Patent law systems around the world continuously adapt to changes in the technological landscape, refining and updating tests, guidelines, and procedures to accommodate new technology. AI, however, is proving to be a particularly stubborn challenge for patent systems. The technical complexity of AI systems, their broad applicability to almost every industry sector, and their inherent overlap with abstract and mathematical concepts mean that assessing whether a given AI invention is patentable is rarely straightforward. As a consequence, the process of trying to obtain patent protection for these inventions can be particularly daunting for applicants.
This article is a practical guide to the approaches of two of the world's largest patent offices – the European Patent Office (EPO) and the United States Patent and Trademark Office (USPTO) – in assessing the patentability of AI inventions.
EPO
Excluded subject-matter at the EPO
The most immediate issue with AI inventions at the EPO is that the European Patent Convention (EPC) states that both mathematical methods and programs for computers "as such" are excluded from patent protection. While at first glance this may seem disastrous for patenting AI-related inventions, the "as such" proviso gives applicants considerable wriggle room: in effect, it means that patent protection can still be obtained for mathematical methods and computer programs if it can be shown that the invention makes a technical contribution over the prior art (according to the EPO's definition of "technical").
A technical contribution may be made by technical features present in a claim alongside non-technical, or excluded features. Additionally, features which, when considered in isolation, would be considered non-technical, may nonetheless in the context of the claim contribute to the solution of a technical problem. Consider for example a claim directed to a method for training a neural network – which in itself may essentially comprise a mathematical method. Such a claim may nonetheless be patentable if it enables the neural network to reduce the power consumption of a data centre (a tangible technical effect). Inventions of this type, containing a mixture of both technical and non-technical subject-matter, are known by the EPO as "mixed type" inventions.
The question of whether or not a mixed type invention makes a technical contribution over the prior art is considered by the EPO in the context of inventive step, and its preferred "Problem and Solution" approach to assessing the presence of an inventive step. The good news is that the EPO has a well-established approach to assessing the technicality of mixed type inventions. The bad news is that the case law defining what "technical" means is both complex and restrictive.
Assessing mixed type inventions at the EPO: the Comvik Approach
The usual approach employed by the EPO for assessing mixed type inventions is known as the Comvik approach, which was established in T 641/00. The Comvik approach consists of two steps (the second of which is simplified here for brevity):
1. First, the claimed subject-matter is assessed for technical character. In the absence of any technical character, the claimed invention is not patentable.
2. In the second step, individual features of the claims are split into three categories: non-technical features which serve no technical function, non-technical features which nonetheless contribute to the technical character of the invention, and technical features. The latter two categories – technical and contributing features – can be used as the basis for demonstrating inventive step.
In general, the first step of the Comvik approach is a relatively "low bar", with the presence of any technical feature being sufficient to pass. As noted above, the second step of the COMVIK approach is considerably more involved in practice than just the categorisation of features. However, a key take away is to remember that both features that are inherently technical, and those which, although not in themselves technical do nonetheless contribute to the solution of a technical problem, can be used to support the presence of an inventive step.
Technical character at the EPO
AI inventions are likely to be considered by the EPO as mixed type inventions, and if you want to ensure your AI invention is patentable then you'll need to look for technical character: technical features, and non-technical features which nonetheless contribute to the solution of technical problem. The EPO defines two ways (or "dimensions") in which the features of an AI invention can contribute to technical character:
1. By virtue of being adapted to a specific technical implementation (i.e. motivated by consideration of internal functioning of the computer). Such inventions may be referred to as "hardware AI".
2. By virtue of being applied to a specific technical field (i.e. application to a field which the EPO deems technical). Inventions that achieve this may be referred to as "applied AI".
Hardware AI inventions include methods specifically adapted to particular hardware configurations, and the specification should set out clearly what form that adaptation takes. For applied AI inventions, it's recommended to ensure that an application contains at least one dependent claim specifying a concrete technical use case, purpose or effect. What exactly constitutes a technical purpose is continually evolving through case law. In some examples, such as control of an industrial machine or process, the assessment can be relatively easy and clear cut. In the digital domain, the question of technical purpose can be more nuanced, but the following are examples that have been accepted by the EPO in the past as technical purposes:
- digital audio, image or video enhancement or analysis
- separation of sources in speech signals; speech recognition
- encoding data for reliable and/or efficient transmission or storage (and corresponding decoding);
- compilation of software;
- encrypting/decrypting or signing electronic communications;
- facilitating data access using data structures.
Plausibility
One particular hurdle in demonstrating an inventive step at the EPO is that it must be plausible that the teaching of an application actually solves the problem it purports to solve. Historically, this question has been far more relevant for our colleagues in the Life Sciences fields, than for tech specialists. However, in the field of AI, it may be far more difficult than is usual in the tech field to establish why an invention acts as it does, and it's often not immediately clear that a given AI method or apparatus will produce a desired result.
Consequently, we recommend that patent applications for AI inventions always include some discussion of plausibility. In some cases, it may be possible to rely on established scientific principles. Frequently however, it is helpful to include evidence that the invention works – for example by comparing outcomes with those of other methods. Discussing plausibility in the application also ensures that the applicant will be able to provide post-published evidence (evidence published after the priority date of the application) to back up the teaching of the application.
USPTO
Subject-matter eligibility at the USPTO
Similarly to before the EPO, applicants at the USPTO often come up against subject-matter eligibility issues when trying to patent AI inventions. Two landmark Supreme Court decisions – Alice and Mayo – are central to the way computer-implemented inventions are currently assessed by the USPTO. These two decisions have resulted in a "two-step" framework – the Alice/Mayo test – which the USPTO applies when deciding whether a Computer Implemented Invention (CII) meets the subject-matter requirements for patent protection.
Step 1: Does the claimed invention fit into one of the four statutory categories?
In the US, there are four statutory categories of invention which are, in theory, eligible for patent protection: processes, machines, manufactures (i.e. manufactured articles), and compositions of matter (i.e. combinations of substances, including composite articles). When it comes to inventions in the field of AI or machine learning, it's usually straightforward to show that such an invention fits into one of the four statutory categories above – generally AI inventions qualify as a process or a machine. With this step satisfied, the USPTO proceeds to Step 2A (Prong 1).
Step 2A (Prong 1): Does the claim recite one of the three judicial exceptions?
Alongside the four statutory categories of invention, the USPTO also recognizes three well-established judicial exceptions to patentability: laws of nature, natural phenomena and abstract ideas. Subject-matter which falls entirely outside of these categories – i.e. for which the answer to the above question is "no" – is generally deemed eligible for patent protection. However, applicants seeking protection for inventions in the field of AI or machine learning often face objections that these inventions are excluded from patent protection by virtue of being one or more of the types of abstract ideas recognised by the USPTO – in particular, mathematical concepts or mental processes. So, for AI inventions, the answer to this question is unfortunately usually "yes". This takes us to Step 2A (Prong 2).
Step 2A (Prong 2): Does the claim recite additional elements that integrate the judicial exception into a practical application?
In assessing this question, the USPTO looks at elements of the claim beyond those which are judicial exceptions – for example, elements of the claim which aren't mathematical concepts or mental processes. It will then consider whether these elements impose a meaningful limit on the judicial exception (in such a way that, for example, the claim does not lead to a complete monopoly over a mathematical equation or relationship), to determine whether the judicial exception is integrated into a practical application.
If the answer to this question is "yes", then the USPTO's subject-matter requirements are satisfied. However, if the claim does not recite additional elements that integrate the exception, we proceed to Step 2B.
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
This step of analysis is effectively the final chance for a set of claims to satisfy the USPTO's subject-matter requirements. If the claim, as a whole, amounts to "significantly more" than just the exception – in other words, if there is an inventive concept in the claim – then the claim is eligible. In this final step, the USPTO considers whether there is any specific limitation in the claimed invention "other than what is well understood, routine, conventional activity" in the field.
Outcomes of the EPO and USPTO approaches
Perhaps surprisingly given their different legislation and procedural approaches, there is some common ground between the EPO and the USPTO – in particular, AI applications stand a better chance of being granted in Europe or the US if the claims tie the underlying technology to a specific tangible application ("technical" in Europe, "practical" in the US). However, the fact remains that the two approaches are derived from completely unrelated legislation, and the differences between them can, in some cases, produce very different results for applicants.
Knowing the ins and outs of the EPO's and USPTO's approaches to AI subject-matter is essential in order to obtain patent protection for AI inventions in either jurisdiction. In our webinar, A Practical Masterclass in Patenting AI at the EPO and USPTO, we explore the nuances of the two systems in more detail, look at some direct comparisons between EPO and USPTO outcomes, and offer some practical tips to applicants looking to protect AI inventions in Europe and the US.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.