through a neural network model, this information can be analyzed to estimate the matching scores between the user and various types of mattresses. The training process involves collecting the sleeping position preferences and body characteristic information of multiple testers, and then combining this data with the intrinsic attributes of the mattresses to calculate the matching scores for each mattress. Accordingly, the specification specifically elucidates the correlation between input data and the trained model’s output and explicitly discloses the model training methodology. Therefore, the specification meets the enablement requirement.

However, Citation 1, as cited in the AI case collection, applies neural networks to calculate matching scores between users and mattresses, incorporating user body characteristic information during the training process. The primary difference is that Citation 1 does not disclose the sleeping position preference information defined in claim 3, meaning the datasets used to train the models are not entirely identical. On the other hand, Citation 2 describes various mattress product designs that categorize users based on their sleeping position preferences (e.g., back sleeping, side sleeping). These patterns determine the materials selected or the support structure configurations for different zones of the mattress. With various combinations of different levels of firmness and support, the mattress’s adaptability and comfort for various sleeping positions are enhanced. 

Claim 3 merely defines the inputting of sleeping position preference information into a neural network model for training, without further defining any particularity in the model’s algorithm or training process (e.g., special pre-processing of input data that distinguishes it from the prior art to improve overall performance). Therefore, compared to Citation 1, although the datasets used to train the models are not entirely identical, there is no distinct difference between the claimed model’s training process and the prior art, and a person skilled in the art would have a motivation to apply the user sleeping position preference information disclosed in Citation 2 as training data for the neural network model in Citation 1. Consequently, the invention of Claim 3 can be easily accomplished and lacks an inventive step.

Summary
Based on the exemplified claim set, major issues of AI-related inventions may be summarized as follows:
 

• For eligibility, aside from cases that concretely performing control of an apparatus, or processing with respect to the control, or concretely performing information processing based on the technical properties of an object, clearly reciting the collaboration between software and hardware is essential to meet the definition of an invention. 

 

• The logical correlation and predictability between input and output data must be ensured; otherwise, the invention may be deemed to fail the enablement requirement due to insufficient technical disclosure. That is, the specification should fully disclose the algorithms, training methods, and data construction, elucidating how they effectively solve the problem to overcome examination challenges arising from the ‘black box’ nature of AI.

 

• If an AI-related invention possesses a unique dataset configuration or training method that yields corresponding effects, it should not be summarily dismissed as easily accomplished; instead, a comprehensive assessment of the technical motivation and the integration of the overall model is required. However, if an invention merely inputs known data into an AI or fuzzy logic system without disclosing specific model or training innovations, thus achieving only general improvements in accuracy or operational convenience, it lacks an inventive step. Therefore, clearly defining the particularity of the model’s algorithm or training process (i.e., how it distinguishes itself from the prior art) increases the likelihood of being recognized as having inventiveness.

It is evident that the examination standards for AI-related inventions in Taiwan are increasingly clear and well-defined, and we anticipate the release of more practical cases to provide further empirical guidance. This evolving clarity is of paramount importance for applicants, as it provides a more predictable framework for strategic application and global IP layout of AI-related innovations.

References 
[1] Patent Examination Guidelines, Chapter 12 (Computer Software-Related Inventions).
[2] AI-Related Invention Cases published by TIPO in 2025.