Protecting artificial intelligence (AI) solutions by intellectual property rights (IPs) is often a combination of patents, agreements regarding data and use and development of systems and software as well as copyright to software. At the time being, especially patenting practices for AI solutions appear to be facing significant changes in Europe.
Inventive Step and Person Skilled in the Art
An essential requirement for patentability is an inventive step, which aims to ensure that a novel technical solution is patented only when it differs enough from the prior art meaning, for example, previously known products and methods. The inventive step according to the European practice is assessed by determining the difference between the invention in question and the prior art and then evaluating how significant the difference is. In general, the difference is significant enough if it is not obvious for a person skilled in the art. In addition to the inventive step, the description of the invention provided in the patent applications must provide enough details that would enable the skilled person to implement the invention on the basis of the application.
Therefore, the person skilled in the art is a key player when assessing both the inventive step and sufficiency of the provided description for the invention. According to the European practice, the skilled person, if simplified a little, is a fictional character who knows everything in the relevant area of technology but is not very clever in applying his knowledge to find new ways of solving tasks. The person skilled in the art is, however, capable of the routine type of work and can carry out experiments.
What is changing in patenting AI?
In the European patenting practice decisions of the European Patent Office (EPO) play a significant role and set directions. Decision T 0161/18 by the Boards of Appeal (BoA) of the EPO stated that the skilled person would not be able to train a neural network needed for the solution described in the patent application since the content of the application was insufficient. The BoA considered that the skilled person would need more information than was disclosed in the application to train the neural network. The decision is, naturally, very unpleasant for the applicant as the application was filed already back in 2005 and the missing information can’t be added anymore.
The EPO’s long-standing practice for inventions has been to assess the inventive step of a presented solution by referring to what is or is not “obvious to the person skilled in the art”. If the presented solution is obvious to the skilled person, it means, in practice, that the Office considers the skilled person having knowledge and skills to end with a solution following the invention described in the application. Examination guidelines of the EPO provide a general guidance on the level of knowledge and skills of the skilled person but no detailed definition on different fields of technology exists.
The above-mentioned decision by the BoA concerns the training of an AI solution, which is essentially a software-based invention. Therefore, it may be concluded, that based on the level of capabilities of the skilled person not being sufficient for training the neural network, AI is understood by the EPO as a field of software technology in which the skilled person suddenly does not have the general level of skills and knowledge that an average software engineer otherwise has. Thus, the skilled person in the AI appears not to be capable to design a data set that could be used to train an AI solution. This suggests that there could be significant variation in the knowledge and skills of the skilled person between different fields of software technology.
On the other hand, the decision states that a data set used for training the AI solution could have enabled a sufficient disclosure of the invention. Therefore, the description of the training method and the data set could be used to define an invention in a manner contributing to the inventive step.
It seems that the “skilled person” in the field of AI is not particularly highly skilled compared to what could be expected from a skilled person in the field of software.
In the future, adding a description of the training data set for implementing the invention should be kept in mind when drafting patent applications regarding AI solutions. There’s also a positive message to patent applicants. It seems that the skilled person in the field of AI is not particularly highly skilled compared to what could be expected from a skilled person in the field of software. Could this mean that obtaining a patent for an AI solution is possibly easier than for other software-based inventions? We don’t know that and estimating the impacts of the recent changes in the long term is difficult. New decisions by the EPO will hopefully clarify the situation but it can be asked if the direction set by this decision is the right one. Can different requirements set for patenting AI solutions in comparison to other software-based solutions be justified? How can we prepare and react to changing requirements?
Planning IP protection, having a strategy, is always important and in case of AI solutions we should aim to be prepared to changing patenting practices and differences between different jurisdictions.
Careful drafting of a patent application helps to anticipate and manage changes in the patenting practices. When patenting AI solutions in the future, it should be ensured that the application contains a description and an example of the training data set to enable sufficient disclosure of the invention.
No major changes can effectively be made to already pending patent applications. Instead, it must be evaluated if the description provided is sufficient and if any changes can be and should be made to obtain a patent.
Also, when protecting AI solutions, other forms of IP protection and agreement matters should be considered. Take care of copyrights to software, trade secrets, and rights to the data sets both in agreements and in your company practices. For example, when a training data set is described in a patent application, the description becomes public at the time of publication of the application. Thus, keeping in mind possible trade secrets regarding training data sets is essential. They are not necessarily a problem if presented suitably. Seeking advice from patent professionals will help you forward.
We are taking part of the SHIFT Business Festival 2021 and the topic of this blog relates to one of the SHIFT 2021 program themes, AI & Digital Trust. Join the festival in August and dive deeper into the theme.
SHIFT Business Festival
25-26 August 2021