Artificial intelligence (AI) has been around for many decades; the term itself was coined in 1955. Yet in the last 10‑15 years there’s been an unprecedented surge in AI-based technologies. These rapid developments have prompted the crucial question, “What are the most effective strategies to protect your AI IP?”
The increasing popularity of AI as a research topic is reflected in patent filings. According to the World Intellectual Property Office (WIPO), by 2019 there had been over 1.6 million AI-related scientific publications and nearly 340,000 AI‑related patent applications. Similarly, the United States Patent and Trademark Office (USPTO) has seen filings relating to AI technology rise from around 9% in 2002 to around 16% in 2018.
IP law, particularly patent law, has had to evolve alongside these changes. Patent law on software and mathematical methods has long been contentious, with the boundaries of the patentable subject matter being tested and probed year after year. Often different jurisdictions have diverging opinions, sometimes subtle, but often more pronounced. While there’s a much-peddled myth that software-based inventions cannot be patented, it’s true they tend to be more challenged during examination than inventions based on more ‘traditional’ technology.
Whilst patents are often regarded as the ‘gold standard’ of IP rights, they’re not the only form of IP protection for AI-based innovations. A diverse IP portfolio underpinned by a well-thought-out strategy may be more effective than one focused on a single type of IP right.
Patents are used to protect technical inventions that are both novel (new) and inventive (not obvious). They provide a 20-year monopoly in exchange for disclosing how to carry out the invention. The patentable aspects of AI-related innovations vary by country and are constantly evolving.
The various patent offices may use different tests to determine whether or not the claimed invention is excluded from patentability, but generally, if the claimed invention provides a technical solution to a technical problem, it may well be patentable.
The EPO’s Guidelines for Examination already include guidance for AI claims. At the time of writing, we’re waiting for updated guidelines from the UKIPO for AI inventions. However, the UKIPO is largely expected to follow the EPO, in differentiating between inventions in the ‘core AI’ technology category versus inventions in the category of ‘AI applications’.
Under the EPO’s approach, protecting ‘core AI’ technology (such as a particular neural network) may be difficult, whilst an AI-based invention with a specific technical purpose may be patentable. It’s not clear what the EPO deems ‘technical’, even for relatively similar problems. For example, applications that classify text using natural language processing (NLP) have been deemed non‑technical as they apparently relate to a linguistic problem rather than a technical one, whereas applications for classifying images or video (e.g. by using pixel information and edges) have been considered as technical.
This seems strange, given that both types of applications can use very similar algorithms; the approach seems to focus on what the applications are being used for, rather than the technology. However, it does reflect current practice at the patent office.
It’s worth bearing in mind that a patent application will publish after 18 months (unless it’s withdrawn well before the publication deadline) regardless of whether a patent is granted. You’ll therefore need to be careful when deciding what should or should not form part of your patent application. It would be wise to consult a patent attorney with specialist expertise in AI, to determine the best approach for protecting your technical innovations.
2 Trade Secrets
Innovations can also be protected as trade secrets. Although not an ‘IP Right’ in the same way a patent or registered design might be, this is nevertheless an important part of a diverse IP portfolio. Trade secrets do not grant you a monopoly right in the same way as a patent, but they do provide protection against the disclosure of sensitive confidential information.
It could be worth considering whether you seek patent protection for some aspects of your AI solution, while other aspects are retained as trade secrets (and therefore omitted from the patent application). However, you need to be careful when doing this. One of the criteria for patent protection is the sufficiency of disclosure, i.e. the inventor must provide enough technical details for the skilled reader of the patent to carry out the invention. It’s also quite common when your patents are being examined, that you need to amend the claims, e.g. to narrow their scope to specific features described in the application. If those particular details are missing from the application, they cannot be used as the basis for amendments.
For example, your patent application might mention the use of a particular artificial neural network arrangement for a particular technical task. There may be certain related hyperparameters that you consider trade secrets and therefore haven’t disclosed in the patent application. You’d need to ensure your patent application still met all-sufficiency and inventive step requirements, and be aware that you won’t be able to specifically claim any of those specific hyperparameters later, for example, to overcome prior art. If those hyperparameters might be routinely chosen by others working in the same area and are not key to the inventive aspect, it might not be necessary to include them. However, if that particular set of hyperparameters leads to a significant benefit, then their configuration may be considered intrinsic to the invention. This is why making use of patents and trade secrets in this way is a delicate balancing act.
Another potential problem with taking the trade secret approach is transparency. In some areas, it’s important for trust and/or confidentiality purposes that certain details around the AI application and its algorithm(s) are available for vetting purposes. This might conflict with omitting details to protect your trade secret.
Copyright protects creative works. In the context of AI-based innovations, the code-based implementation is protected automatically via copyright (including both the source code and object code versions of the software). Copyright is free and generally does not require registration (though some countries, such as the USA, require registration before litigation can commence). However, copyright is a relatively narrow right. It prevents others from directly copying your copyrighted work, but does not give you a monopoly – so it cannot be used to stop a competitor from independently creating their own AI software that functions in the same way.
It’s important not to dismiss the other rights simply because you automatically have copyright. In many cases, once a competitor has seen the product and how it functions, they could easily make a rival product that functions in the same way. If your product has an advantageous technical function, you should at least explore protecting that aspect via a patent. Nevertheless, copyright is still a valuable asset for AI solutions.
4 Registered and Unregistered Designs
Registered and unregistered designs seek to protect the appearance of a product. Registered designs (sometimes referred to as ‘design patents’, e.g. in the US) provide a 25-year monopoly right, while unregistered designs (sometimes confusingly referred to as ‘design right’) are more like copyright, in that they only protect against direct copying.
Although designs are less applicable to AI innovations themselves, they are worth bearing in mind when considering how to protect your end product. Whilst it’s easier to make the case for protecting a hardware product that embodies AI software, it’s also generally possible to protect graphical user interfaces (GUIs) via designs, so these types of rights can also apply to products that exist only in software form.
5 Database Rights
Another type of IP right not often discussed is ‘database rights’, which you might consider as a subset of copyright. If you’ve invested significant time and effort collecting or collating data, then your database may attract IP rights.
This could clearly apply to data sets used in AI systems, for example, training data sets and verification data sets used for a machine learning algorithm. Rights over such databases may be complex, and we recommend seeking specific legal advice.
6 Trade Marks
Finally, you should consider trademark protection for the brand which covers your AI system. This could be the name and/or logo of an AI product, software-as-a-service (SaaS), a smartphone app, etc. As a general rule, when choosing classes for the product, make sure your applications not only cover your current goods and services but also those you have plans for in the near future.
An AI algorithm designed for a particular use may also be used in other areas, so you should consider protecting the mark for those uses too, or using a different mark for those other applications.
For example, an algorithm that improves image classification may have initially been designed for medical diagnostics. However, that same algorithm might be adapted for use in autonomous vehicles to detect road signs. If it’s likely you might explore this and other uses in the next three to five years, it’s worth considering filing the mark in the relevant classes. There may be advantages in building a brand around expertise in image classification generally, rather than in one specific area.
AI-related IP is likely to continue its rapid growth and IP laws will need to keep evolving to keep up with the ‘fourth industrial revolution.
It’s unlikely any one specific right will exclusively protect your AI IP, but using a combination of these different types of rights will help you create a stronger IP portfolio that protects your valuable innovations.
For more details on how we can help protect your AI IP, talk to one of our Intellectual Property team.