We are in the midst of an AI boom, with investment and merger and acquisition activity in the sector increasing exponentially. This new frontier raises various challenges for IP law, where numerous questions exist about whether the existing legal framework is fit for purpose in the age of the intelligent machine.
IP Protection for AI Systems
“Artificial intelligence” is generally used to refer to technology that carries out tasks that normally need human intelligence. Here, we focus on machine learning, a subset of AI that enables computers to learn from data without being explicitly programmed. A machine learning system typically comprises a computational model based on an algorithm (or algorithm stack) with a dataset to train it.
Patents provide monopoly protection for inventions. The main barrier to patenting machine learning, and other AI systems, is that, in most countries, including the U.K. and United States, abstract mathematical methods are not patentable.
While the position varies from country to country, patents are typically available only if the invention is an application of the mathematical method with a technical effect outside the computer (e.g. controlling anti-lock braking or restoring distorted digital images), as well as being new, non-obvious and useful. Despite these hurdles, a 2019 World Intellectual Property Office report states that the numbers of AI patent applications have been growing by an average of 28% year-on-year since 2012.
The software code underlying the AI system may be protected by literary copyright, although its functionality and underlying algorithms may not (as copyright protects the expression of ideas and not the ideas themselves).
Training datasets may also be protected by literary copyright if they are sufficiently creative and, in the European Union, potentially also by the separate “sui generis” database right if they have been subject to sufficient investment. All elements of an AI system (code, algorithms, data etc.) may comprise trade secrets if they are kept confidential.
IP Protection for AI Outputs
The availability of IP protection for the outputs of AI systems is one of the most difficult issues raised by these new technologies. Where the AI system can be characterized as a mere tool employed by a human (akin to a paintbrush), the availability of IP protections is relatively uncontroversial. However, where the AI system is independently creative or inventive without human involvement, a number of difficult issues arise.
From a patents perspective, the key unresolved issue is whether a human inventor is required for an invention to be patentable. An “inventor” is defined in U.K. legislation as the “actual deviser of the invention” and in U.S. legislation as the “individual” who invented/discovered the subject matter of the invention.
Both of these definitions allow room for argument in the context of AI. The general view among patent offices appears to be that a human inventor is required, but this remains untested. However, in August, test applications were reportedly filed in the U.S., U.K., and at the European Patent Office which named “DABUS” —an AI “creativity” machine—as inventor. These pending applications concern two inventions (relating to food containers and flashing lights) that DABUS created autonomously without human involvement, and will likely force an answer to these difficult questions in the near future.
The question of inventorship is also inextricably linked to ownership, because the default position is that the inventor owns the patent. However, as AI systems are not legal personalities capable of ownership, this raises the additional issue of who can (and should) be entitled to the patent: the AI system or its owner, deviser, trainer or user?
From a copyright perspective, the UK extends protection to “computer-generated works”, i.e. works generated in circumstances such that there is no human author. In this case, the “author” (and first owner) is “the person by whom the arrangements necessary for the creation of the work are undertaken”.
This approach is unusual, as most other jurisdictions place human contribution at the heart of the tests for whether copyright protection is available. In respect of training datasets, where these are machine-generated, it is generally accepted that the EU sui generis database right does not apply.
Unique Issues for Enforcement and Infringement
One difficulty when enforcing AI patents is that it is often not clear how an AI system actually works, so it may be difficult to establish and prove infringement. While some commentators have suggested possible solutions (e.g. reversed burdens of proof), none has gained significant traction.
Another healthy debate exists around the legitimacy of using third party copyright-protected content to train AI systems, with varying approaches already emerging in different countries.
For example, the U.K. currently permits “text and data analysis for non-commercial research”. The new EU Copyright Directive (to be implemented in EU Member States before June 2021) follows this approach and also mandates a broader exception for commercial data mining, but only where the relevant rightsholder has not opted-out.
Conversely, following the Google Books case, the more flexible U.S. defense to copyright infringement of “fair use” is likely to permit use of third party content to train AI systems, provided such use is sufficiently transformative and does not compete with the original works.
These differences will continue to cause headaches for international businesses and lawyers alike, as well as potentially influencing the location of the next wave of AI advancements.
This column does not necessarily reflect the opinion of The Bureau of National Affairs, Inc. or its owners.
Yohan Liyanage is a partner and Kathy Berry is a professional support lawyer at Linklaters in London. They advise on a broad range of contentious and non-contentious matters in relation to intellectual property rights.