Patenting Artificial Intelligence

Following its widely attended “Patenting Artificial Intelligence” conference in May this year, the European Patent Office (EPO) has published new Guidelines on the patentability of Artificial Intelligence (AI) and Machine Learning (ML). In this article we will take a look at the patentability criteria for AI/ML inventions in the light of this new guidance, as well as some of the key questions which it still leaves open.

Broadly speaking, the EPO approach is that patent protection is available for any new and non-obvious technical development. To understand how this approach is applied to AI/ML inventions, it is helpful to first consider the key concepts of obviousness and technicality, as both play a critical role in the assessment of patentability.

Obviousness and technicality

It is a fundamental tenet of the patent system that it should not be possible to patent what is obvious. This is as true for AI and ML as it is for any other technology. Hence there is a requirement for an “inventive step”, which is taken to be satisfied if an invention is not obvious to a person skilled in the art. In addition to the non-obviousness criteria, in most patent systems there is also an eligibility criteria which places a further restriction by defining certain subject matter or activities which cannot be patented (even if not obvious). Article 52(2) of the European Patent Convention (EPC) gives a non-exhaustive list of subject matter or activities which are excluded as such, including discoveries, mental acts, business methods etc.  

The EPO approach takes this further by requiring that patent protection is only available for technical inventions, such that a patentable claim should specify a "technical solution to a technical problem". It is by developing this line of jurisprudence over the last 30 years that the EPO Boards of Appeal have been able to extend the patent system of the EPC, which was originally designed to deal with innovations in hardware, to also cope with innovations in software, and so now also innovations in AI and ML. 

Under the EPO approach obviousness and technicality are intertwined together in the assessment of inventive step; only those features which contribute to the technical character of the invention are taken into account in this assessment. On the other hand, if a feature cannot be considered as contributing to the solution of any technical problem by providing a technical effect it is given no significance in the assessment of inventive step; it is ignored. Thus for inventions relating to AI and ML, a crucial question is often which (if any) features of the claimed invention contribute to its technical character? 

Technical character of AI inventions

The new guidance proceeds on the basis that the computational models and algorithms on which AI and ML are based are per se of an abstract mathematical nature. This means that, for example, a neural network as such is considered to be a mathematical rather than a technical feature. However, crucially, the guidance sets out that such features can contribute to technical character in two ways: by application to a field of technology and/or by being adapted to a specific technical implementation. We will look at these two situations, application and implementation, in turn.

Technical application 

It is clear from the new guidance that AI/ML developments will be considered technical when applied to a specific field of technology. A variety of examples of technical applications are provided, including classification of digital images, videos, audio or speech signals based on low-level features, image or video analysis or enhancement (e.g. detecting persons in a digital image, de-noising), speech recognition, and encryption/decryption. (See below “technical application areas” for further examples).

These examples accord with our day-to-day experience with the EPO, and with the case law that has been established by the EPO Boards of Appeal, on which the new guidance is based. Nevertheless that case law is itself relatively limited, and thus the new Guidelines leave many questions unanswered. For instance the guidance is silent on the technicality of some key use cases for AI, such as machine translation and natural language processing, and the case law on these areas is relatively sparse. While some might argue that natural language processing relates to a linguistic purpose rather than a technical one, this seems difficult to reconcile with the fact that that speech recognition is considered technical. Moreover while image classification based on low level features is technical, text classification is ruled out. Some would say this leads to a contradiction – if a new AI/ML development is applicable to both image and text classification, limitation to image classification seems arbitrary and technically irrelevant. 

These issues point to a broader difficulty – many AI/ML inventions are applicable to a number of different technical application areas; how should such inventions best be protected? For such an invention it would seem unfair to require limitation to one specific technical use case or another, or to require an applicant to file multiple divisional applications, one for each technical application. While it is possible that there could in some cases be a drafting solution (e.g. multiple independent claims, or an “OR” statement in the independent claims), the extent to which the EPO would permit this is at present somewhat unclear. It is true that in some cases, such an invention might be considered technical because it falls within the second situation mentioned in the Guidelines, i.e. because it is adapted to a “specific technical implementation”, and that in that case there would be no need to limit to a particular technical application. However this will only apply if particular criteria are met (which we will discuss below) and so does not fully resolve the difficulty.

Technical implementations – protecting core AI

The new Guidelines state that independently of any technical application, a mathematical method (i.e. algorithm) may also contribute to the technical character of the invention “when the claim is directed to a specific technical implementation of the mathematical method and the mathematical method is particularly adapted for that implementation in that its design is motivated by technical considerations of the internal functioning of the computer”. 

The interpretation of this wording is crucial for the patentability of AI/ML inventions that are not limited to any specific technical use case, such as inventions relating to core AI. However its precise meaning is somewhat vague; for example there is a question about whether or to what extent there needs to be a link to hardware in order for a mathematical method to be considered a “specific technical implementation”. Some have suggested that a specific technical implementation requires specifically adapted hardware, but we do not read the new guidance in this way. It is true that the algorithm must be particularly suited for being performed on a computer, but the computer could be entirely conventional. The Guidelines give the example of the adaptation of a polynomial reduction algorithm to exploit word-size shifts matched to the word size of the computer hardware: it is the algorithm which is adapted to the hardware, not the other way around. Thus, algorithms which are specifically adapted for e.g. a memory efficient implementation, or for parallelization across multiple processor cores, or faster training, may in particular cases be argued to fall within the category of a “specific technical implementation”, and so considered technical. Nevertheless the boundary delineating what is and what is not a “specific technical implementation” remains somewhat unclear, and more case law is needed to clarify the patentability criteria in this area, in particular for inventions relating to core AI.

Final thoughts

The new EPO Guidelines provide helpful guidance on the situations in which AI/ML will be considered technical, referring in particular to developments adapted to a specific technical implementation or applied to a specific field of technology. However as mentioned above, a difficulty arises in circumstances where an AI/ML innovation has multiple technical use cases and does not fall within the guidance relating to “specific technical implementation”. This points to a difference between AI/ML and many other innovations in that many AI/ML developments are applicable to multiple fields. In denying that AI/ML is itself a specific technical application area, the new Guidelines leave room for ongoing debate and further case law on the patentability of innovations in this developing field.


Examples of technical application areas

  • classification of digital images, videos, audio or speech signals based on low-level features (e.g. edges or pixel attributes for images);
  • the use of a neural network in a heart-monitoring apparatus for the purpose of identifying irregular heartbeats;
  • controlling a specific technical system or process, e.g. an X-ray apparatus or a steel cooling process;
  • determining from measurements a required number of passes of a compaction machine to achieve a desired material density;
  • digital audio, image or video enhancement or analysis, e.g. de-noising, detecting persons in a digital image, estimating the quality of a transmitted digital audio signal;
  • separation of sources in speech signals; speech recognition, e.g. mapping a speech input to a text output;
  • encoding data for reliable and/or efficient transmission or storage (and corresponding decoding), e.g. error-correction coding of data for transmission over a noisy channel, compression of audio, image, video or sensor data;
  • encrypting/decrypting or signing electronic communications; generating keys in an RSA cryptographic system;
  • optimising load distribution in a computer network;
  • determining the energy expenditure of a subject by processing data obtained from physiological sensors; deriving the body temperature of a subject from data obtained from an ear temperature detector;
  • providing a genotype estimate based on an analysis of DNA samples, as well as providing a confidence interval for this estimate so as to quantify its reliability;
  • providing a medical diagnosis by an automated system processing physiological measurements;
  • simulating the behaviour of an adequately defined class of technical items, or specific technical processes, under technically relevant conditions.