One issue with patenting AI-related inventions is how to claim subject matter that will be of commercial value to the applicant. This is not as straightforward as in many other technical fields because of the way AI inventions are developed and used. Typically, this requires a stage of selecting and/or designing a machine learning model, the careful curation of data for training the model, training, testing and refining the model. The result can be a complex algorithm that cannot always practically be defined in words, as is needed for a patent claim.
The ideal protection for the applicant would often include protection for the trained model, which would enable the applicant to have a monopoly over the provision or use of the model itself. If the protection from the patent only covers the development aspects, the patent may not be usable to stop third parties distributing or using copies of the model.
While the EPO’s case law on AI-related inventions is at an early stage, there is some established EPO practice that could be helpful in informing us how to protect AI-related inventions. In general, the EPO considers there to be two types of claim: a claim to a product and a claim to a method. However, there are two special provisions that blur this distinction and enable a product to be protected using method terminology, rather than describing the product itself. This might provide some promising opportunities for protecting AI inventions by protecting a model obtained by the design and/or training process.
Protection of a product via a method claim
Article 64(2) EPC states that:
‘If the subject-matter of the European patent is a process, the protection conferred by the patent shall extend to the products directly obtained by such process.’
This means that, even though a claim is to a method, the claim can include protection for a product obtained using that method.
However, this additional product protection does not apply to all method claims. The EPO considers there to be two types of method claim:
1. the use of an entity to achieve a technical effect
2. a process to produce a product
A method claim of type 1 does not result in the production of a product, but is only a method that uses a pre-existing entity. As confirmed indirectly by the decision of the Enlarged Board of Appeal G 2/88, only a method claim of type 2 is given the additional product protection. See EPO Guidelines For Examination F-IV, 4.16.
When drafting AI-related method claims, it is therefore important to ensure that the claim will be treated as a type 2 claim, i.e. a claim that results in a product being produced. It is arguable, if not yet established, that a claim to, e.g. a method of training a model would be considered as a type 2 claim. The argument would be that the trained model is something that was produced. A contrary point of view, however, could be that the method is merely using a training process on a pre-existing model to achieve a technical effect.
Protection for direct products of processes is implemented in national law of the EPO member states, and the EPO itself does not rule on matters of infringement, so this question only comes up at the EPO when an applicant tries to change a method claim to a product claim after grant. This is only allowed for a type 2 claim, as the scope of protection of a European patent cannot be extended after grant. EPO case law on the matter may therefore be a long time coming. We are also not aware of national case law on the point of “direct product” in relation to AI at this time, but we will review national case law in relation to the protection for direct products of methods in general to see how this can inform AI drafting in a future instalment.
In summary, while there is nothing definite here, yet, the possibility for protecting a trained model by way of a method claim to training the model makes it, we think, worthwhile for applicants to include “method of training” claims in their application. It is probably also worthwhile to include alternative claim wording, for example in the description, and broadly recite methods of “producing a model” and the like, also making sure that the trained model is the direct result of the process. For example, a claim including both steps of training and using a model would be undesirable for this and other reasons.
The second special provision is a “product-by-process” claim. This is a claim explicitly to a product, but, unlike a normal product claim in which the features of the product are described, the product is defined by the way it is made.
The EPO will only allow a product claim to be expressed in a product-by-process form if ‘it is impossible to define the claimed product other than in terms of a process of manufacture’ (Guidelines F-IV, 4.12). This situation could well apply to AI inventions, where the model is produced by a training process that can be described, while the resulting model (a complex collection of parameters in conjunction with an algorithm) often cannot practically be described.
When it comes to examining a product-by-process claim for patentability, the focus is on the resulting product, rather than the method of production. So, it is necessary to show that the resulting product is itself new and inventive over prior art products, however they were produced. This may turn out to be difficult for AI inventions – although it is likely that different training methods lead to different models, different runs of the same training method also do so, which raises the question of whether AI models can be clearly defined in this way. Equally, when a distinction over the prior art can be encapsulated in the model structure, this may be preferable in any case. However, there may be a place for product-by-process type claims to come to the rescue if it turns out that the way the model is trained is essential for the invention, so that reference to a specific training method may save the day.
With the above in mind, it may be worth including claims (or claim-like wording in the description) of the type “computer-readable medium or media (CRM) encoding instructions that define a machine learning model trained/produced using any preceding claim” and also computer systems comprising such a CRM.
While the case law on AI inventions is in its early stages, it is a good idea to try distinct types of claims to ensure that the applicant obtains the best possible scope of protection. Even if current EPO practice would reject some of these claims, it is possible that future case law developments will cause them to become allowable later. Without applicants trying to push the boundaries, new case law is likely to be generated more slowly. However, if distinct types of claim wording are not present in the application as filed, it is unlikely that the applicant would be able to make use of these options later in prosecution or in litigation after grant, due to Europe’s strict requirements for amendments to be unambiguously disclosed in the application as filed. Progress at the EPO in terms of protecting AI inventions may well be driven by applicants who take an innovative approach to claim drafting.
If you have any questions, please contact Alexander or your usual Kilburn & Strode advisor.