With the global technology race to create large scaled AI learning systems utilising Artificial Neural Networks (ANNs) and Reinforcement Learning, there has been a general reluctance of patent offices around the world to recognise and protect this significant new technology.
Recent developments from patent offices in the UK, Europe and the US indicate a restrictive trend amongst patent offices in the recognition in patent rights for ANN technology. Early indicators are that the Australia and New Zealand offices are likely to follow this trend.
While these patent office pronouncements produce a complex labyrinth of decisions, patent practitioners are left with the impression that, by careful crafting of their patent applications for ANNs to account for these pronouncements, they will suddenly turn their patent claims from patent ineligible to patent eligible.
While we discuss here the current patent office positions in the Untied Kingdom, Europe, United States, Australia and New Zealand, no doubt future divergent positions may evolve.
United Kingdom
In July 2024, the UK Court of Appeal in Comptroller-General of Patents, Designs and Trade Marks v Emotional Perception AI Ltd [2024] EWCA Civ 825, found that a patent application directed to an artificial neural network system utilising reinforcement learning came under the peculiar legislative exclusion of being for “a computer program as such” and was therefore excluded under the relevant UK legislation implementing the corresponding European position. The court’s reasoning sought to equate the utilisation of ANN network weights with the creation of a computer program.
Just six days after the decision, the UK patent office issued revised guidelines for examiners on examining patents for ANNs, requesting examiners apply the earlier standard software tests to all ANNs applications.
A major problem with the decision is its broad definition of computer as ‘a machine that processes information’. As argued below, this may be extended to encompass all electronic circuitry.
For example, most large scale digital chips implementing digital logic are designed via a Hardware Description Language (HDL). HDLs are very close to high level language software such as C code. Hence, HDLs may fall under the court’s definition of a computer in the processing of information and therefore arguably could be invalidated in line with the same reasoning. The numbers used in the HDL code could be argued as equivalent to ANN weights. Following on from the UK court’s reasoning, all digital hardware claims may now be argued patent ineligible.
Further, the UK decision extensively discusses that a neural network is designed to simulate the human brain, which works in an analog manner. Spare a thought for the Analog ANN equivalent (no, not the human brain!). The old analog computer design is being resurrected in the Analog Deep Learning arena due to its potentially producing a higher energy efficient ANN. As this is an equivalent ANN implementation but in the analog domain, it too must vulnerable. However, it is hard to distinguish an analog ANN from any other analog electric circuit, which is processing information in an analog manner (e.g. a simple electronic analog sensor). There is therefore a danger the UK decision might encompass analog ANNs and by extension, all analog circuitry!
The AI gold rush is also driven by scaled ANNs, which are implemented most efficiently by matrix multiplications running on GPU hardware for maximum speed and flexibility. The matrix values being equivalent to the weights of the ANNs. Hence, all large scale ANN patents are now likely vulnerable under this test.
Europe
The EPO position also appears to take a similar stance, with the Technical Boards of Appeal of the EPO taking a similar approach to ANNs (see decision T 702/20 Mitsubishi Electric Corporation 2022 – Sparsely connected neural network).
United States
The US position is not without particular complexities. In late July, the US Patent Office release an updated guidance for immediate use by US examiners: 2024 Guidance Update on Patent Subject Matter Eligibility, including on Artificial Intelligence. The guidance discusses the equivalent boundaries of ANNs for patentable subject matter under the equivalent US legislation (35 U.S.C. 101).
After a detailed discussion of the relevant applicable Alice/Mayo test, the guidance also attempts to apply its own guidance rules to three hypothetical examples in the ANN machine learning area. Each of the three hypothetical examples, however, include claims that the USPTO considers both eligible and ineligible in the same application.
This guidance would appear to advise patent applicants that, through careful drafting of claims, inventions can be turned from ineligible to eligible!
Australia & New Zealand
Whilst no high level decisions have been made in Australia on ANNs, these Patent Offices have shown a general reluctance to grant software patents. For example, in the recent Patent Office case of Apple Inc. [2024] APO 23, the Australian Patent Office rejected a software patent application (whose parent had already been accepted) through the application of its own previous patent office tests and a selective interpretation of court precedents.
When dealing with the Australian Patent Office, it would again be wise to carefully draft any ANN specification and claims to ensure the application steers away from patent ineligible claims, and towards patent eligible claims.
In New Zealand, the patent office closely follows the UK precedents and legislation and, therefore, the Emotional Perception decision is likely to be highly influential, necessitating the same levels of claim scrutiny.
Conclusion
Overall, the signals are that ANN patent applications are in for some rough weather. The guidance issued by various patent offices need careful examination by patent drafters to ensure applications are steered towards eligible claims.
One bright note is that, hopefully, the new Large Language Models such as Chat-GPT can provide clear direction to patent drafters trying to avoid patent ineligible claims! Just make sure you don’t use them with client materials.
And no, this article was not written with the assistance of the ANN trained Chat-GPT.
How we can assist
Given the global nature and speed with which ANNs and AI systems in general are evolving, if you are developing a potentially patentable application, ensure you engage an informed adviser to assist.
Spruson & Ferguson has some of the most highly qualified and engaged ICT attorneys in the region. Don’t hesitate to reach out to our team.