The art of claim drafting when protecting AI systems
With AI’s recent breakthrough in machine learning, now more than ever inventors are looking for ways to protect AI systems. But obtaining patent protection depends on the right claims.
What is AI and how does it work?
Artificial intelligence (AI) is the simulation of human intelligence processes by machines, principally computer systems. These processes include learning (the acquisition of information and rules for using information), reasoning (using rules to arrive at estimated or definitive conclusions) and self-correction.
Patent Offices new practices for AI’s
Overwhelming figures from the World Intellectual Property Organisation depict AI’s technology trends showing more than 300,000 active AI related patent families. This shows that the number of inventions including some form of AI is rising rapidly.
Reaching a definitive number like this is not an easy task, as at present, the EPO classification system does not yet have a specific class for AI. As a result, one must use classes for software and neural networks when patenting inventions involving some form of AI. Because of the difficulty of going through different classifications when searching AI-related applications, it is expected that in the coming years the EPO will introduce a specific class for AI, as the number of applications continues to rise.
For now, the EPO has merely updated its guidelines for examination to reflect the increasing volume of AI related applications. These extended guidelines have been in force since November 2018, and can be found in the new chapter G-II, 3.3.1.
The new guidelines established that inventions relating to AI shall be examined in the same manner as inventions involving mathematical methods, which are considered as non-technical/abstract, and hence unpatentable. This would lead to the presumption that the EPO is of the opinion that claim features relating to AI, for example artificial neural networks, are also non-technical/abstract. Thus, these features alone cannot satisfy the meaning of an invention according to Article 52(1) EPC. However, this obstacle may be overcome if some other technical means are included in the claim, for example one could claim the processing hardware employed to carry out the AI technique.
Protecting a training method
As a training method is essentially what makes a software AI, it is important to examine how this new addition to the guidelines will be applied.
There are not yet any EPO decisions that specifically deal with the protection of a training method. The reason for this being that all the AI related applications have been filed within the last years and only a few of them have been appealed, which can take two to three years to be handled by the Boards of Appeal. So we can expect a lot of such decisions in the next coming years. At the moment it is hard to envisage the protection of a training method because you would have to argue a technical effect.
Interestingly, there is a statement in the new section in the guidelines which provides that “where a classification method serves a technical purpose, [the step] of [...] training the classifier may also contribute to the technical character of the invention if [it] supports achieving that technical purpose”
It has been speculated that the EPO’s rationale is that, if you have a classification method, that protects the complete AI method and you include a training step, this may contribute to inventive step.
Bearing in mind that a classification method aims to identify the category of a new observation amongst a set of categories, a proposed way to protect a training method would be:
- To claim an AI system that is configured to execute a novel and non-obvious classification method serving a specific technical function, based on a classifier,
- And to claim a non-obvious method for training the classifier of the AI system according to claim 1, wherein the training supports the technical purpose.
Thus, if you can show that the training enhances the technical purpose of the invention once you have claimed a feature which supports a technical purpose, one would receive protection for the training method. However, this has not been decided yet, but appears to be a viable option, although close attention must be taken in relation to clarity per T 2026/15.
Upon reflection and careful interpretation of the new EPC guidelines, it can be concluded that the EPC provides for several options to protect almost all aspects of AI systems. In order to know how these options work out exactly and what will be considered as technical and non-technical AI, we will have to wait for the EPO’s Boards of Appeals upcoming decisions.