I. Introduction
Artificial intelligence (“AI”) is one of the most exciting technologies of our time. Although AI has been a field of research for over sixty years, it is only in recent years that it has begun to realise its potential.
One factor in AI’s coming of age has been the development of new machine learning algorithms, through which a computer can learn to perform a particular task without being explicitly programmed. The growth of machine learning algorithms has been fuelled by the availability of vast quantities of data from which those algorithms can learn, and ever more powerful computer hardware with which to process that data. (In this article, the term “artificial intelligence” is used to refer to a broad range of technologies that includes machine learning algorithms.)
The rise to prominence of AI has been accompanied by a significant increase in the number of patent applications for AI. For example, the EPO has seen more than a fifty per cent increase in the number of European patent applications that broadly relate to AI over the period from 2010 to 2014 (the most recent year for which the EPO’s statistics are available), with around six thousand such applications in 20141. Over the same period, the EPO has seen nearly a threefold increase in patent applications for so-called “core AI” technology.
This article explains the EPO’s practice for examining patent applications relating to AI, and discusses the challenges faced when trying to patent AI at the EPO.
II. New Guidelines for Examination in the EPO
Although there has been a steady stream of European patent applications relating to AI for many years, the EPO had not published any guidance on its practice for examining such applications until last year. The uncertainty faced by patent applicants in this area was compounded by the dearth of case law specifically relating to AI.
The November 2018 edition of the Guidelines for Examination in the European Patent Office (hereafter referred to as “the Guidelines”) set forth the EPO’s practice for examining AI-related inventions for the first time.
In accordance with the new Guidelines, inventions relating to AI are examined in the same way as inventions involving mathematical methods2. The EPO’s rationale for treating AI in this manner stems from an observation that many artificial intelligence and machine learning techniques are based on computational models and algorithms. In the EPO’s view, such computational models and algorithms are inherently of an “abstract mathematical nature”, and so should be treated in the same way as other mathematical methods.
The EPO’s decision to treat AI as a particular species of mathematical method creates a presumption that claim features relating to AI, such as an artificial neural network or a support vector machine, are non-technical. Hence, these features alone cannot result in a claim being seen to define an invention within the meaning of Article 52(1) EPC, although this hurdle can easily be overcome by including some other technical means in the claim (such as by reciting the processing hardware used to implement the AI technique, or by reciting a “computer-implemented method” in the preamble of the claim). More importantly, features relating to AI cannot support the presence of an inventive step when the Comvik approach3 is applied.
However, the presumption that claim features relating to AI are non-technical can be overcome if the claim falls into either of two “safe harbours” that are set out in the Guidelines4. The first safe harbour applies when a claim is limited to a “technical application” of a mathematical method. The second safe harbour applies when a claim is directed to a “technical implementation” of a mathematical method. The two safe harbours are discussed in more detail below.
If the claim falls within either of the safe harbours, then the claim features relating to AI are deemed to be technical. Assuming that those AI-related features are novel, they can be relied upon to demonstrate the presence of an inventive step over the prior art.
A. Technical Application
A claim falls within the first safe harbour when it is functionally limited to a specific technical purpose.
The Guidelines give several examples of technical purposes that, if recited in the claim, would allow the claim to benefit from the first safe harbour. The most relevant of these technical purposes to the field of AI are quoted verbatim below:
- controlling a specific technical system or process, e.g. an X-ray apparatus or a steel cooling process
- 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
- providing a medical diagnosis by an automated system processing physiological measurements
The Guidelines emphasise that the technical purpose must be specific. Merely reciting a generic purpose, such as “controlling a technical system”, is not enough.
B. Technical Implementation
A claim falls within the second safe harbour when it is directed to a specific technical implementation of the mathematical method.
The Guidelines explain that this safe harbour applies when the mathematical method is particularly adapted for that implementation, in the sense that the design of the mathematical method is motivated by technical considerations of the internal functioning of a computer.
The Guidelines do not give an example of the second safe harbour that specifically relates to AI. However, an example of a claim that might fall within the second safe harbour is claim 1 of European Patent No. 1 569 128. That claim related to a “computer-implemented method for processing a computer application”, but did not recite a specific technical purpose of the method. The claim did, however, include a detailed recitation of how a central processing unit (“CPU”) interacts with a graphics processing unit (“GPU”) to perform a machine learning technique, and specified various types of data that are communicated between the CPU and GPU. Although this patent was granted by the Examining Division in 2015, without being considered by the Boards of Appeal, it serves as an example of an AI-related mathematical method that the EPO found patentable.
The two safe harbours may be exclusive. Thus, a claim that falls within the second safe harbour need not recite any technical purpose. It is possible for a claim to recite an overtly non-technical purpose (such as a business-related purpose) and nevertheless benefit from the second safe harbour5.
III. Patentability of Some Exemplary AI Technologies
There is very little case law that specifically relates to AI. The following paragraphs review some of the small body of case law that exists, together with the new Guidelines, to give some examples of AI technologies that are, and are not, regarded as patentable by the EPO.
- Core AI
“Core AI” is a term used by the EPO to refer to the fundamental building blocks of AI and machine learning, as opposed to the applications of AI. For example, an artificial neural network6 would be regarded as a core AI technology.
The new Guidelines suggest that it will now be difficult to patent innovations in core AI. A machine learning algorithm that is new and non obvious may nevertheless be deemed to lack an inventive step, as the algorithm could be regarded as a mathematical method that lacks the technical features needed to support an inventive step. Although an inventive step objection could be overcome by amending the claims to take advantage of either of the safe harbours discussed above, this would of course have the undesirable effect of sacrificing some scope of protection. Furthermore, such amendments will not be possible when the application does not disclose either a technical purpose to which the algorithm can be applied or a specific technical implementation of the algorithm.
The Guidelines’ treatment of artificial intelligence and machine learning techniques as mathematical methods has some basis in case law. In T 22/12, the Board found that a support vector machine7 was a mere mathematical method that did not provide a technical effect.
Nevertheless, the EPO has a long history of granting patents for core AI. As just one example, European Patent No. 0 554 083 B1, which was granted in 1999, claims a “neural network learning system” that learns a probability density for relating input data to output data. The claims do not recite a specific technical application of the neural network learning system, nor do they recite any details of how the neural network learning system is implemented in hardware. If these claims were to be examined in accordance with the new Guidelines, they could well be said to be directed to a pure mathematical method that is devoid of an inventive step.
- Natural Language Processing
Natural language processing (“NLP”) refers to techniques that allow a computer to interpret inputs, and to generate outputs, in human languages such as English, Spanish or Chinese.
A well-known application of NLP is in virtual assistants, such as Amazon’s Alexa and Apple’s Siri. Virtual assistants use a range of NLP techniques, such as: speech recognition, to transcribe a user’s speech into text; natural language understanding, to interpret the meaning of the text; and natural language generation, to produce a response in the language of the user.
The EPO considers some NLP techniques to be more patentable than others. At one end of the spectrum lies speech recognition, which the Guidelines expressly recognise as a “technical purpose”, and which is readily patentable. At the other end of the spectrum lies natural language understanding, which is considerably more difficult to patent.
The difficulties in patenting NLP, and other technologies involving linguistics, are longstanding. In T 52/85, the Board considered a system for automatically generating a list of expressions whose meaning was related to an input linguistic expression. The Board held that the relationship between the input and output expressions was not of a technical nature, but was instead a matter of their “abstract linguistic information content”. The Board consequently found that the claimed subject-matter was unpatentable.
In another relatively old decision, T 1177/97, the technology at issue related to machine translation. The Board again found the claimed subject-matter to be unpatentable, stating “Features or aspects of the method which reflect only peculiarities of the field of linguistics, however, must be ignored in assessing inventive step.” This statement is often quoted by examiners when applying the Comvik approach to inventions in the field of natural language processing. Although the Board in T 1177/97 also held that “information and methods related to linguistics may in principle assume technical character if they are used in a computer system and form part of a technical problem solution”, it is hard in practice to convince the EPO that a technical problem is solved by the linguistic aspects of an invention.
- Classification Algorithms
Classification algorithms are able to categorise input data into one of a number of distinct classes. For example, classification algorithms are commonly used to analyse medical images, and categorise the images according to whether they do, or do not, indicate the presence of a disease.
The answer to the question of whether classification algorithms are patentable hinges upon the type of data that is being classified.
According to the Guidelines, classifying text documents solely in respect of their textual content is not a technical purpose and, therefore, is unlikely to be patentable8. The Guidelines refer to T 1358/09, which related to classifying a document based on the frequency of occurrence of a particular term. The Board in T 1358/09 held that this technique was not technical, and thus lacked an inventive step, stating:
“Classification of text documents is certainly useful, as it may help to locate text documents with a relevant cognitive content, but in the Board's view it does not qualify as a technical purpose. Whether two text documents in respect of their textual content belong to the same "class" of documents is not a technical issue.”
There are a few other decisions in which the Boards of Appeal have found algorithms for classifying text documents to be unpatentable. For example, in T 22/12, the Board found that a method of classifying electronic messages to detect “junk” messages lacked an inventive step. Although the Board acknowledged that an email has technical properties, this was not enough to save the application from refusal because only the textual content of the email was classified. It is interesting to note that another Board reached a different conclusion in T 1028/14, which also related to detecting undesired messages. In T 1028/14, however, a message was classified based upon factors such as the IP address from which it originated, rather than based upon the textual content of the message. Thus, the data that was used to classify a message in T 1028/14 could be said to have a more “technical” nature than that in T 22/12.
Whereas text classification is usually regarded as non-technical, the EPO takes a more favourable view when other types of data are classified. In particular, the Guidelines acknowledge that the classification of digital images, videos, audio and speech signals is a technical purpose.
T 1286/09 lends some support to the Guidelines’ position that image classification is technical. The technology at issue involved manipulating various properties of an exemplar image, and then using the resulting image to train a classifier. The Board reversed the first instance decision that this subject-matter lacked an inventive step over the prior art. However, the technical character of the subject-matter was never in question and, therefore, the Board did not explicitly confirm that image classification should be regarded as a technical purpose.
The EPO’s strict approach towards text classification is hard to reconcile with its more lenient approach towards classification of other types of data. One possible explanation might be that in one of the earliest decisions on the patentability of computer programs and mathematical methods, T 208/84 (Vicom), a method of digitally filtering an image was found to be technical because the image was considered to be a physical entity. Thus, the early case law may have created a view that images are technical, physical entities, whereas text documents are nothing more than their linguistic content.
- Recommendation Systems
Recommendation systems provide suggestions for content that is likely to be of interest to a particular user. For example, video streaming services may use recommendation systems to suggest movies that a user might enjoy. As another example, merchants may use recommendation systems to suggest products that the user might wish to purchase. Recommendation systems learn from the user’s previous behaviour to predict what content is likely to interest the user in the future.
In T 306/10, the Board held that a recommendation system lacked an inventive step and stated:
“In the Board’s view, the selection of an item, for example a song, for recommendation to a user does not qualify as a technical purpose. From a technical point of view it is irrelevant what songs are recommended to a user”.
Although recommendation systems are of great commercial importance to many online businesses, they are extremely difficult to patent.
V. Claiming AI-Related Inventions
AI-related inventions may have three distinct, and potentially patentable, aspects:
- Generating training data for use in training a model, such as an artificial neural network;
- Training the model using the training data; and
- Using the trained model to analyse data.
Each of these aspects may be performed by a different party, and those parties may be located in different jurisdictions. Thought should be given to drafting separate independent claims for each of the three aspects, where appropriate.
In addition to claiming each of the three aspects listed above, one should also consider claiming the trained model itself. Such a claim may have commercial value if, for example, the trained model can be deployed separately from the software or device that uses the model to analyse data.
Claiming the trained model is difficult in practice. The trained model may be no more than a set of numbers representing, for example, the weights of an artificial neural network. The trained model may thus be difficult, or even impossible, to define in concrete technical terms. A possible way to overcome this difficulty might be to draft a product-by-process claim. For example, if the patent application claims a method for training a model, then it is simple to draft a claim directed to a model trained in accordance with that method. However, given that a trained model may just be a set of numbers, it might be difficult to demonstrate that the model is new and inventive.
Even if a product-by-process claim is not allowable, a well-drafted claim to a method of training a model will confer protection on the model itself under Article 64(2) EPC. This emphasises the importance of having an independent claim directed to the training method.
VI. Conclusion
As the number of patent applications relating to AI increases, the new Guidelines represent the EPO’s attempt to draw a line between those AI technologies that are patentable and those that are not. In some cases, the Guidelines will assist applicants by providing new lines of argument that can be raised during examination. In the case of core AI, however, they may make it difficult to patent the type of technology that has been patentable for many years.
Footnote
1 The EPO’s statistics were presented by Yann Ménière, Chief Economist at the EPO, during his keynote speech at the EPO’s “Patenting Artificial Intelligence” conference on 30 May 2018.
2 Guidelines for Examination in the EPO, Chapter G-II, 3.3.1
3 T 641/00 and Guidelines for Examination in the EPO, Chapter G-VII, 5.4
4 Guidelines for Examination in the EPO, Chapter G-II, 3.3
5 See T 2330/13, for example.
6 An artificial neural network is a machine learning technique that maps input data to output data using layers of interconnected nodes, similar to the connections between neurons in the nervous systems of living beings.
7 A support vector machine is a machine learning algorithm that classifies input data, such that the data is assigned to one of two distinct classes.
8 Guidelines for Examination in the EPO, Chapter G-II, 3.3.1
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.
