Three reasons why patent applications for applied AI inventions do not succeed at the EPO
AI is being adopted as a tool for solving the most challenging problems in a wide range of technical fields, including healthcare and life sciences.
At the EPO, AI inventions applied to all fields of technology are considered to be computer-implemented inventions. As for computer-implemented inventions generally, how a patent application for an applied AI invention is drafted has a considerable impact on its success in examination.
The technicality requirement is perhaps the most well-known (and most notorious) requirement for computer-implemented inventions at the EPO and is is assessed for applied AI inventions in the same way as it is for other types of computer-implemented inventions.
As summarised by the Enlarged Board of Appeal in G 1/19, the technicality requirement comprises two distinct hurdles. The first asks whether the invention falls within the definition of ‘invention’ in Article 52 of the European Patent Convention (EPC). To do so, there must be at least one feature that is technical when considered in isolation. Once the first hurdle is overcome, the second occurs during assessment of inventive step under Article 56 of the EPC, where the features supporting inventive step (ie, those distinguishing the invention over the closest prior art) must contribute towards a technical effect serving a technical purpose.
The immediate question arising from the two hurdles is: what does ‘technical’ mean? Unfortunately, what the EPO considers to be technical is not rigidly defined, not even in case law, which can present a challenge for particular (though certainly far from all) applications of AI. That said, any subject matter listed in Article 52(2) of the EPC, which includes programs for computers and mathematical methods, is not deemed to be technical.
Despite being a combination of a program for a computer and mathematical method, applied AI inventions seldom fall at the first hurdle. Their general implementation on a computer implies the use of a processor, which is sufficient to overcome this.
For the second hurdle, it is much less clear-cut for applied AI inventions, and this is where a well-drafted patent application can be a considerable advantage. While full assessment of inventive step is fact-dependent, the second hurdle may generally be overcome where the AI is used to solve a technical problem in a field of technology. The EPO gives the examples of the use of a neural network in a heart-monitoring apparatus for the purpose of identifying irregular heartbeats and the classification of digital images, videos, audio or speech signals based on low-level features (eg, edges or pixel attributes for images) as technical applications of AI.
In view of the second hurdle, it is best practice, when drafting, to state prominently in the application the technical problem solved by the applied AI invention, and to avoid statements relating to problems that are not technical. It is also very useful to include any considerations for the applied AI invention arising from the specific field of technology or science .
The claims of applied AI inventions should be drafted so that they either explicitly state, or at least directly imply, the particular technical application. Often this involves defining the technical use of the output of a trained AI model. Without this, the EPO may conclude that a claim does not contribute towards a technical effect serving a technical purpose across its entire scope, which would cause the claim fails the second hurdle. The March 2022 version of the EPO Guidelines for Examination refers to a notable exception to this: when a measurement from external physical reality is used as the input and the application of AI results in an indirect measurement that calculates or predicts the physical state of an existing real object. In this scenario, a technical effect is held to arise, regardless of what use is made of the indirect measurement. However, this scenario will not be relevant to all applied AI inventions.
The sufficiency requirement, which arises from Article 83 of the EPC, is where considerations for applied AI inventions begin to deviate from computer-implemented inventions.
Article 83 of the EPC stipulates that a patent application must disclose the invention in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art. For computer-implemented inventions, this requirement is generally considered to be satisfied by a functional description of the invention. Rather than describing how the steps of a computer-implemented invention are programmed in a computer, it is usually sufficient to define the function of each step in a patent application. However, for applied AI inventions, particularly those that rely on trained AI, it has been shown by T 161/18 of the Board of Appeal that a functional definition is not necessarily enough to meet the sufficiency requirement.
To make sense of this distinction, it is useful to consider what the person skilled in the art is required to do, in order to carry out the invention beyond that which is disclosed in the application.
For computer-implemented inventions more broadly, the person skilled in the art must program the defined functionality on a computer before the invention can be carried out. As the act of programming itself is not considered by the EPO to be a technical endeavour, it is perhaps no surprise that a computer-implemented invention is deemed to be sufficiently disclosed without mention in the application of how the functionality is programmed on a computer.
For applied AI inventions, while an element of programming is still necessary to turn defined functions into a practical reality, this alone is usually not enough. Although a trained AI model might be thought of as a black box with certain functionality, typically more than an act of programming is needed to recreate that functionality. The architecture of the model being trained and the algorithm used for training may play important roles in defining functionality, but perhaps most significant are the trained parameters of the model or, more commonly, the training data used to arrive at these parameters, because the nature of the training data has a direct impact on the functionality learnt by the black box. Accordingly, for patent applications for applied AI inventions, it is usually necessary to disclose the training data (or trained parameters) in the application to meet the sufficiency requirement.
The requirement to disclose training data immediately raises the question of how detailed it should be. The EPO’s view is that it depends on the nature of the invention. Meeting the sufficiency requirement in some cases might require disclosure of the specific training data, but in others, describing the type used for training (eg, “images of human faces”) could be enough. For drafting patent applications for applied AI inventions, then, the nature of the training data required to solve the technical problem to which AI is being applied is likely to the deciding factor how detailed the training data needs to be in the application.
Where training data is difficult for the person skilled in the art to obtain, as in T 161/18 of the Board of Appeal, then the patent application should disclose how it can be acquired in a quantity that is adequate for training. Such details help to show that the application is not speculative and can be put into effect.
Plausibility relates to whether a technical effect of the invention is made credible (plausible) from the patent application as filed, and often arises for chemical inventions. Unlike the technicality and sufficiency requirements, there is no statutory basis in the EPC for an invention being plausible but, nevertheless, plausibility has featured in numerous Board of Appeal decisions under either inventive step (Article 56 of the EPC) or sufficiency (Article 83). Although there are ongoing questions about the extent to which a technical effect of the invention must be made plausible from the patent application, as filed in view of G 2/21 of the Enlarged Board of Appeal, plausibility remains an important consideration when drafting applications for chemical and biological inventions.
While plausibility is not commonly raised as an issue outside of these types of inventions, we have heard rumblings of it being applied to computer-implemented inventions. For example, in T 2147/16 the Board of Appeal took the view that the claims lacked inventive step because the alleged technical effect was not specifically and sufficiently documented in the patent application as filed (although, in that case, the emphasis was more on the technicality requirement). Moreover, G-II, 3.3.2 of the Guidelines for Examination, which has recently been updated, states that where an alleged improvement is not achieved because a computer-implemented simulation is not accurate enough for its intended technical application, this may be taken into account in the assessment of inventive step or sufficiency.
Plausibility has not yet arisen for applied AI inventions, specifically. However, this could be a logical extension given the nature of applied AI inventions, particularly those that use a trained AI model. As the latter might be thought of as a black box with certain functionality, it may be difficult to determine, without further information in the patent application, whether and how the black box solves the technical problem at issue. Therefore, to mitigate potential future plausibility issues, it may be helpful to include data in the application showing the applied AI invention achieving its intended technical effect, or, at the very least, an explanation as to why the technical effect is achieved by the invention.
This is an Insight article, written by a selected partner as part of IAM's co-published content. Read more on Insight
Copyright © Law Business ResearchCompany Number: 03281866 VAT: GB 160 7529 10