ニュース&リソース

01.02.22

AI Inventorship – Can an artificial intelligence (AI) code be listed as an inventor on a patent application?

Background

Dr Stephen Thaler created an AI system called “DABUS” and subsequently filed patent applications at a number of patent offices, including the EPO, naming DABUS as an inventor.

Dr Thaler argued that he has created a system that has feelings and sentience, stating on his website:

“So that Thaler’s artificial inventors could appreciate their creations, he equipped them with learning rules to bind memories, contained within a series of nets, together to produce not only complex concepts, but also the consequences of said concepts, what psychologists would call affected responses…..In other words, feelings or sentience was the result”

Peer review of such findings has been sceptical or apathetic and hindered because the code behind DABUS has not been made public.

So how has the issue managed to get this far without any actual evidence that the AI system is capable of inventorship?

Well, patent offices do not assess inventorship. In the preliminary opinion of the EPO Boards of Appeal they stated that “the issue of how the invention was made….is outside the competence of the EPO”.  The issue being decided by the different patent offices and courts is purely the formal aspect of the whether a non-human can be named as an inventor.  The question of whether or not an AI system is actually the inventor or how that invention was made is not being assessed.

The cases so far

The legal systems of South Africa and Australia have found that DABUS can be named as an inventor, although the case is currently under Appeal in Australia.  The US and the UK have both refused to allow DABUS to be named as inventor.  In Germany it was ruled that a named inventor must be a natural person, but an AI system responsible for the underlying invention can be additionally named.

At the EPO, oral proceedings on the matter (J0008/20-3.1.01) were held recently where the previous decision of the Receiving Section was upheld, and the EPO decided that an AI system cannot be named as an inventor on patent applications.  The designated inventor on a European patent application must be a person with legal capacity.

Dr Thaler’s team submitted an auxiliary request arguing that a natural person had the right to a European patent by virtue of being the owner and creator of the AI system.  However, the EPO stated that the origin to the right to the European patent had to conform with the provisions of the EPC and that was not the case with the applications in question.  Dr Thaler’s team have already announced that they plan to continue their arguments in a divisional application.

The future of AI inventorship

Development in AI continues to attract interest with big players in the space making rapid advancements.  Google has perhaps the largest and most important AI investment strategy and their DeepMind company is continuing to navigate through various national patent schemes without concern for the issue of AI inventorship.  DeepMind’s AlphaFold uses neural networks to predict protein structures and is offered open-source under a Creative Commons Licence with the code available to be downloaded from the DeepMind website.  AlphaFold appears to be the subject of at least one patent application (WO 2021/110730) however the company takes the sensible approach of being more concerned with who has the rights to the IP than the AI inventor debate.

Other advances in the AI realm are being made by Amazon, Apple and IMB, though at the moment they are usually limited to a specific field of technology, see the Tempo technology developed by MIT for predicting breast cancer, and FedEx’s DoraSorter.

Tesla is also working on the “Tesla Bot”, a humanoid robot, since CEO Elon Musk has been slowly pushing the company towards becoming more of an AI/robotics company.  The Tesla Bot is “on track to become the most powerful AI development platform” according to Andrej Karpathy, Tesla’s Director of Artificial Intelligence.  However, the project is seen more as a way to attract AI and robotic talent to Tesla with little further details provided at the Tesla AI Day presentation last summer.

The traditional way of training an AI model to interpret something is to give it huge volumes of labelled examples, for example providing a picture of common fruit or vegetable with the different parts labelled or provide a spoken conversation with the words transcribed.  However, this process can be laborious and requires large manually created databases.  Therefore, new systems are now being created that are self-supervised, they involve models that can work from large quantities of unlabelled data, such as books and videos, and they can build their own structured understanding of the rules of that system.  For example, by reading thousands of books it will learn the position of words and grammatical structure without anyone first informing the system what each section of text means.

However, Meta (also known as Facebook) is working on an AI that can learn on its own from spoken, written or visual materials.  The research named data2vec was originally built on an AI framework but will start from scratch, so you can give it a book to read, images to scan and speech to sound out and, after a while, it will start to learn each of those things.  Early results show that data2vec outperformed similar sized models dedicated to a single modality.

It is also interesting to note that the code for data2vec is also opensource and is available via the Meta website.

With further development in the area the topic is attracting global interest and we can expect further cases and legal reforms to the existing patent laws to accommodate the innovations.  Check back with us for updates in the future.

If you have any questions on this topic, please contact Emma Bevan.

Share this page:
Back
In this section