Artificial Intelligence & Patents: what’s the connection?

Artificial Intelligence is everywhere these days and it is no different in the world of patents. On March 22, 2023 the Administrative Council of the European Patent Organization approved a common practice in regards to examining Computer Implemented Inventions (CIIs) and Artificial Intelligence (AI) between the EPO and most of its contracting and extension states.

A working group started in January 2022 with the aim to establish a common understanding of terms and to outline a common practice among EPC contracting and extension states and the EPO. Several aspects are playing a role in the very fast expansion of AI that we see today. For example, processing power has increased significantly and better AI models are available as a tool. Combine that with big data, cloud computing, 5G, etc. and AI can solve technical problems in almost any technical domain. This resulted in the past years in a significant rise of the number of AI-related patent applications.

The figure below reflects the publications in IPC class G06N for applications which have a publication in EP:

The increase started in 2015 and, although we see a small decline in 2020, it’s expected that the growth will further increase in the coming years.

The common practice first clarifies that the EPC contracting and extension states and the EPO have a common understanding of what CIIs, AI and computer programs are:

CIIs, AI and computer programs

These terms have the meaning given in the art by the skilled person as follows:

• A CII is an invention involving at least one feature that is implemented by a computer program.
• AI is intelligence demonstrated by a machine, in particular producing behaviors perceived as intelligent by humans. AI includes, for example, machine learning and neural networks whose behavior is largely determined by learning from data, and
• A computer program is a set of instructions executed by programmable hardware.

The common practice subsequently clarifies that subject-matter lacking technical character is excluded from patentability. Therefore, the AI must have features which make a technical contribution. It’s commonly known that non-technical features are excluded from patentability when taken in isolation, but these non-technical features can still contribute to technical character. This is the situation for the so-called “mixed-type inventions”. Mixed-type inventions comprise both technical and non-technical features, e.g. mathematical steps related to AI. The common practice clarifies that non-technical features can contribute to an invention’s technical character when they interact with technical features to provide a technical solution. The common practice also clarifies that the knowledge of the skilled person, which is an important factor when deciding on inventive step, includes commonly known AI tools.

Another aspect touched in the common practice is the requirement of sufficiency of disclosure. It’s known that in new research and technology areas (such as areas involving AI), patent applications are sometimes highly speculative resulting in a claimed invention which is insufficiently disclosed. The common practice clarifies that when a technical effect depends on mathematical methods and training datasets, the level of detail in the application must be sufficient to reproduce this technical effect.

Where for example training datasets are used in machine learning algorithms and contribute to bringing a technical effect, the characteristics of the training dataset required for reproducing this technical effect need to be disclosed. The specific training dataset itself employed by the inventors should generally not be disclosed. Taking this common practice into account, we cannot emphasize enough that applicants should provide sufficient details in the patent application during drafting to avoid that the patent application is later on refused due to not fulfilling the requirement of sufficiency of disclosure.

We provide two examples of patented inventions related to AI, one in the field of dental care, the other one in the field of road safety.

Example of patented inventions: Dental Care

The first example in the field of dental care is patent EP3618752. It describes a method for designing dental drilling templates using 3D scans and an artificial neural network.

Caption: Simplified representation of the computer implemented AI method from EP3618752, taking into account figure 2 of EP3618752.

The patent’s prosecution history highlights the potential to patent AI inventions in Europe and clarifies the argumentation used by the European Patent Office (EPO). In particular, the patent faced opposition due to concerns about its inventiveness. The basis of this opposition was that the prior art already describes the use of AI, including artificial neural networks, on surface and volumetric models in other dental applications. The opposing party mentioned restorative dentistry procedures or crown creation activities in particular. Hence, the opposing party deemed the application of similar networks for the design of drilling templates based on a 3D surface and a volumetric model obvious.

However, the European Patent Office (EPO) disagreed. It considered the combination of surface and volumetric models for designing drilling templates to be inventive: whereas the use of AI for dental restorations and crone creation is known, the drilling of holes and the design of a drilling template is another, specific application of AI.

So it is a practical demonstration that the application of AI with a technical effect in a new domain is one of the ways in which an AI invention is patentable.

Thus, it is a practical demonstration that the application of AI with a technical effect in a new domain is one of the ways in which an AI invention is patentable.

Some might suggest that this argument was made in view of general common knowledge as considered in 2018 and that it may not hold up today. In particular, the current practices mentioned earlier in the article seem to hint to the fact that common AI tools are considered part of the knowledge of the skilled person. If you were to interpret this very narrowly, the application of AI to the manual process of designing drilling templates may today be considered obvious and not inventive.

We see two reasons that this would not be the case, particularly in relation to “mere application” and “mere automation”. Firstly, applying AI to a novel (sub)domain often goes beyond the mere application of common AI tools; it requires effort to train the algorithms or to make them work with the appropriate input or produce the right output. Secondly, the application of AI will probably have additional technical effects beyond a mere automation of the design process; these will have to be assessed according to the common practices to assess inventiveness of computer implemented inventions (see above).

Remember, your patent description needs to provide sufficient information to use this kind of arguments during examination, opposition or litigation. Therefore, it is important to work with a patent attorney who understands computer implemented inventions and can help you with this!

Example of patented inventions: Road safety

The second example in the field of road safety is patent application EP3791376A4 with title “Method and system for vehicle-to-pedestrian collision avoidance” for which the grant is intended. More specifically, the present invention relates to a method and a system for collision avoidance between vehicles and pedestrians.

In this patent application there is provided a method for vehicle-to-pedestrian collision avoidance, which method comprises physically linking at least one vehicle to at least one Long-Term Evolution (LTE)-capable user equipment (UE) terminal; physically linking at least one pedestrian to at least one Long-Term Evolution (LTE)-capable user equipment (UE) terminal and determining a spatiotemporal positioning of each terminal determined from Long Term Evolution (LTE) cellular radio signals mediated by at least three Long-Term Evolution (LTE) cellular base stations (BS) and at least one Location Service Client (LCS) server. This at least one Location Service Client (LCS) server includes an embedded artificial intelligence algorithm comprising a Recurrent Neural Network (RNN) algorithm to analyze the spatiotemporal positioning of both the terminals and determine a likely future trajectory of the at least one vehicle and the at least one pedestrian so as to maximize a reward metric based on Reinforcement Learning (RL) analysis. The at least one Location Service Client (LCS) server communicates the likely future trajectory of the at least one vehicle and the at least one pedestrian to the at least one terminal physically linked to the at least one pedestrian; the at least one terminal physically linked to the at least one pedestrian including an embedded Artificial Intelligence algorithm comprising a Conditional Random Fields (CRFs) algorithm to determine if the likely future trajectory of the at least one pedestrian is below a vehicle-to-pedestrian proximity threshold limit; and, if the proximity threshold limit is reached, the terminal physically linked to the at least one pedestrian communicates a collision-avoidance emergency signal to the at least one pedestrian and to the at least one vehicle that meet the proximity threshold limit.

This invention first was found novel by the examining divisions of the EPO over the cited stated of the art and although not really mentioned by the examining division, due to the fact this is obviously present, the technical effect of the present invention could be considered to be the vehicle-to-pedestrian collision avoidance by means of alarms in case of reaching a proximity threshold limit.

In summary, this patent describes a comprehensive system and method that combines LTE technology, spatiotemporal positioning, Artificial Intelligence algorithms (RNN and CRFs), and communication protocols to predict and prevent vehicle-to-pedestrian collisions by providing timely collision-avoidance signals to the involved parties wherein the Artificial Intelligence algorithms predict a future trajectory of the vehicle as well as for the pedestrian and subsequently at expected intersection of such trajectories generates an alarm signal.

When comparing this situation with other jurisdictions, for instance with the United States, it is to be noticed that a patent is allowed as well for this invention with a similar (broad) claim-scope as granted in Europe.

In the examination proceedings of the patent application it was stated that: The prior art of record is different from the claimed invention because in the claimed invention the at least one terminal physically linked to the at least one pedestrian including an embedded Artificial Intelligence algorithm comprising a Conditional Random Fields (CRFs) algorithm to determine if the likely future trajectory of the at least one pedestrian is below a vehicle-to-pedestrian proximity threshold limit; and, if the proximity threshold limit is reached, the terminal physically linked to the at least one pedestrian communicates a collision-avoidance emergency signal to the at least one pedestrian and to the at least one vehicle that meet the proximity threshold limit. This in view of the other limitations of claim 1 resulted in the claimed invention being novel and non-obvious and hence a patent is allowed for this present invention in the US as well.

Moreover, a patent is granted as well for the present invention in following further jurisdictions: CN, IL, KR and JP as well having obtained a substantial similar claim-scope. And, a well drafted patent application requires sufficiency of disclosure of the invention.