Richard Yong WANG  Partner, Attorney-at-law and Patent Attorney

                                

Since the Alice case in 2014, software and artificial intelligence (AI) patents in the United States have endured a decade-long period of marginalization. However, entering 2025, marked by the appointment of the new Director of the United States Patent and Trademark Office (USPTO), John Squires, the U.S. has demonstrated a radical transformation in legislative initiatives, examination standards, and administrative governance. This paper aims to analyze the shift in the U.S. approach to AI patent examination from "preventing abuse" to "avoiding wrongful rejection., explores the strategic intent behind defending national competitiveness by relaxing the thresholds for patent-eligible subject matter (Section 101), and furthermore, based on the technical nature of AI as an independent scientific field, re-examines the current provisions in China's patent examination system regarding the eligibility of algorithm and model improvements and proposes recommendations for updates.

 

I. Background of AI Patent Examination in the United States

Over the past decade, the U.S. patent system has long been in a state of wavering and cautious response when facing the development of AI technology. The 2014 Alice v. CLS Bank case established a two-step test for examination, originally intended to curb overly broad business method patents. However, in practice, it gradually evolved into a "stringent threshold" for AI patents. A large number of patent applications involving machine learning and neural networks were rejected at the Section 101 stage for patent-eligible subject matter after being labeled as "abstract ideas," without even proceeding to the substantive examination stages of novelty (Section 102) or non-obviousness (Section 103).

 

For a long time, USPTO examiners have also adopted an oversimplified logic in their examination practices: as long as an invention involves model training or parameter optimization, it is equated with unpatentable mathematical methods. This perception severely overlooks the substantial innovation of AI at the level of engineering implementation, such as in system architecture, resource scheduling, and high-dimensional parallel computing. This systemic bias of “algorithms as abstractions” has made it increasingly difficult for cutting-edge AI technologies to obtain protection, significantly dampening the industry’s enthusiasm for investing in foundational AI infrastructure and algorithms.

 

II. 2025: A Pivotal Year for the Strategic Shift in U.S. AI Patent Policy

 

Beginning in the second half of 2025, the USPTO has been breaking this long-standing pattern of examination through a series of administrative and judicial actions.

 

A landmark event is the Desjardins case. This case addresses a highly typical technical challenge in the field of machine learning—"Catastrophic Forgetting," or solving AI's "memory loss" problem. In existing technologies, when an artificial intelligence model (such as a large model) learns a new task, it often loses the ability to perform previous tasks. The invention in question improves the functionality of a machine learning model through specific parameter value adjustments, enabling the model to retain old knowledge while acquiring new knowledge. This invention is directly relevant to multi-task models, continual learning, and the industrial practicality of current large models, representing a substantive improvement in artificial intelligence technology.

 

However, the examination process of this case has been exceptionally challenging, fully reflecting the stringent restrictions placed on software patents in the United States over the past decade.

 

During the examination phase, the examiner considered that its claims involved parameter adjustments and mathematical calculations. Employing the traditional logic of "algorithms as abstractions," the examiner categorized them as a "mathematical algorithm," deemed an unpatentable "abstract idea," and concluded that they failed to meet the requirements for patent-eligible subject matter under Section 101 of the U.S. Patent Act.

 

When this case was appealed to the Patent Trial and Appeal Board (PTAB), the PTAB did not limit itself to reviewing only the grounds raised by the examiner. Instead, it proactively introduced additional grounds for rejection, reinforcing the determination of a "mental process." The PTAB held that because the steps involved mathematical logic, they could theoretically be performed by the human mind, thereby further blocking any path to patent grant. This reflected the highly vigilant and even predisposed-to-rejection attitude of the examination system at that time toward AI patents with mathematical characteristics.

 

The turning point in the fate of this case came with the appointment of the new USPTO Director in September 2025. Shortly after assuming office, the new USPTO Director, John Squires, took an extraordinary step by directly intervening in the case and overturning the previous rejection.

 

In the final ruling, John Squires explicitly stated that the invention does not merely perform mathematical operations but improves the overall functioning of the machine learning system through parameter control. He concluded that this constitutes a substantive improvement to computer technology, rather than an abstract theory.

 

This reversal essentially performs a "de-abstraction" of AI technology, marking a formal shift in U.S. patent examination from a purely mathematical-logic perspective back to an engineering-improvement perspective. It re-establishes the standard for evaluating AI patent claims from an engineering standpoint rather than a purely mathematical one.

 

On August 4, 2025, the internal memorandum issued by the USPTO to examiners also conveyed some revolutionary signals:

 

1）The "Over 50% Certainty" Principle: Examiners are not allowed to issue rejections under Section 101 unless they have more than 50% certainty that the claims constitute an abstract idea. This effectively downgrades Section 101 from a "substantive filter" to a "weak examination condition."

 

2）Redefining the Boundaries of "Mental Processes": It explicitly states that actions such as "real-time clustering of embedding vectors," "running convolutional neural networks on multi-frame videos," and "training neural content" cannot be simply classified as "mental processes" that the human brain can perform.

 

Meanwhile, to address the average backlog of 22.5 months in examination, the USPTO has launched the AI-assisted search program and the streamlined claim review program: leveraging AI tools to generate search reports before the first Office action to improve application efficiency. Additionally, starting November 20, 2025, the authority to institute IPR proceedings will be centralized under the Director, aiming to reduce repetitive and harassing challenges against patents and enhance patent stability through a "One-and-Done" policy.

 

III. Signals and Implications of This Policy Shift

 

The core rationale behind the USPTO's reform is that the standards for patent eligibility have escalated to a matter of national security.

 

Director Squires explicitly stated that tightening AI patent examination standards leads to job losses and industry relocation, thereby strengthening competitors, while relaxing standards creates jobs and safeguards national security. This perspective directly links the scope of AI patent protection to global technological leadership. Director Squires also invoked the legal tradition of the 1854 Morse Telegraph Case to defend his policy, arguing that patent law must align with technological trends. Protecting the once-cutting-edge and seemingly abstract telegraph technology established America's industrial standing, and the same holds true for AI today.

 

This series of shifts will have profound implications for AI-related industries. As AI, software, and related business method patents are no longer considered "high-risk, low-return" areas, the difficulty of obtaining patent rights has significantly decreased. Companies will reassess their patent strategies in the United States, leading to a reevaluation of the strategic value of AI patents. With the focus of examination shifting, Section 101 is no longer a barrier to innovation, and the emphasis has clearly moved to the examination of novelty under Section 102 and non-obviousness under Section 103. Consequently, businesses will increasingly view patent applications as a critical means of accumulating core intangible assets, significantly boosting their enthusiasm for filing AI-related patents. At the same time, the improved expectation of patent stability will enhance investor confidence in AI startups. By reassessing and strengthening their patent portfolios in the U.S., companies may more effectively attract capital investment, thereby advancing the commercialization of cutting-edge technologies.

 

IV. Implications and Reflections on China's AI Patent Examination System

 

In light of the institutional reconstruction by the United States in the field of AI patents, which adopts both offensive and defensive measures, China's current patent examination system, particularly regarding the review of AI, big data, and algorithms, urgently requires re-evaluation.

 

The author believes that after decades of sustained development and continuous evolution, artificial intelligence research has matured into an independent, systematic scientific discipline. It is now understood that algorithmic improvements in AI are not merely arbitrary accumulations of mathematical logic. Rather, they are designed to address scientific challenges such as distributed system stability, large-scale parameter convergence, and multimodal data alignment. Solving problems within this scientific domain itself constitutes a contribution to humanity's technological repository. To simplistically and mechanically exclude such innovations from patentable subject matter would be tantamount to tying our own hands in this most cutting-edge field.

 

1. Re-examining the Regulations on the Eligibility Review of Algorithms and Models

In China's current examination practice, when reviewing inventions involving computer programs, the core remains assessing whether the "algorithm forms an integral part with the technical problem, technical means, and technical effect." If the improvement of an algorithm is merely reflected in the optimization of logical processing and does not address a specific, recognized technical problem, (for instance, if it is not closely integrated with a concrete application field such as industrial control, medical imaging, etc.), it is highly likely to be considered a mere model or algorithmic improvement and thus not regarded as a technical solution.

 

And the United States, as demonstrated by the Desjardins case, is gradually downplaying the requirement to link specific application areas, tending to believe that as long as the performance of the computational model itself is improved (such as reducing catastrophic forgetting), it is itself solving the "technical problems" of computers. Therefore, it can be considered that the United States is gradually accepting that "improvements to machine learning models themselves" can constitute technological improvements.

 

The current competition in AI technology among nations is, at its core, a competition of institutional systems. It must be recognized that AI technology has evolved from "application-layer innovation" to the stage of "fundamental scientific innovation." China's patent system should respond by appropriately expanding the scope of eligible subject matter, recognizing solutions to AI scientific problems as technical contributions, and incorporating them into the patent protection framework. This would incentivize Chinese AI developers, researchers, and enterprises to engage in long-term, fundamental research and investment in underlying algorithmic architectures. If our examination rules continue to adhere to the outdated notion that "algorithms must have specific industrial applications," while the United States has already begun protecting improvements to the models themselves, future global AI standards and foundational patents will inevitably lean in favor of the United States.

 

2. A Reconsideration of the Definition of "Technical Field" Should Be Undertaken

In China's long-standing patent examination practices, there has been a tendency to confine the "technical field" to traditional industrial domains such as physics, chemistry, and machinery. For "pure model problems" in the AI field (such as optimizing computational costs for long-text attention mechanisms), if they are not associated with specific applications like image processing or speech recognition, they are often challenging to be recognized as technical problems

 

As artificial intelligence evolves from an enabling technology into an independent scientific discipline, its evaluation criteria should shift in a timely manner from "dependence" to "autonomy." That is, it should be acknowledged that endogenous challenges within the AI scientific field (such as vanishing gradients, memory overflow, or catastrophic forgetting in AI models) are themselves technical problems. This is because enabling technologies emphasize "tool attributes" and "external contributions," whereas an independent scientific discipline emphasizes "disciplinary essence" and "internal progress." Viewing AI as an independent scientific field signifies that the law begins to recognize the value of solving endogenous problems in digital space, no longer mandating that they provide "proof of dependence" on the physical world. Improving these intrinsic performance challenges of the models is equivalent to reducing wear or improving thermal efficiency in traditional machinery. If the patent system continues to adhere to the outdated standard of "must be tied to traditional industrial fields," it may artificially downgrade innovation in digital space due to cognitive lag. In fact, the threat posed by "memory overflow" to server clusters is no less significant than that of "pressure overload" to a boiler; the constraint of "vanishing gradients" on algorithmic accuracy is no less impactful than that of "material aging" on mechanical longevity. In the digital age, optimizing algorithm performance equates to deeply harnessing the potential of computing machinery.

 

Therefore, to protect China's fundamental AI innovations and advance China’s digital strength, consideration should be given to appropriately expanding the definition of the technical field by establishing AI as an independent scientific and technological discipline on par with fields such as physics and chemistry. Improvements in areas like data structures, model architectures, and training schemes within this field should enjoy patent examination thresholds with equal status to those of other disciplines. Here, the latest U.S. "engineering perspective" can serve as a reference: as long as an algorithm, through parameter control or architectural optimization, improves the overall operational manner or processing efficiency of a computer system, the algorithm itself constitutes a technical means. Metrics such as increased convergence speed, enhanced generalization capability, reduced computational complexity, and saved storage space should be directly recognized as "beneficial technical effects" within the meaning of patent law. This means that even without being linked to a specific application domain, making a model stronger, faster, and more stable is in itself a contribution to technology.

 

 

3. The Standards for Subject Matter Eligibility Review Should Be Relaxed

In the past, the initial screening conducted through Article 2.2, of the Patent Law (eligibility of subject matter) in China’s patent examination practice has effectively served an overly stringent “gatekeeping” function. While this approach proved effective in addressing early and simple business methods, it now faces challenges when dealing with complex AI technological innovations. Due to the often subjective and ambiguous nature of its judgment criteria, this practice is highly prone to being overly dismissive or rejecting eligible inventions.

 

Therefore, it is recommended to shift the focus of examination backward, appropriately weaken the preliminary screening function of the subject matter eligibility standard, and position it as only filtering out "purely mathematical games" or "purely commercial rules" at the primary threshold. For AI inventions involving complex underlying logic or optimization of model architecture, they should be comprehensively allowed to proceed to the substantive examination stage. Through more objective and rigorous examination of novelty (Article 22.2) and inventiveness (Article 22.3), it should be determined whether an invention truly proposes non-obvious improvements, thereby achieving precise filtration of "low-quality patents" and "frivolous patents."

 

At the same time, to address the pain point of uncertainty in subject matter eligibility standards, it is recommended to introduce more operational practical guidelines. Drawing on the approaches demonstrated by the USPTO in its July 2024 AI examples and the August 2025 Director’s Memorandum, the Examination Guidelines could explicitly exclude advanced computational processes often misinterpreted as “mental activities” through a checklist format. For instance, dimensionality reduction of high-dimensional embeddings, real-time clustering analysis, and multi-layer convolutional feature extraction—although expressed mathematically—are in essence complex computational engineering tasks beyond the capability of the human mind. By incorporating such examples, both examiners and applicants can be provided with clearer expectations, thereby avoiding endless legal debates over “whether something constitutes a technical means.”

 

Furthermore, consideration could be given to establishing a "presumption of eligibility in favor of the applicant." That is, if an invention demonstrates a preliminary indication of technical character, examiners should presume by default that it meets the eligibility requirements, and instead focus their efforts on searching for prior art documents and assessing the inventive step of the technology. This delegation and deferral of the examination standard would significantly shorten the review cycle and improve the quality of granted patents.

 

V. Conclusion

In the global AI competition, the "flexibility" of the patent system serves as critical infrastructure for attracting high-quality innovation resources—talent, capital, and computing power—from around the world. U.S. patent policy is currently at a decisive turning point, transitioning from the restrictive Alice era toward actively supporting AI innovation. Through a series of case rulings and revisions to examination guidelines, the U.S. has substantially lowered the threshold for protection, aiming to consolidate its global leadership in AI.

 

For China, artificial intelligence is not only the core driving force for industrial upgrading but also the cornerstone for securing the nation’s future trajectory. We must recognize that protecting foundational scientific contributions and improvements in algorithmic architecture within the AI field is a necessary prerequisite for fostering the long-term development of this critical technology. Gradually expanding the scope of patent eligibility for AI, and shifting the focus of examination from “stringent subject matter eligibility review” to “substantive technical examination,” represents a strategic choice for China to protect domestic innovation, attract global capital, and cultivate New-generation Productive Forces in the context of future international technological competition

 

Author:

 Mr. Richard Yong Wang

 

Mr. Wang received his bachelor's degree in 1991 from the department of computer science of East China Normal University and his master's degree from the Institute of Computing Technology of the Chinese Academy of Sciences in 1994. In 2005, he received degree of master of laws from Renmin University of China. Mr. Wang joined Panawell in January 2007.

 

In the past years, Mr. Wang has handled thousands of patent applications for both domestic and foreign clients, and he has extensive experiences in application drafting, responding to office actions, patent reexamination and invalidation proceeding, patent administrative litigation, infringement litigation, software registration and integrated circuit layout design registration. As a very experienced patent attorney and attorney-at-law, Mr. Wang also participated in many patent litigation cases on behalf of a number of multinational companies as leading attorney. Mr. Wang's practices include computer hardware, computer software, communication technology, semiconductor devices and manufacturing process, automatic control, household electrical appliances, and etc.