Technology Trends – Why Patent Your Hidden AI?

Patent protection for AI technology is available around the world. Recent articles in this series have examined patentability of AI subject matter, and how best to prepare your AI patent application for success in Europe. However, many AI algorithms may be capable of generating huge value for their creators without ever being publicly disclosed. Faced with the upfront cost of securing patent protection, companies developing such “hidden” AI solutions may well be asking themselves whether seeking patent protection for this technology is a worthwhile investment.

As AI expands rapidly into all areas of life, national and international bodies are racing to try to keep abreast of the ethical, legal, and national security implications of this expansion, and this continuing evolution has a direct bearing on the value of patenting AI technology, hidden or not. In the sections below, Gemma Robin, Frances Wilding and Lisa Williams take a look at some of the most interesting developments, and their implications for the value of seeking patent protection for hidden AI.

Trustworthiness of AI systems is one of the key issues for the sector to address in driving public acceptance of AI. Some very public controversies in recent years over accountability and bias in AI algorithms have raised the profile of this issue, and brought the subject of Explainable AI to the centre of attention, as one part of the solution to providing trustworthy AI.

Explainable AI is the research field that seeks to unpack the “black box” nature of AI systems, providing understandable explanations for the decisions made by these systems. Research in this field is advancing at pace, and companies able to offer AI solutions that are demonstrably fair, and that can be proven to act in the way their creators intended, are likely to find that these qualities are of considerable advantage. Of course, if an AI algorithm is never publicly disclosed, then no matter how elegant and trustworthy the solution may be, its creators will be hard pressed to reap the benefits of this in terms of take-up of the technology and potential revenue streams outside of their core market.

As sensitivities over the use of AI continue to grow, transparency in how training data is collected, the contents of training data sets, and the fundamental workings of proprietary AI technology, are becoming serious questions for companies in the AI space. It is possible that disclosure of AI technology may in future become a qualifying requirement for deployment in a wide variety of industrial and commercial use cases. This requirement could in due course extend far beyond sectors of particular sensitivity, including for example legal and healthcare settings.

Rendering AI algorithms open source is one option, but protection in return for public disclosure of an invention is at the heart of the patent system. Patent protection thus offers a route towards the transparency that can support public confidence in AI and stimulate innovation, without sacrificing hard earned commercial advantage.

Detectability refers to the ease with which a patentee can determine whether or not a competitor may be infringing their patent. Detectability of AI models is another area of AI that is advancing rapidly. A few years ago, arguments could be made that there was no point in patenting AI as the patents would be unenforceable. This was because you simply couldn’t detect, for example, whether a particular method of training or data processing had been performed by a competitor.

Technological advances are rendering this type of argument increasingly invalid, as marked development in the field of machine intelligence is transforming the way in which AI is trained and used. For example, it is no longer the case that training is a one-time process that is hidden away or performed behind closed doors. Real-time and continuous online training methods are now widely used, and significantly increase the opportunities for detectability. It is interesting also that this move to online and real-time training has come about relatively recently. Over the 20 year lifetime of a patent, the way in which AI is trained and deployed will likely change again, and what is now impossible to detect may well become detectable.

The commercial drivers for deployment of AI across distributed systems, and for making AI solutions available as a service, are also providing increased opportunities for detectability, for example through reverse engineering training datasets, or even the underlying AI models themselves, based on increased access to AI model output. AI verification techniques are consequently attracting new interest, with water-marking already proving successful in verifying AI ownership, and new verification techniques, such as AI fingerprinting, under development.

In light of these advances, relying upon keeping AI solutions hidden may become a less attractive, or less viable, option in years to come. Indeed, in addition to the increased deployment opportunities, and the risk of unwanted disclosure, continuing developments in the field look likely to incentivise the use of patent protection for those who may otherwise have been inclined to rely on keeping their AI confidential.

Standardisation in the field of AI is in its early stages but is already the subject of extensive study on the part of national and international organisations. The work programme of the International Standards Organisation (ISO) AI working group (ISO/IEC JTC 1/SC 42) details the wide-ranging standardisation projects currently underway there. These projects range from functional safety to quality evaluation of AI systems, and from explainability to risk management, bias, and ethical concerns. In Europe, the European Commission has already adopted a package including a proposed legal framework for AI, the Artificial Intelligence Act, and the mapping of international standards to this framework is ongoing. National standardisation efforts are also well established in EU member states, as well as in the US, China and elsewhere.

In may be that in the not so distant future, being able to disclose your AI solution in order to demonstrate compatibility with the relevant standard will become a desirable, or even a necessary step on the road to deployment. Securing patent protection for your AI technology can ensure that this disclosure does not reduce your commercial advantage.

To have an idea of the potential impact of standardisation upon patenting strategy for AI inventions, we need look no further than the telecommunications sector. Telecoms is an area with a very mature system of international standards, and patentees in this sector have been adapting to the implications of international standards for many years. It is worth noting in this context that AI is considered as a foundational technology in many industry visions for 6G and beyond, and so is likely to start appearing in future telecoms standards in addition to standalone work for AI systems in general.

Patents covering technology included in telecoms standards are considered to be “standard essential patents” (SEPs), meaning that any company that is obliged or wishes to comply with the relevant standard necessarily infringes the patent, and is required to pay royalties. Standardisation is consequently a huge driver of patent value in the telecoms sector. In addition, in an ever more connected world, issues of SEP licensing for telecoms patents are becoming important to industries as diverse as automotive manufacturing, consumer goods manufacturing, and financial services. In view of the future prevalence of AI, it is not unreasonable to imagine that this pattern of expanding relevance of SEP licensing could be repeated with the introduction of standardisation in AI.

With work on the founding standards for AI technology continuing, the value of patents for AI innovation has the potential to increase dramatically in the coming years, both in establishing what you have invented when, and in providing a monopoly to exploit that invention.

AI patents, or AI patent applications, can offer particular advantage when collaborating with third parties. Increasingly, larger companies are teaming up with expert AI solution providers, and working together to apply expert AI knowledge to specific applications. In negotiations with these larger companies, the AI experts need to showcase their abilities, and to demonstrate what their AI products can already achieve, in preparation for close collaboration with the larger company to find a solution tailored to their needs. How is this possible without fear of the existing knowledge being misappropriated? Non-Disclosure Agreements (NDAs) are essential in this situation and can go some way towards protecting the information disclosed. However, there are many possible pitfalls in drafting and enforcing NDAs (often centred on what exactly is disclosed). Having a patent application on file adds safety in providing a record of previous AI innovation, and demonstrating what the AI expert is bringing to the table.

Since the term “artificial intelligence” was first coined in 1956, there has been a dramatic shift from government funded research in the early days of AI to the present situation that is heavily dominated by the private sector. In addition to the incredible technical innovation in the field, and vast advances in computing power and storage that have facilitated that innovation, the rise of venture capitalism has transformed the AI sector. VC has opened up entirely new avenues for funding, and resulted in intensified research and hugely diverse AI projects. The computing power now leveraged by huge cloud-based technologies, combined with the emergence of deep learning neural network algorithms and other techniques, are driving the latest renaissance in AI, fuelled in large part by VC investment.

In the wake of Microsoft’s $ 20 billion acquisition of Nuance – a leader in AI-powered speech to text technologies, with a serious approach to patenting – it would be reasonable to anticipate an even more active and competitive market for AI-powered startups. Investors have a fresh incentive to invest in startups with a patent-heavy AI focus.

Patent protection has also been shown to correlate with increased venture funding, and as set out in the joint 2021 report from the EPO and the EUIPO, “Intellectual property rights and firm performance in the European Union”, firms with IP generally see higher revenue. Although obtaining patents has associated costs that may be daunting to startups, the investment generally has a positive return, particularly when viewed against future fundraising.

Whether it is drawing a line in the sand to prove what you have created, establishing financial value associated with your AI innovation, or preparing for future requirements or opportunities, patents for AI inventions can add huge value to an enterprise, regardless of whether or not such inventions were ever destined to be publicly disclosed.

It is also worth remembering that patent protection lasts 20 years. Over the lifetime of a patent for which an application is filed today, the AI landscape, whether technological, legal, ethical, or regulatory, is likely to change beyond recognition. Securing patent protection for your AI innovations is a key step in preparing for the challenges and opportunities that this change will bring.