Executive summary
A decision to adopt AI can raise fundamental and moral issues for society. These are complex and vital issues that are not typically the domain of lawyers.
We use cookies and other similar technology to collect data about you to allow us to deliver our online services, measure our website audience and improve your browsing experience. Full details on the cookies we use are set out in our Cookies policy. Please click OK to signify your consent to our use of cookies.
You can withdraw your consent by clicking “manage cookies” and following the instructions shown.
Part of our Artificial Intelligence briefing
Global | Publication | December 2019
By Seiko Hidaka, Counsel, and Will Garton, Associate, Norton Rose Fulbright LLP1
The product of artificial intelligence (AI), based on the machine learning algorithm “Generative Adversarial Network,” is being cited as one of the most sophisticated AIs ever created. This type of AI gives us a glimpse of what is to come – a world where the machine creates new and better things autonomously. But what is the role of intellectual property (IP) in such a scenario? Understanding how AI operates, how current IP law is applied to AI and generated works, and how it impacts upon the ability of industry to protect their AI-related work is important, and will require close monitoring of industry voices and governmental policies over the years ahead.
Inventorship (in the case of inventions), authorship (in the case of copyright works), and more generally the ownership of IP created by AI is a hotly debated issue. The UK Intellectual Property Office (UK IPO) has made its position clear in respect of patents by adding in its formalities manual that it is not acceptable to designate an ‘AI Inventor’ as the inventor “as this does not identify ‘a person’ which is required by law. The consequence of failing to [identify a person] is that the application is taken to be withdrawn under s.13(2)”. The US Patent and Trademark Office (USPTO) has issued a request for comments from the public to a series of questions concerning IP protection relating to AI-generated works so that it can inform itself of the impact of AI for the development of future guidance and policies in this area. Inventorship/authorship and ownership considerations feature near the top of the list of questions.
The issue needs consideration because it goes to the heart of IP rights, which is to “contribute to the promotion of technological innovation and to the transfer and dissemination of technology, to the mutual advantage of producers and users of technological knowledge and in a manner conducive to social and economic welfare, and to a balance of rights and obligations.”2 AI machines/technology are not humans. As such, they do not require incentives such as IP rights. Machines would just as efficiently create works currently protectable by IP rights without incentives as they would with incentives, or without IP protection. The absence of IP protection, however, may disincentivise humans from creating – some would argue that humans require incentives such as IP rights in order to invest in developing highly advanced AI with the ability to invent and create.3 So how should the balance be struck?
Within the field of AI, machine learning allows computers to execute tasks which are not specifically pre-programmed, disrupting many sectors. A well-known advanced machine learning algorithm, “Generative Adversarial Networks” (or “GAN” for short), has gained popularity for its ability to create seemingly inventive and creative products and images independent of human input. In what follows, we describe some examples of GAN at work, look at how the present laws might apply, and consider how laws might be shaped in the future to accommodate advanced AI systems such as GAN.
GAN is a machine learning algorithm which operates through two neural networks: the “generator” and the “discriminator.” The generator network generates new synthetic datasets, based on random input variables. The discriminator network is loaded with a training dataset (for example, all the existing products of a particular company). The discriminator network assesses each new instance of data created by the generator network to decide whether it is genuine.
A constant feedback loop telling each network how close it was to fooling the other network acts to indirectly “train” both networks to optimise their respective functions – that is to say, to make the generator network better at generating synthetic datasets closer to the training dataset, and the discriminator network better at spotting these fakes. The networks are, in other words, competing and learning from each other, causing both networks to move closer to an equilibrium.
Eventually the generator network is able to generate what could be described as a “passable” synthetic dataset which the discriminator network does not determine to be fake. The passable synthetic dataset could be said to be “inspired” by the training dataset, but it is a whole new instance of data created by the AI.
The GAN algorithm can be applied to create new data in fields as diverse as engineering, corporate strategy and economic predictions. Three different kinds of applications of the GAN algorithm are described below.
“DABUS,” which stands for “device for the autonomous bootstrapping of unified sentience,” is an AI developed by Stephen Thaler, based on the GAN algorithm (Dabus AI). A team led by Ryan Abbott, a professor of law and health sciences at the University of Surrey, has reportedly filed patent applications in jurisdictions around the world (including to UK IPO and the European Patent Office (EPO)) for products produced by the Dabus AI - a beverage container based on fractal geometry and a flashing light for search-and-rescue missions. These applications mark a challenge to the international patents regime because Dabus AI is said to have invented the invention.
The London-based start-up called Benevolent has reportedly created an AI-based on the GAN algorithm (Benevolent AI) to aid drug discovery and screening. The Benevolent AI uses data from research papers, patents, clinical trials and patient records to infer relationships between biological entities such as genes, diseases, and candidate drugs. Benevolent AI recently partnered with AstraZeneca to collaborate on the use of AI for drug discovery and development of new treatments for chronic kidney disease and idiopathic pulmonary fibrosis.4
Obvious, a French art collective, is reported to have created an AI-based on the GAN algorithm (Obvious AI). In October 2018, the Obvious AI produced a number of art works, including the Portrait of Edmond Belamy – the first AI-created artwork ever to be sold at action. While the relevance of such an AI application in industry is perhaps limited, image synthesis is one of the most popular applications of GAN, and it raises interesting questions about the copyrightable works produced by an AI.
The European Patent Convention 1973 (EPC) provides a process for obtaining a European patent which, once granted, acts as a bundle of national patent rights in each country designated in the patent application. Under Article 60(1) of the EPC, the “right to a European patent shall belong to the inventor…” and “if the inventor is an employee, the right to a European patent shall be determined in accordance with the law of the State in which the employee is mainly employed”. Another provision, Article 62 of the EPC, ensures that inventors have the right to be mentioned before the EPO – this is a moral right, which is inapplicable to AI systems. Furthermore, Rule 19 requires the inventor to be designated by reference to family and given names, and by Rule 21, the EPC lays states that the designation of an inventor can be rectified only with the consent of the wrongly designated person. These provisions, taken together, are likely to have formed the basis for the position that EPC envisaged that inventors should always be humans leading to the rejection of the application in which Dabus AI was designated as inventor.
In addition to the wording of relevant provisions, there are also substantive challenges. The EPC and its associated case law do not provide a specific definition of who an inventor is. The UK legislation provides a hint. Section 7(3) of the Patents Act 1977 (PA 1977) defines an inventor as "the actual deviser of the invention" – that is, the party who made a contribution to the conception of the invention.
A research report commissioned by the EPO and published in February 2019 found that similar laws and concepts applied around other major jurisdictions, such as the US, China, Japan, Korea and other major EU states. The report concludes that the current European legal framework is “suitable for addressing the inventorship…of inventions involving AI activity” by viewing AIs, regardless of their intelligence, as merely “tools.”5
Applying this to our Dabus AI example, it would seem that Dabus AI would be merely a tool even if it were instrumental (if not decisive) in the creation of products using the GAN algorithm. In its preliminary opinion, the EPO took the same approach by commenting that “right to an invention and a moral right to be designated as inventor can belong only to a natural person” and “machines do not have legal personality and cannot own property... a machine cannot own rights to an invention and cannot transfer them”. It therefore came as no surprise when the EPO refused the applications in the minutes of its oral proceedings, concluding: “After hearing the arguments of the applicant… the EPO refused [the patent applications] on the grounds that they do not meet the requirement of the EPC that an inventor designated in the application has to be a human being, not a machine.”
The UK IPO’s position appears aligned with this position. However, a spokeswoman for the EPO is recently reported by the BBC to have said that there is a “global consensus that an inventor can only be a person who makes a contribution to the invention's conception in the form of devising an idea or a plan in the mind.”6
Questions also arise in relation to the ownership of the resulting patent.7 At present, AI systems cannot own, or have the right to a patent. That is the preserve of those conferred with legal personality. If only persons with legal personalities can be owners of a patent created by an AI, then any of the contributors to the outcome of the AI system could be the owner. In the vast majority of commercially significant cases, ownership would likely depend on the terms of contract between those involved in the particular arrangement.
If we consider the artwork created by Obvious AI, there are similar questions to be asked of authorship and ownership of AI-produced works for the purposes of copyright.
EU copyright law is made up of multiple directives with the Information Society Directive8 at its core, all implemented by domestic legislation within individual member states to create a relatively harmonised body of law. All of these directives implement the provisions of the Berne Convention, which has been interpreted by the Court of Justice of the European Union (the CJEU, the highest court in the EU) as requiring originality for a work to attract copyright protection in the sense that the work is the “author’s own intellectual creation.”9
The CJEU recently reiterated (by reference to established case law) that, in order for an intellectual creation to be regarded as an author’s own, it must reflect the author’s personality, which is the case if the author was able to express his creative abilities in the production of the work by making free and creative choices.10
Such an interpretation of originality seemingly excludes works produced by advanced AI systems such as the GAN algorithm. Indeed, the Computer Programs Directive11 provides that:
It follows that the question of legal personality – already problematic with respect to patents – is also a problem in relation to copyright.
In the UK there is a provision for works that are “computer generated”. The Copyright, Designs and Patents Act (CDPA 1988) defines a computer-generated work as one “generated by computer in circumstances such that there is no human author of the work".12 For such works, the author is taken to be the person “by whom the arrangements necessary for the creation of the work are undertaken” (section 9(3), CDPA 1988). However, this provision may possibly not be consistent with EU case law, which requires human input, or “intellectual creation.”13
For Obvious AI, it may be said that the GAN is simply using statistical models to extrapolate aesthetic principles from the training dataset that is input by the programmer.
If that is the case, the necessary “arrangement” is still being performed by the programmer when selecting the training dataset and inputting it to the discriminator network, which could well entail intellectual creation.
In many ways, however, GAN designs itself based upon the two neural networks interacting with each other. Therefore, as we come closer to an AI system that is truly autonomous, or an AI system designed entirely by another AI system, what amounts to “necessary arrangement” may become harder to identify.
The German legal position distinguishes between automated systems (more likely to be eligible for copyright protection) and autonomous systems (less likely to be eligible). These conceptual distinctions could become useful when considering the position of AI at the advanced end of the spectrum.
The level of debate surrounding inventorship/authorship and ownership of works created by AI would suggest that current laws need to be looked at with an eye to future developments. Cognisant of this issue and other grey areas in law concerning the use of AI in patent protection, as mentioned above, the USPTO published a request for comments on the Federal Registry on patenting AI inventions. This was followed up by a similar request for comments on the Federal Registry in relation to other forms of IP rights.
The USPTO’s request for comments for AI inventions comprises twelve questions concerning wide-ranging patent-related issues regarding AI inventions for the purpose of evaluating whether further examination guidance is needed to promote the reliability and predictability of patenting such inventions.
Among the USPTO’s 12 questions, there are two inventorship/ownership questions:
Question 3: Do current patent laws and regulations regarding inventorship need to be revised to take into account inventions where an entity or entities other than a natural person contributed to the conception of an invention?
Question 4: Should an entity or entities other than a natural person, or company to which a natural person assigns an invention, be able to own a patent on the AI invention? For example: should a company who trains the artificial intelligence process that creates the invention be able to be an owner?
In the USPTO’s extended request for comments for AI in relation to IP other than patents, the first two questions concerned authorship and the fifth question concerned ownership:
Question 1: Should a work produced by an AI algorithm or process, without the involvement of a natural person contributing expression to the resulting work, qualify as a work of authorship protectable under US copyright law? Why or why not?
Question 2: Assuming involvement by a natural person is or should be required, what kind of involvement would or should be sufficient so that the work qualifies for copyright protection? For example, should it be sufficient if a person (i) designed the AI algorithm or process that created the work; (ii) contributed to the design of the algorithm or process; (iii) chose data used by the algorithm for training or otherwise; (iv) caused the AI algorithm or process to be used to yield the work; or (v) engaged in some specific combination of the foregoing activities? Are there other contributions a person could make in a potentially copyrightable AI-generated work in order to be considered an ‘‘author?’’
Question 5: Should an entity or entities other than a natural person, or company to which a natural person assigns a copyrighted work, be able to own the copyright on the AI work? For example: Should a company who trains the artificial intelligence process that creates the work be able to be an owner?
The US has approached the question as a public policy issue, rather than relying on courts to interpret the wording of laws that were created significantly before AI emerged in the state we know it today.
Part of the reason it has done so may be because the involvement of AI in protected works might not be entirely compatible with the foundation of the IP system. The Patent and Copyright Clause of the US Constitution states that “[The Congress shall have power] to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries.” There may well be economic drivers behind the US approach too, because a keenly-sculpted IP policy can provide economic dividends from an industry reliant on advanced technology such as AI.
In relation to copyright in artificially-generated works, at the recent AIPPI Congress 2019 in London, delegates from around the world were asked to consider whether humans should be “essential” for the creation of an original work of art. The answer of the “vast majority” was positive, concluding that AI-generated works should only be eligible for protection by copyright if there is human intervention in the creation of the work. The basis for such a view was that copyright protection should not be awarded to provide an incentive for works without human input. The majority of participants at the conference also felt that the current legislation in their own jurisdictions could be improved in order to give greater certainty and clarity regarding the legal conditions for the protection of AI-derived work. The USPTO is certainly paving the way for the development of a suitable framework with future capabilities of AI in mind with the objective of providing certainty and clarity for businesses.
The increasing complexity and prevalence of AI in business is likely to generate many IP issues, which in turn will drive legal developments in IP law. If AI-derived work products can attract some form of IP right, then the contractual terms are likely to decide who will own the resulting IP right and so who can exploit the results generated by the AI.
This situation is somewhat analogous to other scenarios where a valuable intangible output is created, but the IP rights in that output are complex or uncertain. A common example is the licensing of datasets, where database right or database copyright may or may not arise in the dataset, depending on the circumstances in which the dataset was created. In these scenarios, “contract is king” and, while it is common for the licensor to assert ownership of any IP that may arise, the licensor will mainly rely on detailed contractual restrictions on use rather than IP rights to protect its position.
It follows that, if parties are seeking to enter into an agreement for the development and licensing of an AI system based on something autonomous (like GAN) which is expected to evolve and produce new works, they should consider what the following express terms in the contract should be:
In reality, the negotiating powers of the players involved is likely to dictate who gets to own any IP rights that there may be.
A decision to adopt AI can raise fundamental and moral issues for society. These are complex and vital issues that are not typically the domain of lawyers.
AI is a field of computer science that includes machine learning, natural language processing, speech processing, expert systems, robotics, and machine vision.
AI will need to meet certain minimum ethical standards to achieve sufficient end user uptake, varying according to the type of AI and sector of deployment.
Courts in a number of countries have already had to address a range of legal questions in relation to the automatic nature of machines and systems.
The key question businesses need to consider is whether deploying AI will result in a shift of ethical and legal responsibility within their supply chain