Artificial Intelligence as a Tool in Patent Claimed invention
By Matthias Kügele, Partner and Swiss Patent Attorney, and Marc Bompart, Ph.D., European and French Patent Attorney – Katzarov, Patent & Trademark Attorneys
For several years now, artificial intelligence has permeated our lives, advancing daily in terms of quality, applications, and ubiquity. It is therefore unsurprising to observe that it has established itself as one of the most patented technologies in the world, or at least one of the major technologies in terms of patent application filings. This distinction, obvious to professionals, is far less so to the layperson, yet it is all the more important given that in the field of patent applications relating to an AI tool, filing a patent application is far from systematically resulting in the granting of a patent.
Indeed, whether claimed as a standalone invention or used as a tool within a technical process, AI raises fundamental questions for intellectual property offices. Between reinterpreted patentability criteria, enhanced description requirements, and geographic disparities in grant rates, patent law is facing a profound transformation.
Examples of Innovation Sectors Involving AI as a Tool
The integration of artificial intelligence (AI) into technical fields as diverse as healthcare, finance, manufacturing, energy, and transportation has generated an increasingly dense and structurally complex intellectual property (IP) landscape.
The presence of AI across virtually all economic sectors is reflected in the scope of protection of patent applications relating to AI tools, as defined by their claims.
These economic sectors include, for example, the medical field and more specifically diagnostic assistance systems, imaging analysis, and therapeutic prediction, which represent a first major axis of innovation involving AI. Hundreds of patent families covering machine learning algorithms applied to oncology or radiology have been filed.
AI applied to medical imaging for the detection of cancerous lesions, organ segmentation, histopathological analysis, etc., represents one of the most active areas in terms of filings. However, it faces a dual exclusion: that of mathematical methods on the one hand, and that of diagnostic methods practiced on the human or animal body under Article 53(c) EPC on the other.
The drafting strategy must therefore ensure that claims are anchored in image processing as a technical signal, rather than in the diagnostic act itself. A claim directed to a medical image processing system that generates a lesion probability map at the output of a convolutional neural network, with specific architectural parameters producing unexpected sensitivity for a given pathological class, will be more robust than a claim targeting the diagnosis of said pathology.
A concrete example is US patent 10,957,041, which claims a convolutional neural network system for the detection of diabetic retinopathy from fundus images, with a clinically validated sensitivity exceeding that of an ophthalmologist in mass screening conditions.
In another domain, transportation and autonomous driving constitute a particularly dense area for patent application filings: claims focus on the interaction between environmental perception, real-time decision-making, and multi-sensor data fusion.
The technical effect, such as safety, reduced reaction times, and management of degraded situations, is here intrinsically embedded in a controlled physical process.
A concrete example is European patent EP3678109, which claims a method for training a recurrent neural network to predict the trajectory of mobile agents (pedestrians, vehicles) in an urban environment, combining LiDAR, camera, and HD mapping data, the claimed technical effect being a reduction in simulated collision rate of more than 40%.
In this context, the legal complexity lies primarily in the density of the existing patent landscape, with massive portfolios held by manufacturers, suppliers, and digital players. Consequently, claiming specific methods for managing uncertainty or detecting out-of-distribution situations may constitute a potentially fertile axis of differentiation.
In the manufacturing industry, AI systems are used for predictive maintenance, process optimization, and quality control, with claims targeting neural network architectures embedded in production systems. These innovations present an a priori favorable patentability profile since their technical effect is directly measurable (reduction in failure rate, improvement in energy efficiency, reduction in rejects).
However, offices, and in particular the EPO, will scrutinize whether the shift from a classical control approach to a machine learning approach constitutes a routine variation within the reach of the skilled person, or whether the specific combination of model architecture, reward function, and real-time constraints produces an unexpected result in terms of inventive step.
A concrete example is European patent EP3557355, one technology of which covers a reinforcement learning system applied to the dynamic regulation of an industrial furnace, where the AI agent continuously adjusts temperature and flow parameters to reduce energy consumption while maintaining part compliance.
The Fintech sector is also seeing a proliferation of patent application filings related to fraud detection, automated credit scoring, and algorithmic portfolio management.
AI applications in this sector sit at the intersection of several exclusions: mathematical methods, business methods, and mental acts.
Patentability is not excluded in principle, but it requires rigorous characterization of the technical effect. For example, a claim directed to the detection of anomalies in a real-time transactional data stream, with a technical effect on signal processing or the management of the system’s computational resources, will be more robust than a claim targeting fraud rate reduction as such.
A concrete example is US patent 10,679,198, which claims an LSTM recurrent neural network trained on transaction sequences to detect fraud patterns in less than 50 milliseconds, incorporating a mechanism to justify each blocking decision.
Finally, human-machine interfaces (NLP, vision, voice) constitute the most active segment by volume of patent application filings, driven by major American, Chinese, and European digital players. For example: Chinese patent CN112307768 claims a multimodal transformer architecture for the simultaneous understanding of text, images, and audio signals in an embedded voice assistant, with claims specifically anchored on the reduction of inference latency on mobile processors, this hardware performance gain conferring the required technical character.
The common thread among these examples of granted patents is that they propose innovations endowed with a technical character, thereby making it possible to solve a technical problem.
Technical Character of AI-Related Inventions
Indeed, the patentability of an AI-based invention faces, in most jurisdictions, the exclusion of abstract methods, computer programs as such, and mental methods, and it is precisely the notion of technical character that constitutes the Gordian knot of AI claims.
At the European Patent Office (EPO), case law and, more recently, examination guidelines require that the invention produce a technical effect going beyond the normal interactions between a program and the computer on which it runs. Concretely, a neural network that improves the classification of medical images by reducing the false positive rate may claim a concrete technical effect related to diagnostic accuracy.
In the United States, the Alice/Mayo test derived from Supreme Court rulings (Alice Corp. v. CLS Bank, 2014) requires that claims cover more than an abstract concept implemented on a generic computer. The USPTO has published successive guidelines to clarify the application of this test to AI, but legal uncertainty remains high, and many applications face initial objections on the basis of this test.
In practice, patent drafters adopt several strategies: anchoring claims in measurable effects on the physical world, specifying the technical architecture of the system (processing layers, training parameters, input data structure).
Training Data, Reproducibility, and Sufficiency of Disclosure
Beyond the technical aspect, one of the most delicate requirements for patents relating to AI systems is that of sufficiency of disclosure, known as enablement in US law. This requirement demands that a person skilled in the art be able to reproduce the invention from the patent text alone, without additional inventive effort.
However, AI models, and in particular large language models or deep neural networks critically depend on their training data, hyperparameters, and optimization conditions. Failing to disclose these elements often renders the invention non-reproducible.
Generally speaking, current European law is designed for deterministic inventions (if I do A, I get B). AI is by nature probabilistic and evolving. The EPO has therefore begun formulating objections on this basis when claims relate to specific model performance without the data or training procedure being described with sufficient precision.
The question of training data also raises cross-cutting intellectual property issues: such data may be protected by copyright, sui generis rights over databases, or trade secrets. Applicants thus find themselves in a dilemma between the requirement for sufficient disclosure and the protection of their competitive advantage.
An emerging approach consists in describing not the data itself, but the selection criteria and statistical characteristics of the training sets, as well as the validation protocols enabling the claimed performance to be achieved.
Filing/Grant Ratio: Comparison of Europe, China, the United States, and Japan
Comparative statistics reveal significant disparities in the way the four major jurisdictions handle AI patents.
In the United States, the USPTO records the largest number of AI filings in absolute value, with a strong concentration among major technology companies. However, due to patentability obstacles under § 101, the grant rate for applications with a strong software component hover around 50 to 60%, according to studies conducted by the USPTO itself, significantly below the office’s overall average.
In Europe, the EPO displays a generally lower grant rate for purely IT-based inventions, but technically well-anchored applications are far better. The EPO published a report on AI patents in 2022 indicating that filings in this domain had increased fivefold between 2013 and 2022, reaching 9,000 annual applications. The abandonment or refusal rate nevertheless remains high, approaching 30 to 35% for applications with purely algorithmic content.
In China, the CNIPA (formerly SIPO) has recorded the highest global volume of AI filings since 2017, driven by national policies incentivizing innovation. However, experts note a significant gap between filings and grants: many Chinese families are the subject of multiple applications on closely related variants, and the average quality of granted patents is considered heterogeneous by comparative WIPO studies.
In Japan, the JPO stands out for its pragmatic approach: the revised guidelines of 2019 relaxed the criteria for AI inventions, leading to an increase in the grant rate, estimated at approximately 70%, among the highest of the major jurisdictions for this domain.
Conclusion
AI, as a tool claimed in patents, is redrawing the boundaries of industrial property. Offices are compelled to adapt their examination practices to a rapidly evolving technology, while applicants must master both the local legal subtleties and the technical imperatives of reproducibility. In this context, a coherent international filing strategy, taking into account the divergences between the EPO, USPTO, CNIPA, and JPO, is becoming a decisive competitive advantage.
About Katzarov. Katzarov’s patent attorneys combine deep technical expertise with extensive EPO, USPTO and international prosecution experience to help innovators secure robust, defensible protection for AI-related inventions. Matthias Kügele, Partner and Swiss Patent Attorney, and Marc Bompart, Ph.D., European and French Patent Attorney, advise clients on patent strategy for artificial intelligence and computer-implemented inventions.
Frequently Asked Questions
What is “AI as a tool” in patent law?
In patent law, AI as a tool refers to artificial intelligence used within a technical process rather than claimed as a standalone invention. Patent applications using AI as a tool span medical imaging, autonomous driving, manufacturing, fintech, and human-machine interfaces.
Can AI inventions be patented at the European Patent Office (EPO)?
Yes, AI inventions can be patented at the EPO when they produce a technical effect going beyond the normal interactions between a program and a computer. The EPO requires demonstration of technical character, for example, a neural network that reduces the false positive rate in medical image classification claims a concrete technical effect related to diagnostic accuracy.
How does the USPTO assess AI patent eligibility?
In the United States, AI patent eligibility is governed by the Alice/Mayo test from Alice Corp. v. CLS Bank (2014). This test requires that claims cover more than an abstract concept implemented on a generic computer. The USPTO has published successive guidelines to clarify the application of the test to AI.
Which jurisdiction grants the highest percentage of AI patent applications?
Japan’s JPO leads with an approximate 70% grant rate following its 2019 revised guidelines that relaxed criteria for AI inventions. The USPTO grants around 50–60% of applications with strong software components. The EPO sees abandonment or refusal in 30–35% of applications with purely algorithmic content. China’s CNIPA records the highest filing volume but with heterogeneous grant quality.
What does “sufficiency of disclosure” mean for AI patents?
Sufficiency of disclosure (enablement in US law) requires that a person skilled in the art can reproduce the invention from the patent text alone, without additional inventive effort. For AI patents this is particularly challenging because models depend on their training data, hyperparameters, and optimization conditions. Failing to disclose these elements often renders the invention non-reproducible.
How can applicants address training data disclosure for AI patents?
An emerging approach is to describe not the data itself, but the selection criteria and statistical characteristics of the training sets, as well as the validation protocols enabling the claimed performance to be achieved. This balances sufficiency of disclosure with protection of training data as copyright, sui generis database rights, or trade secrets.
What are examples of granted AI patents?
Concrete examples include US 10,957,041 (convolutional neural network for diabetic retinopathy detection), EP3678109 (recurrent neural network for trajectory prediction in autonomous driving), EP3557355 (reinforcement learning for industrial furnace regulation), US 10,679,198 (LSTM for transactional fraud detection in under 50 ms), and CN112307768 (multimodal transformer for embedded voice assistants).
Why should AI patent claims in medical imaging be anchored in image processing?
Medical AI patents face a dual exclusion at the EPO: mathematical methods on the one hand, and diagnostic methods practiced on the human or animal body under Article 53(c) EPC on the other. A claim directed to a medical image processing system that generates a lesion probability map with specific neural network architectural parameters is more robust than a claim targeting the diagnostic act itself.
How does AI patentability compare between Europe, China, the United States, and Japan?
The four major jurisdictions handle AI patents very differently. The USPTO has the highest filing volume but moderate grant rate (50–60%) due to § 101 obstacles. The EPO refuses 30–35% of applications with purely algorithmic content but grants well-anchored technical applications. CNIPA in China has the largest global filing volume but heterogeneous grant quality. JPO in Japan has the highest grant rate (around 70%) since its 2019 revised guidelines.
What strategies improve the chances of an AI patent grant?
Patent drafters anchor claims in measurable physical-world effects, specify the technical architecture of the system (processing layers, training parameters, input data structure), and ensure the AI’s role addresses a concrete technical problem rather than an abstract method. A coherent international filing strategy, taking into account the divergences between EPO, USPTO, CNIPA, and JPO, is a decisive competitive advantage.
