In the coming decades, machine learning is likely to be the primary driving force behind a Cambrian explosion of applications in robotics and software automation.

Author of ‘The Rise of the Robots’, in an article in the Financial Times in 2016

“The near-term future is likely to be transformed not by general purpose robots or AI systems but rather a nearly limitless number of specialised applications,” Ford added. Inventors and patent practitioners are now grappling with how IP rights can protect those applications as well as developments in fundamental machine learning techniques – and the challenges they pose for the way IP systems work.

In May this year, more than 350 people attended an EPO conference on patenting AI, and in the same month WIPO hosted a meeting with IP offices on ICT strategies and AI for IP administration. WIPO Director General Francis Gurry discussed the topic further in an interview in the WIPO Magazine, published in September. It is clearly a subject that is being discussed at the highest levels.

This, the first of a series of articles looking at AI and IP, focuses on using patents and other IP rights to protect AI-related inventions and machine learning. Future articles will look at the role of AI in trade marks and in the IP workflow.

Protecting AI innovation: patents or trade secrets?

Patenting of artificial intelligence technology is a relatively recent phenomenon, but is growing rapidly – applications increased by an average of 43% a year from 2011 to 2016 according to the EPO’s recent report, Patents and the Fourth Industrial Revolution. And AI technologies can be applied in many different industries, from life sciences to retailing and transport to entertainment.

But AI technology also poses challenges to patent applicants and the patent system generally. These include: When is an invention patentable? What types of invention are excluded from patentability? How can you detect whether a particular AI technology is used by competitors? And what prior art is relevant?

Many businesses therefore also believe there is potential to use trade secrets laws to protect AI inventions: AI is hard or impossible to reverse engineer, in particular if implemented in the cloud, and provided secrecy is maintained, AI inventions should be good candidates to be kept as trade secrets in many cases.

Industry views: sensors, robotics & drug discovery

We spoke to representatives of two companies who are active in R&D about the trends, challenges, strategies and future of patents for AI technologies.

Gareth Jones, vice president, intellectual property for Benevolent AI, which develops and applies AI to scientific innovation, says: “AI innovation and therefore AI patent filing volumes are rising at a rapid pace. The early leaders are clear in terms of quantities, but where the valuable assets lie is probably too early to determine.”

He adds that selecting whether to use patents or trade secrets to protect AI innovations involves similar considerations as for other software innovations. “Trade secrets are useful for inventions that can realistically be kept confidential for a long period of time, including factors such as detectability, staff turnover, cybersecurity, external collaboration, etc,” says Gareth.

Chris Nalevanko, associate general counsel and head of IP strategy at Zoox, a company in the San Francisco bay area who is using unique AI, robotics, and product design to develop an autonomous mobility service, adds: “As a small company, you want to smartly grow your patent portfolio and take into account when something should be protected as a trade secret instead. But trade secrets might have less usefulness in some scenarios: if you move on from a technical implementation, you might not care about it as far as trade secrets are concerned. But if you patent it, it might have defensive value even if it’s not something you do yourself anymore.”

“For patents, it’s great that you have the 20-year exclusivity period. But by its nature you’re giving some things away and telling the public what you’re doing. Where a lot of what you’re doing is secretive, that’s something you need to consider. You need to consider if the claims that issue are something you could find out if somebody else is doing,” says Chris, who notes that patenting has other advantages too: “One thing to be cognisant of is AI is a very academic field and you have a lot of researchers and engineers who like to publish. Patenting in this space can be a recruiting tool to let them know publishing in some form is something you do allow and that you are working on innovative problems.”

Need for a concrete application or implementation

When it comes to filing patent applications, there is some uncertainty over what constitutes eligible subject matter in certain jurisdictions. However, the EPO applies a fairly consistent test, and provided the invention has a real-world application then it should not be excluded. It follows that AI applications where AI is specifically adapted for the application should be patentable.

Chris says similar criteria apply in the US: “Algorithms or structures that you will be using in your AI may be known, but what you do with the output, how you use them, and how you combine the output with other technical aspects, for example, may be the interesting thing. So in our space, for example, we would focus on how we combine the output of the AI with other technical aspects and software or hardware developments in autonomous driving or vehicles.” From a practical point of view, he adds: “We try to focus the claims on a concrete application or implementation, for example autonomous driving or vehicles. It can help the examiner have a concrete view of what’s going to help get over some of those Section 101 issues.” The recent guidance published by the USPTO following the Berkheimer case has been helpful in this respect.

However, given the transformational nature of AI and to encourage investment in and dissemination of fundamental application independent developments in AI, it would be desirable for AI and machine learning to be recognised as technology in its own right. Patent offices, including the EPO, recognise mathematical inventions in other fields of seemingly abstract technology, such as cryptography, without a second thought and have done so for a long time. There is no reason why AI and machine learning should be treated differently. Cryptography enables computers to do things more securely, while AI and machine learning enable computers to do more things previously not possible. The patent system should not discriminate against fundamental AI and machine learning technology and hopefully the patent offices will recognise this in light of the growing economic and societal importance of these fields.

Hurdles to overcome in AI patenting

Another challenge for patenting AI and machine learning is an abundance of prior art. Of course, as Gareth says that is a challenge with any software invention, and not specific to AI. However, work in AI and machine learning has for a long time been a mostly academic endeavour until the advent of sufficiently powerful computing technology, in particular the use of graphical processing units for the matrix manipulations underlying many AI and machine learning techniques. This means that by the time a commercial case for patenting became clear, many of the fundamental techniques had been published in academic journals.

While the regulatory environment and abundance of prior art represent hurdles to patenting AI innovation, the true test for AI patents will be seen when they are enforced. Given the complexity of the technology, the fact that the output of AI and machine learning systems usually contains very little information on the computations used to produce it and in particular in the context of processing that can happen safely out of scrutiny in the cloud. As Gareth says: “Detecting infringement may be difficult for many AI inventions, and therefore the ability to enforce will be one of the big challenges for AI patents.” Chris agrees: “You will see some more evidence of detectability and scope when AI patents start to get litigated.”

Will the patent system need to adapt?

One of the decisions AI innovators face is: what to patent? The rapid advances in the technology provide lots of opportunities for filing patents, but in some cases other rights such as trade secrets may be more relevant. Gareth says it is key to think about “protecting the right inventions”. He adds: “Protect inventions that provide the best leverage: those that will be widely used, easy to detect, hard to design around, and truly novel and inventive.” Chris echoes this: “There is a lot of misnomer on the street about what AI is. When you dig into the details of deep learning or neural networks, it brings it back to more traditional notions of software programming and computer science. It’s a bit more concrete than just asking: is AI patentable?”

One question that AI poses that is probably different to other technologies is whether it represents a more fundamental challenge to established concepts in patent law and the way the patent system operates. In his interview in the WIPO Magazine mentioned above, Francis Gurry said “it is clear that AI will have an impact on traditional IP concepts” such as “composer”, “author” and “inventor”. He added: “For example, the life sciences generate enormous quantities of data that have significant value but don’t constitute an invention in the classical sense. So we need to work out the rights and obligations that attach to them.”

Gareth of Benevolent AI agrees that IP systems may have to adapt: “Patent systems need to determine the likely impact of non-human inventors, particularly the subsequent changing definition of a skilled person, the economic and societal impact of such changes, and the implementation changes required to ensure continued fulfilment of the intended purposes of patent systems.” We will look at some of these challenges in a subsequent article in this series, to be published next month.

This article was co-written by James Nurton