Examining the intersection of AI technology with patent eligible subject matter is a challenge for patent offices around the world. Many offices, but in particular the EPO and USPTO, have well established processes for examining inventions, including those related to AI, which may encompass a mix of eligible and excluded features. Regardless of the underlying legislation they are intended to enforce, these processes generally seek to distinguish that which is anchored in the technical, and consequently patentable, from that which is abstract, purely theoretical, mathematical, or business related, and consequently excluded from patent protection in many jurisdictions.
Those of us practicing in this field are well used to crafting claims and arguments that try to characterize and present the invention in front of us in a way that maximises our chances of navigating these examination processes successfully, and so obtaining commercially valuable protection for the invention. However sometimes it simply isn’t possible, particularly for advances in the field of Machine Learning itself, to frame an AI invention in such a manner as to achieve a useful claim scope at the level of generality the invention deserves. After wrestling with patent eligibility for so many challenging and brilliant inventions in this field, we would argue that there are commercial, scientific, and ethical arguments for broadening access to patent protection for AI inventions by recognising AI as a technical field in its own right, and thus acknowledging innovations in this field to be inherently eligible for patent protection.
Keeping the patent system fit for purpose – a commercial case for patenting AI
For the purposes of patent protection, AI inventions are often categorized as being Applied AI or Core AI. Applied AI is the application of an ML model to a specific technical problem, while Core AI refers to innovation in the field of machine learning itself. Core AI innovations may have applicability across a wide range of technical and industrial fields, and may in themselves make the difference between an AI solution to a problem being viable or not. However, Core AI inventions are typically the most susceptible to being excluded from patent protection, considered in Europe as a mere mathematical method or computer program, or in the US as an abstract idea.
The very nature of AI, and its acceptability and application in our world, raise huge challenges that will only be solved by innovation in the nature of machine learning algorithms themselves, how they work, and how we understand them. It would seem that excluding such advances from patent protection not only risks penalising actors investing in this area, but disincentivises such innovation.
AI is built on data, and there is currently a huge amount of research around the explainability and trustworthiness of AI models, seeking for example to address the very real challenges of bias when training an AI model on data collected from an imperfect world. AI’s reliance on data also raises challenges related to the availability of that data. Consider for example the use of AI in management of catastrophic failure events in industry, transport, or other sectors. Thankfully, there is relatively little data from such events on which to train a model, so how can we leverage AI innovation to nonetheless take advantage of all that AI has to offer to predict and mitigate such events? In a totally different field, how can we use AI to look for solutions to medical challenges where data is limited? This question is relevant for example when seeking to ensure existing AI models are fit to serve populations who may be underrepresented in medical data, or when seeking solutions for rare diseases or genetic disorders, for which data is by definition scarce. The answer to these challenges may well lie in innovative pre-processing of the limited data available, generation of synthetic data, adaptation of existing models or training methods to achieve greater accuracy with smaller training data sets, or in some other approach entirely. All such innovations are likely to be in the field of machine learning itself, with broad applicability and huge commercial relevance, and yet they may well face significant hurdles in securing patent protection.
Moving on from the challenges associated with data, AI is computer implemented, meaning the fundamental resources for its operation are frequently processing power, memory, bandwidth for the transfer of data, and time to complete the processes underlying the method. Underpinning all of these is the energy to power the hardware on which any computing system resides. While the growing implementation of AI models in IoT and other constrained devices raises particular difficulties, finding ways to run AI models faster, and more economically and ecologically, is a key challenge for the future of all AI systems. Innovation that enables useable results while reducing demands on time, processing, memory, or communication resources is of huge commercial importance, and again has potential relevance across the vast range of fields of application of AI technology. However, without limitation to a specific hardware or field of application, such innovation faces an uphill struggle to establish that it is eligible for patent protection.
Faced with the prospect of patent protection that may be limited to a specific use case, and consequently of limited commercial value, applicants may consider bypassing the patent system entirely, and seeking to rely on keeping their core innovations secret, or moving to an Open-Source model. While Open-Source is popular with the developer community, major players such as OpenAI appear to be moving away from the Open-Source model, and incoming regulation, notably in the form of the EU’s AI Act, is likely to limit the options for companies in this space looking to keep the details of their AI systems secret.
Potential applicants in this area may therefore be faced with the unhappy choice between patent protection that is highly limited, and protects only a fraction of the commercial scope of their innovation, or no protection at all. Broadening access to patent protection for AI innovation would incentivise AI research, and ensure the patent system remains fit for purpose to serve innovators in this field.
More than the sum of its parts – a scientific case for AI
As European Patent Attorneys, our approach to this question is inevitably somewhat skewed to the relevant law and practice around excluded subject matter under the EPC (mathematical methods and computer programs “as such”, and the “two dimensions” for technicality). However, we suggest that many of the considerations offered below are relevant for other jurisdictions.
You only have a to take a brief look into what is meant by “AI” to start coming across statements along the lines of “AI is just maths”. The assertion may be that AI is very complicated maths, or alternatively that it is simple maths but executed at huge scale. However, while a mathematical understanding may be fundamental to working with and designing AI systems, to say that AI systems themselves are nothing more than maths seems highly reductive when contemplating the extraordinary capabilities of these systems. AI functionality may be describable using mathematical concepts and language, but so are many scientific, engineering, and other processes which we have no problem at all in understanding as technical. Just because an AI system may operate entirely in the digital domain, does that mean it is not technical? Can a purely mathematical method really be considered capable of evolving, learning, reasoning, predicting, or taking account of context and uncertainty as ML models are capable of doing? Reinforcement Learning algorithms have been balancing the relative merits of exploration and exploitation of a state-action space for many years, but are these concepts truly compatible with a process that is nothing more than a mathematical method?
We would suggest that, for the purposes of the patents system, the mathematical functionality underpinning AI systems should be viewed as building blocks that can represent individual components of the underlying functionality of a model. However, arranging these building blocks in such a manner as to achieve a desired result, a result that may encompass logic, critical thinking, uncertainty, reasoning and qualitative knowledge, is a process that requires inventive skill, and results in a creation that goes far beyond what could be considered as a mathematical method “as such”.
A proper exploration of the relation between mathematics and AI is a question well beyond the scope of this article, but the characteristics of AI systems support the idea that AI, through its capacity for evolution and reasoning, in fact transcends the mathematical models and computer programs through which it is described and implemented. We would therefore argue that AI in and of itself embodies a further technical effect when compared either to a computer program or mathematical method as such, and that AI is, or should be, by its very nature a technical field, and inventions within this field should be eligible for patent protection.
Incentivising disclosure – an ethical case for patenting AI
As Generative AI models dominate headlines, and discussions over the safety and legality of these hugely impressive machines continue, it is interesting to see safety cited as a reason both for the disclosure and the withholding of details of how these machines actually work. OpenAI’s GPT-4 is a good example. Many in the AI community are calling for detail about the design, implementation and training of GPT-4, so as to evaluate and mitigate potential harms and risk, as well as asses the human and environmental costs of GPT-4. OpenAI however, in addition to highlighting the need to maintain a competitive advantage, have suggested that the potential harms that could be caused by models like GPT-4 mean these models may be simply too powerful to risk putting them into the public domain. Ilya Sutskever, OpenAI’s chief scientist and co-founder, has gone so far as to say “I fully expect that in a few years it’s going to be completely obvious to everyone that open-sourcing AI is just not wise”.
Safely managing the incredible advances taking place in AI technology is a concern for all of us. While international regulation and standardisation will be necessary, the controlled disclosure, and balance of public information with commercial protection, that are offered by the patent system have the potential to play an important role in mitigating some of the safety concerns around AI technology. In this context, widening access to patent protection for AI innovation would be desirable.
Final thoughts
This short article can only scratch the surface of the case for a more inclusive approach to patent protection for AI inventions. We are under no illusions as to the challenge that such a shift in approach to patenting of AI inventions would represent. However, it remains our opinion that the coming ubiquity of AI innovation, its vital importance to the global economy, and its extraordinary potential to shape our future, all place AI in a category of its own in terms of scientific innovation. For this reason, the question of a more radical shift in our approach to the patentability of AI inventions merits serious consideration.
This is for general information only and does not constitute legal advice. Should you require advice on this or any other topic then please contact hlk@hlk-ip.com or your usual HLK advisor.