A guide to AI and patents – 2024 update

Introduction

Much has changed since we first released this guide to artificial intelligence (AI) and machine learning (ML) back in 2019. At that time, ChatGPT did not exist, and ML remained mysterious to many in the world of science. In the last five years, the general understanding of ML has improved enormously, along with its adoption in many different aspects of technology. From a patent perspective, practice has also developed, although many of the fundamentals remain unchanged.

Today, AI/ML is used in fields as diverse as drug discovery, immunotherapy, automotive, finance, telecommunications and manufacturing – the list is too extensive to feasibly recite here. This crossover means that it is essential for advisors to upskill to ensure they can properly advise applicants. AI/ML is no longer confined to the fields of mathematics, statistics and computer science. Advisors must be able to identify the relevant issues so that they know when to seek specialist advice.

In this paper, we have tried to set out some useful guidance points for practitioners, regardless of their technical background. We hope it is relevant for both ML experts and complete novices – let’s call it a ‘primer’.

Let us begin by saying what we will not cover. This paper is not concerned with crystal ball gazing or policy considerations but is instead a pragmatic guide to the state of play in 2024. Ideas on non-human inventorship or giving personhood to AI are better suited to academic writing. Similarly, we will not touch on the ethics of AI or the philosophical role Intellectual Property can have in improving public trust in machines. We will also not go into matters closer to home, such as changes that will (or should) occur to the hurdle for determining inventiveness as a result of incorporating AI into the skilled team.

As AI becomes more widespread and invades the public consciousness, there is a mistaken supposition that systems are being developed to replicate the thinking of human specialists. This anthropocentric view of ‘intelligent’ systems is dangerous when considering AI patents. The systems of today are increasingly outperforming human experts, not by copying high-performing people but by exploiting the distinctive capabilities of new technologies, such as massive data storage capacity and brute-force processing. As we go through the current AI issues, we caution the reader to be wary of this kind of bias. AI inventions do not occupy a special category and can be considered on the same basis as others in the field of computer technology.

AI/ML ‘Primer’

First, what do we mean by AI? Generally, when people use the term ‘AI’, they really mean ML. We will predominantly use the term ‘ML’ here, but let’s start with some terminology.

Artificial intelligence

The modern definition of AI is ‘the study and design of intelligent agents’, where an intelligent agent is a system that perceives its environment and takes actions that maximise its chances of success.

Artificial general intelligence

AI is the intelligence of a machine that can successfully perform any intellectual task that a human being can. In effect, it is true intelligence and not data analysis. Most academics believe we are decades from this level of technology.

Machine learning

ML is a subset of AI. It is the programming of a digital computer to behave in a way that, if done by human beings or animals, would be described as involving the process of learning. ML uses statistical techniques to give computer systems the ability to ‘learn’ (e.g. progressively improve performance on a specific task) from data, without being explicitly programmed.

Typically, the process of ML involves building a mathematical model from a set of input training data and then applying that model to a set of test data to provide a prediction or output.

Often, the most important part of an ML development process is performing feature analysis. This is the selection of a subset of relevant features, such as variables or predictors, for use in model construction. For example, a programmer may choose which features of a set of cars should be given to an ML model to help the model identify which car will be fastest.

Deep learning

Deep learning is a subset of ML and is based on learning data representations as opposed to task-specific algorithms. With deep learning, it is generally not possible to identify ‘how it did it’. Since there is generally no standard definition of the term, we encourage readers to avoid its use where possible.

Large language models

Large language models are a category of ML models that process and generate text data in the form of human language. Large language models are typically trained on huge corpora of text data and use statistical techniques to identify patterns and relationships in language. Due to their flexibility and recent improvements in hardware and parallelisation, the popularity of large language models such as ChatGPT has increased dramatically in both popular culture and the patent space since our original 2019 guide.

Patentability

As far as patents are concerned, we can identify three categories of inventions: inventions on AI, inventions using AI as a tool and truly AI-created inventions. Most applications will fall into the second category as the use of ML becomes more prevalent. In most jurisdictions, the patentability of an invention in any one of these categories is considered similar to applying the general principles of patent eligibility, as the inventions are considered computer implemented (or mathematical methods).

Patentability in Europe

The fundamental principles of patenting AI inventions at the European Patent Office (EPO) have remained largely unchanged in recent years, with most inventions still falling within the three primary categories of AI patent application:

AI patents: Type 1 – ‘Core AI’

These are patents that relate to algorithms. The inventions involve the design of AI as a tool rather than its application to a particular problem and are rarely encountered by most patent practitioners. More commonly, applicants use an existing AI tool in a particular application; indeed, most currently used AI algorithms are based on academic papers, some of which are fairly old. Applicants need to demonstrate the invention’s technical character and show that it is more than a mathematical method.

These techniques are probably difficult to protect since they are unlikely to be tied to a particular technical advantage. Therefore, the EPO is likely to view this category of inventions as excluded from patentability on the grounds that they relate to no more than a mathematical method.

AI patents: Type 2 – Generating a training set or training a model

The EPO considers that the steps of generating a training set may contribute to the technical character of the invention if they support achieving a technical purpose. An objection of lack of technical character could arise if an examiner does not see a clear link between the training set and a technical result.

AI patents: Type 3 – AI as a tool

This type of patent is the use of AI in an applied field, defined by way of technical effects. This is the most likely avenue for success for applicants. In fields such as autonomous vehicles and healthcare, AI might be claimed as a tool for using a training set to provide a solution that yields a technical advantage.

2024 EPO Guidelines

Despite the assessment of the patentability of AI inventions remaining largely unchanged, the EPO updated their guidelines in March 2024 to provide applicants with further guidance regarding the level of detail required to ensure that the invention is sufficiently disclosed.

In the updated guidelines, the EPO provides a specific example of insufficiency in the AI/ML sphere, indicating that insufficiency may apply to AI/ML inventions having mathematical methods or training datasets that are not disclosed in sufficient detail to reproduce the technical effect over the whole range claimed.

The updated guidelines also clarify that the level of proof that an AI/ML invention has achieved a technical effect lies somewhere between a mere allegation of such an effect and a comprehensive proof of the effect. This proof may be readily apparent to the skilled person or disclosed in the form of explanations, mathematical proof and experimental data or the like.

Regarding training data, the updated guidelines further indicate that in cases where a technical effect is dependent on characteristics of the training data itself, those characteristics must be either disclosed or determinable by the skilled person without undue burden. The EPO does, however, state that this does not amount to a requirement to disclose a full training dataset, which in many cases would be disadvantageous to applicants.

Patentability in the UK

For a brief period, the November 2023 decision of the High Court regarding Emotional Perception appeared to have significantly changed the landscape in terms of the kind of AI inventions that are considered patentable in the UK in comparison with the EPO. However, the Court of Appeal overturned this decision in July 2024, thus returning the state of play largely to how it was before the High Court’s original decision.

In the context of Emotional Perception’s application, which involved the use of a trained artificial neural network (ANN) to provide media file recommendations to a user, various aspects of the statutory exclusions to computer programs and mathematical methods were considered. The key takeaways from the decision of the Court of Appeal are summarised below.

In the first important aspect of the decision, the judge arrived at a definition of a computer as ‘a machine which processes information’, consequently defining a computer program as a ‘set of instructions for a computer to do something’. On that basis, in the case of Emotional Perception, it was considered that the trained ANN did engage the exclusion to computer programs as such, the ANN being considered a computer and the trained weights and biases being a computer program. Consequently, many type 3 inventions that apply such trained networks will likely only overcome the exclusion of computer programs if their contribution is ‘technical’ in nature, similar to the EPO’s requirement of a technical effect.

The best guidance for determining whether a contribution is technical in nature remains the ‘five signposts’ shown below. These signposts were considered by the courts in the Emotional Perception and Halliburton decisions. In Halliburton, the modelling/simulation of a drill bit was considered patentable because it was for a technical purpose. Applying this approach to ML, it is likely that ML inventions that are deemed to involve a computer program will be patentable if it can be demonstrated they are for a technical purpose (similar to the treatment of type 3 inventions as identified by the EPO). Type 1 and type 2 inventions are less likely to be looked upon as favourably at the Intellectual Property Office of the United Kingdom (UKIPO) as they are at the EPO; to be considered patentable by examiners at the UKIPO, an ML invention will most likely have to result in a positive conclusion from at least one of the signposts:

(i). whether the claimed technical effect has a technical effect on a process which is carried on outside the computer

(ii). whether the claimed technical effect operates at the level of the architecture of the computer; that is to say whether the effect is produced irrespective of the data being processed or the applications being run

(iii). whether the claimed technical effect results in the computer being made to operate in a new way

(iv). whether the program makes the computer a better computer in the sense of running more efficiently and effectively

(v). whether the perceived problem is overcome by the claimed invention as opposed to merely being circumvented.

To conclude this section, it is worth noting that the UKIPO has issued updated guidance for examiners based on the Court of Appeal’s decision indicating that ANN-implemented inventions should be treated in the same way as any other computer-implemented invention. Therefore, any future ANN inventions will be examined by applying the well-established Aerotel test to identify whether they relate to excluded subject matter and using the five signposts when considering whether a computer program makes a technical contribution.

Identifying the invention

It is clear from the above that identifying patentable inventions will not be straightforward where inventions involve the application of ML techniques. The following section splits the method of implementing an ML process into six steps:

Step 1: Identify the problem you want to solve
Step 2: Decide which data you need
Step 3: Select a type of ML model
Step 4: Gather the data
Step 5: Train the model
Step 6: Use the trained model to make predictions

Step 1: Identify the problem you want to solve

To find a patentable invention, the applicant will need to show that they have identified a problem that is not obviously solved with ML. Since ML is versatile, the mere application of ML will likely be considered obvious, as will similar ‘problem-inventions’. Applicants may have to demonstrate there is a prejudice against using ML in the art for some reason. It will likely be challenging to obtain patent protection for this category of invention since the use of ML may be considered obvious for any problem that is capable of being analysed on a computer.

Step 2: Decide which data you need

Patentable inventions may result from the non-obvious selection of a particular parameter (’feature’) in the data used to train a model or the structuring of that data to achieve a technical result. In practice, it may be considered obvious to consult ‘domain experts’ to gather the data.

Step 3: Select a type of ML model

While patentability is possible here, it will likely be important to tie the use of a model to a particular technical process. One should be careful to provide an enabling disclosure and link any claims to the particular technical process.

Step 4: Gather the data

It may be possible to demonstrate an invention in data gathering if it involves a particular technical constraint. The requirements specification analogy used by the EPO could be useful here – that is, what tasks would a business person ask the engineer to perform? If there is a technical challenge being addressed to meet the requirements set by the business person, this may indicate the presence of a technical invention.

Step 5: Train the model

This is likely to be considered the application of a model and may well be determined to be routine if there are no unique aspects to the application of the model.

Step 6: Use the trained model to make predictions

Technical character may be conferred by this step of the process. There will be no invention for the use of a trained model for what it is intended, but the technical purpose of a method is important for determining whether there is an invention (for example, the use of model output to control technical process X).

Drafting considerations

Attempting to describe all the considerations that go into drafting the ‘perfect’ ML patent application will probably be the subject of a comprehensive textbook, updated monthly as practice changes and technology advances. However, we have listed some points to look out for (though this is by no means an exhaustive list):

Clarity

Poorly drafted patent applications are likely to fail for lacking clarity if they merely amount to a sprinkling of buzzwords related to the application of AI/ML. As an example, identifying the scope of a claim reciting ‘applying deep learning’ will be difficult and impractical – is a neural network implicitly essential to this claim, or is the use of classifiers excluded?

Location of infringement

As can be seen from the six steps listed above, an ML process can be categorised into discrete sub-processes. It is common for those sub-processes to be conducted in a distributed manner and potentially across borders. For example, data might be gathered by a smartphone in the UK and sent to a cloud server to apply a model to the data in Ireland using a model trained in the US. Therefore, multiple steps of a claim would be carried out in different jurisdictions on a claim to the whole process. Wherever possible, claims should be drafted to cover steps conducted by discrete entities, as well as methods and computer program products.

Technical effect/output/purpose

To infer technical character, it may be necessary to include a technical purpose in a claim, but this may unnecessarily limit an invention to a particular use even though it is potentially more widely applicable. Similarly, the specification will likely need to clearly demonstrate how the algorithm could be used for a technical benefit and how the output is used. Some applicants may be reticent about including this information, but practitioners will need to push back on this to ensure that they can draft a comprehensive description.

Enablement/plausibility

Continuing this theme, it may be important to detail aspects of the algorithm that some applicants would prefer to keep secret. In some circumstances, it may be enough to simply say ‘inputting the data to an artificial neural network’ if the skilled person would understand how to put the wider invention into practice. However, it is important to consider the level of detail disclosed carefully, as more detail is likely to be required, depending on the interplay between the ML algorithm and the application in question. Given the recent update to the EPO guidelines, it is important to consider the scope of the claims, where the technical effect lies in the invention and what is known to the skilled person at the time of filing to determine a suitable level of disclosure for both the ML method itself and the training dataset. The level of detail required will, of course, vary from case to case, but in general, reference to applying AI/ML methods without any specific details of the implementation should be avoided if possible.

Pseudo-code and/or detailed mathematics

Patent offices are increasingly approving the inclusion of pseudo-code and/or detailed mathematics in applications. There are significant pros and cons to this, and it is important to balance the disclosure requirements, the potential for limitation and the release of trade secrets when considering how much to include in an application.

Detectability

In filing strategies, one may want to consider carefully how possible it will be for third parties to detect the use of an invention. For example, if the invention is in the way the model is trained and only a model or coefficients are shared with third parties, will it be possible to reverse engineer how the model was trained?

Replicability

One may want to consider how the model is trained when preparing an application. Hypothetically, if an invention is trained once on a training dataset and then a set of coefficients or a model is applied repeatedly, where the invention lies in how the model is trained, is the value of the application affected?

Conclusion

Providing effective advice to applicants using ML is rapidly becoming an essential part of a patent attorney’s skill set, and we can see no evidence of this trend abating.

The realities of ML mean that patent protection cannot always be a silver bullet. While copyright, trade secrets and database rights are beyond the scope of this article, it is important to mention that these are fundamentally important assets to businesses using ML techniques. The EU trade secrets directive (implemented in the UK as The Trade Secrets [Enforcement, etc.] Regulations 2018) and copyright protection afforded to software implementations must not be overlooked. Maintaining proper version control, repositories, documentation and a trade secrets register is essential for businesses.

Hopefully, we have provided some useful tips and pointers to enable discussion. While the approaches of patent offices around the world are changing and settling, common threads run through them. We are always on hand to discuss any issues further.

We look forward to harmonisation efforts and a settled approach as the law is tested by applications over time and by applicants willing to push the boundaries.