Withers & Rogers studied 3,000 European patents, revealing that inventive step and clarity objections are the main stumbling blocks to patentability
As advances in generative AI continue to impact innovation and the patent ecosystem, an understanding of how patent applications directed to generative AI are examined before the European Patent Office, and the obstacles that exist, is becoming increasingly important.
In 2024, generative AI remains a leading topic of interest in the technology sector, largely due to the ability of such models to understand and reason from complex natural language inputs. The opportunities that come with this for innovators include reduced development time, more efficient processes, and the opening up of new areas of innovation.
Withers & Rogers and IP Quants researched more than 3,000 European patent applications focused on generative AI published between 2020 and mid-2023. We uncovered several interesting trends, which we have compared to an earlier study of machine learning patent applications in the EPO. We have tracked the objections that generative AI patent applications received for novelty, inventive step, sufficiency, and clarity.
Our study reveals that inventive step and clarity are likely to be the biggest obstacles for these innovations in the EPO. We have also highlighted the key EPO patent rulings that patent applicants of generative AI innovations must comply with if they are to attain a successful grant.
Generative AI at the EPO
Our research focused on the approach taken to examining machine learning patent applications at the EPO, identified trends in objections raised during the examination of more than 3,000 European patent applications published between 2018 and 2021.
When repeated with a focus on generative AI applications at the EPO, published between 2020 and mid-2023, the research revealed several interesting differences.
Figure 1: Machine learning and generative AI patent application trends in the EPO
First, the proportion of cases receiving a novelty objection under Article 54 of the European Patent Convention (EPC) increased for generative AI applications, no doubt owing to the rise in prior art within the field over the last few years.
Furthermore, the proportion of cases receiving inventive step objections under Article 56 EPC remained static. This should be encouraging for applicants; it shows that the EPO has not raised the bar for inventive step for patent applications in this area, although it does remain a potentially difficult hurdle to overcome.
The research also showed that the proportion of cases receiving a sufficiency objection under Article 83 EPC decreased for generative AI applications but remains higher than would typically be expected for computer-implemented inventions.
The proportion of cases receiving clarity objections under Article 84 EPC also increased significantly for generative AI applications. This is due to the state of flux surrounding terminology in this field and the way in which the use of generative AI is claimed. Particularly, attempts to define an invention that uses generative AI by referencing the model without an indication as to how the model is used – a so-called result to be achieved.
Whilst each of these data points reveal interesting aspects worthy of analysis and discussion, it is likely that inventive step and clarity will be the two main obstacles for generative AI applications in the coming years.
We now turn to critical rulings from the EPO’s highest authority on patents, the Enlarged Board of Appeal (EBA), that will impact generative AI applications as they move through the patenting process.
Inventive step
Recent decisions from the Board have provided some insight as to how the EPO is likely to consider inventiveness for generative AI based inventions.
Over the last 20 years or so, the EPO has developed a substantial body of law to determine whether an invention is deemed to be ‘technical’. This is often with reference to a requirement of the presence of a technical effect or a feature having technical character. A recent EBA decision, G 1/19, has confirmed this approach for all software-based inventions, including generative AI.
G 1/19, issued on 10 March 2021, related to a computer simulation method that simulated the movement of pedestrians through an environment. Whilst the method related to simulations, the decision is very much written with an eye to AI. G 1/19 confirmed the recent efforts of the EPO in stating that AI-based inventions should be deemed as a form of ‘super software’ and are no different to any other form of software. The EBA was at pains to point out that there is no basis to treat different types of software differently and that all existing case law regarding software patentability still stands. This is illustrated by our research which showed a consistency in the percentage of inventive step objections.
The G 1/19 decision also provided commentary around what qualifies as a technical effect. When determining the technical nature of a claim, the intended use of the invention, which can either be explicit or implicit, is decisive. The technical effect of a feature can also be implied or explicit. Therefore, if an integer of a claim covers both technical and non-technical uses (explicitly or implicitly), that integer can be considered non-technical.
This is becoming an increasingly important objection for both generative AI, and machine learning inventions more broadly. The EPO can and does point to passages in a description to say that an integer of a claim covers both technical and non-technical uses to attempt to limit the technical nature of the claim. In the case of generative AI applications, this can make obtaining a broader position more difficult, particularly if a description covers both technical and non-technical uses. In such cases, the use of a generative AI model does not impart technical character to a claim and instead the task of the model must be considered. Is the generative AI model being put to a technical use?
Another decision, G 2/21, has brought further developments in this field, despite also being in an unrelated area of technology (chemistry). In this case, the EPO allowed for the possibility to file evidence, that was published after the filing date of the application, to demonstrate the presence of a technical effect.
In the case of generative AI inventions, the technical advantage of an invention can be difficult to quantify or may not be accepted by an examiner. G 2/21, issued on 23 March 2023, makes it possible to submit evidence to demonstrate a technical effect – an advantage for generative AI inventions where further technical effects may occur beyond the use of the generative AI model. For example, an invention that uses generative AI to create labels to classify images might have a further technical effect relating to the later use of the labelled images. For instance, this could be an increase in accuracy in subsequent image classification. Should an examiner dispute the technical effect, it is possible to submit evidence to demonstrate the presence of such an effect.
Clarity
Article 84 EPC sets out the requirement that claims shall define the matter for which protection is sought and that they shall be clear and concise and supported by the description. Of particular importance to generative AI applications is that the definition of the matter for which protection is sought, as set out in the claims, must be as precise as the invention allows, and claims which attempt to define the invention by a result to be achieved are not allowed.
Consider the simplistic example of a method for categorising network traffic which comprises the step: “using a generative AI model to categorise a network packet as safe or harmful”. In this case, the “using” step merely repeats the problem that the claim intends to solve – categorising a network packet. It is unlikely that the recitation of the use of a generative AI model would avoid or overcome a result-to-be-achieved objection since the generic use of generative AI does not lift the claim beyond claiming the underlying technical problem. Indeed, the technical problem merely shifts to the ‘sub-problem’ of how to use the generative AI model to categorise a network packet.
In the analogy between a standard computing processor and a generative AI agent, a computer-implemented invention claim for solving problem ‘X’ would not be drafted with the step of using a computing device to solve X. Instead, the claim would typically include generalisations of the steps required to enable the computing device to solve problem X. When considering a claim directed to the use of generative AI to solve a sub-problem, it can therefore be better to break down the steps that enable the generative AI model to solve the sub-problem (just as a computer-implemented invention claim focuses on the steps required to enable the computer device to solve the problem). Typically, this involves claiming the structure and function of the input to the generative AI model, along with any specific environment configurations.
In the example given, this could be achieved by functionally claiming the input to the generative AI model. For example, including the step of “generating a prompt operable to cause the generative AI model to categorise a network packet as safe or harmful based on one or more examples of each category provided in the prompt”. This outlines a solution to the generative AI sub-problem by providing a prompt containing examples of each category. In providing this step, the subsequent step of “providing the prompt to the generative AI model to categorise the network packet as safe or harmful” no longer attempts to define the invention in terms of a result to be achieved.
Conclusion
Generative AI is a fast-moving field that promises a lasting impact on society and the patent profession. Whilst existing EPO guidelines and case law present several hurdles, applicants seeking patent protections for their innovations do not seem deterred.
However, generative AI remains a field in its infancy and questions persist over the relevance and application of patent law to this area. Indeed, whilst the above discussion is based on the current state of practice before the EPO, it remains to be seen if the guidelines and case law will evolve in this area over the coming years to provide further clarity and guidance to applicants.