European Patent Office provides guidance on AI patentability

The European Patent Office (EPO) recently issued guidance on the patenting of artificial intelligence (AI) and machine learning (ML) technologies. What does this mean for businesses in the field?

Summary

These guidelines offer some much-needed clarity for would-be Skynet creators on whether the mathematical processes and statistical models which underpin the technology will be what hinders their patentability. The main points to take from the guidance are:

• The computational models and algorithms which comprise AI/ML technologies constitute mathematical methods per se, and are not regarded as patentable under the European Patent Convention;
• The use of expressions such as “neural network”, “support vector machine”, and “reasoning engine” (and it is reasonable to assume other, similar terms) to describe an AI invention are likely to provoke greater inspection, as the EPO regards them as often acting as a smokescreen for products that are otherwise devoid of technical character; and
• Despite the difficulties, AI/ML technologies which provide a solution to a technical problem remain patentable.

In summary, the EPO’s guidance appears to suggest that patenting AI/ML technologies will be no different to patenting other algorithmic technologies.

Background

AI is a hot topic in almost every industry; from transport, to manufacturing, to even the legal profession where AI is already available to clients. It is a disruptive technology that will only become more prevalent as time goes on.

Both AI and ML rely upon the statistical interpretation of vast amounts of data. AI interprets the data from its environment to take an action which maximises its chances of success. ML is a subset of this field, and seeks to use data to progressively improve its performance of a given task (thereby “learning” how to complete it), without having been programmed how to do so.

On 30 May 2018, the EPO held its first conference discussing the challenges and opportunities around patenting AI. One of the actions the EPO took following this conference was to update its Guidelines for Examination to include guidance on the inspection of AI/ML patent applications.

The Guidance

The guidance directly addresses the focal issue of AI/ML technologies – that the processes which operate to produce an AI/ML product “are per se of an abstract mathematical nature” and thus are subject to the guidance on mathematical methods, irrespective of whether they can be “trained”. This means that the particular processes, algorithms and models which comprise an AI/ML system cannot be patented due to a lack of technical character.

However, as is the current position with mathematical methods such as algorithms and computer programs, an AI/ML product could be patented if it achieves a technical purpose. Specific examples given in the guidance included a heart-monitor which utilises a neural network to identify irregular heart patterns, and the use of AI to classify digital images, videos, audio or speech signals based upon low-level features.

A distinction was made regarding the use of AI/ML for the purposes of classifying data. The classification of documents through interpretation of their textual content was given as an example of a purpose that was linguistic, rather than technical. Additionally, the classification of data without an indication of a future technical application was another instance which fell below the required technical standard.

Jargon-lovers should take particular note of the guidance, as the EPO stated that the use of buzzwords such as “support vector machine”, “reasoning engine” or “neural network” would provoke careful inspection, due to a propensity for applicants to use such expressions to refer to “abstract models devoid of technical character”.

How to proceed?

Applications for AI/ML patents are only likely to increase over the coming years, so what does this new guidance mean for those who want to contribute to the robot uprising?

Firstly, the EPO reiterates the established practice, being the need to demonstrate technical function. Technology which provides a solution to a technical problem will remain patentable, and it is necessary for applicants to draw attention to this. For example, and as previously mentioned, classification of documents or data with no further technical purpose does not appear to reach the required level.

Secondly, applicants might study successful patent applications to understand how technical effect can be best demonstrated in AI/ML technologies to not only maximise their chances of getting a patent application through, but to shape their patent claims to capture different ways in which the inventions can be implemented. Inventions involving AI/ML are being intensively researched and there is significant potential for further, and as yet untapped, use cases as the platforms which can exploit those inventions – such as the Internet of Things – are evolving.

Finally, inventions reliant upon algorithms are generally challenging to enforce because of the difficulty in showing infringement, and many businesses therefore prefer protecting their algorithms through secrecy. However, there is ever-increasing discussion around the need to place a degree of regulation over algorithms that are being put to use. Calls for greater transparency regulation have been made due to the occurrence of discriminatory decision-making by AI selection systems. Further, on 29 October this year, a High-Level Expert Group for Artificial Intelligence was established to propose draft guidelines to the European Commission covering ethical topics such as transparency and fairness in AI/ML systems. Several leading scientific organisations and scholars in the US urged the White House to display greater openness in the development of the nation’s first artificial intelligence policy following a closed summit on Artificial Intelligence for American Industry in May this year, arguing that closed policymaking would result in critical privacy, accountability and fairness issues being ignored. Even public figures such as Elon Musk and Tim Berners-Lee have expressed the need for transparency in AI systems and the decision-making process of the technology.

Thus far, many businesses have chosen to protect their algorithms by keeping them secret. A push towards openness might be met with resistance without an intellectual property system which allows some degree of protection, thereby curbing further developments and advancement. It is not certain how or if the intellectual property rules can change to influence the balance between competing interests as developments and potential regulations unfold. However, it is certainly an area to maintain a close watch over.

The authors would like to thank Dan Harman, trainee solicitor, for his assistance in preparing this article.