Thanks to Katharine Stephens (Bird & Bird) for this guest post:
On 27 September 2019, WIPO held a Conversation on IP and AI; a fascinating day of presentations and discussion on the impact of AI on IP systems, IP policies, IP rights management and international cooperation on IP matters. The list of presentations and the presenters' biographies give you a view of the breadth and depth of the discussion. However, this report has to start with an apology as it reports on the legal discussions only and in the very briefest terms and misses out all mention of the presentations on the use of AI by the EPO and other patent offices.
Francis Gurry, Director General of WIPO opened the day by stressing the need to engage now with the issues raised by AI, noting that Governments are starting to engage on strategy, opening up data for commercial use, and the need for regulation. The day's "conversation", together with WIPO's recently published landscape of the patent data in AI and its development of tools using AI applications for administration of IP (which were described and in some instances demonstrated during the course of the day), was all part of WIPO's engagement with those issues.
The impact of AI on the IP system and IP policy
There was a very upbeat note sounded by Andrei Iancu, Under Secretary of Commerce for IP and Director of USPTO, at the start of the first panel discussion. AI has the potential to be the most disruptive technology we will see in our lifetime, but that does not mean that it has to be seen in a negative way and we were asked to consider some of the many positive examples of AI, such as personalised medicine. As to policy, AI raises hugely important questions on IP law. For example, if a machine invents, is it an inventor? If it is, who owns the IP? As AI innovations are a black box, how does that satisfy the requirements for disclosure? Is a new right needed to protect the datasets used to train AI systems so that companies are encouraged to share them? These are just some of the questions on which the USPTO has requested comments (deadline for responses extended to November 8, 2019).
In the same session, Nuria Oliver, Chief Data Scientist in Data-Pop Alliance and Chief Scientific Advisor to the Vodafone Institute, noted that the way AI-generated content was not only dependent on the data input (the same algorithm might generate a work of art or a cartoon depending on the input), but was something of a black box. This was echoed by Zeng Zhihua, Director General in Automation Department and Patent Examination Cooperation Guang Dong Center, CNIPA. In particular, Zeng noted that if AI was to be used by patent offices, its decision-making processes had to be transparent.
AI and patents
Belinda Gascoyne, Senior IP Law Counsel, IBM, noted that IP laws were written when there was only a human author in contemplation. These are now being put to the test; for example, the University of Surrey has announced that its creation machine, DABUS, is named as the inventor of a new food container, the subject of patent application GB1816909.4. Belinda also discussed the issue of the patentability criteria for AI-related inventions and the lack of coherence as between countries and even within some countries as between patent office and courts. She noted that the EPO had amended its Guidelines for Examination in November 2018 to include a section dealing with AI, but the referral pending before the EPO's Enlarged Board of Appeal in G01/19 could pose a threat to the stability of the test of what is patentable in Europe. Belinda agreed with AIPPI's 2017 resolution on the patentability of computer implemented inventions ("CIIs") which stated that "patents should be available, and patent rights enjoyable, without discrimination for inventions in all fields of technology, including CIIs". This was echoed in the "refreshing" approach taken by the JPO in their recently updated examination guidance and the case examples pertinent to AI-related technology.
Beat Weibel, Chief IP Counsel, Siemens, contrasted new inventions that incorporate AI technology and inventions made by AI. The former were CIIs and should not be treated any differently from normal CII inventions. AI-generated inventions should be patentable, but the question was who should be the inventor? One could "pretend" that a natural person is the inventor (not a good solution in the long run); or nominate a machine as an inventor (also not a good solution as machines do not have rights or duties); or, and this was his preferred solution, expand the definition of inventor to include the legal person who controls the AI system.
Zhixiang Liang, Vice President and General Counsel, Baidu, underlined the need to ensure respect for data privacy and safety. He referred to the book 'AI Superpowers' and noted that AI, for a big company such as Baidu, meant super-responsibility.
Socio-economic and ethical impacts of AI
The IPO's Chief Executive and Comptroller General, Tim Moss, moderated the third panel. Nuria Oliver again spoke and introduced us to the acronym FATEN to describe 5 dimensions of the ethical principles raised by AI, namely:
A: (human) autonomy and accountability
T: trust and transparency
E: (b)eneficience, education and equality
Tom Ogada, Executive Director, African Centre for Technology Studies, rightly warned the audience about the widening technology gap between countries and was hopeful that, not only could AI improve productivity, but that it could also increase numbers of jobs.
AI and copyright
In the afternoon we were treated to a discussion on whether AI will change human creativity and how AI-generated works should be protected, chaired by Karyn Temple, Register of Copyrights and Director, US Copyright Office.
Pierre Sirinelli, Professor of Private Law and Criminal Science at the University of Paris took us through the arguments on whether copyright exists in AI created works. His starting point was that the form and identity of a work produced by an AI system was the same as one created by a human. However, if we maintain that originality is an imprint of personality then such a work will not be protected. Although it might be possible to sidestep that test by saying that an author makes arbitrary choices and a machine does the same thing, this would ultimately not work because people make subjective decisions whereas machines are cold and objective. Looking to countries where a work receives copyright when effort and investment is made, this still does not help because that effort and investment goes into creating the AI system, not the work which is generated by the press of a button. Another problem in the copyright system is that a work has to come from a physical person. It might, again, be possible to get around this requirement by looking at those copyright systems which immediately transfer copyright such as, in an employment scenario, from the employee to the employer. However, this is still not a solution since there is no person creating the work and so no one who can make the transfer. He concluded that if we want protection for AI-generated works, we must either find a right outside copyright – a new sui generis right – or abolish the link to a physical person within our copyright rules.
Pravin Anand, Managing Partner, Anand and Anand, started by telling us that the Indian courts have considered that idols and animals are legal entities for the purpose of owning copyright. If an animal or idol can be a person for the purposes of copyright then a machine should also qualify. He backed this up by considering a number of deeming provisions under Indian law: the author of a film is the person who took the initiative and took responsibility for its creation; in relation to computer-generated works, the author is the person who has caused the work to be created; and an employer owns a work created by an employee in the course of his/her employment. He thought that a similar deeming provision could be used for AI.
Andres Guadamuz, Senior Lecturer in Intellectual Property Law, the University of Sussex and author of "Do androids dream of electric copyright?", asked the question: since machines are getting so good at creating works, if those works remain in the public domain, what will happen to works generated by human authors? Will the humans be able to compete in the marketplace? He plays a game with his students which he calls "bot or not" where he shows them photos and paintings and plays music and reads poems to them. Over the last 5 years, the students had found it increasingly difficult to identify those works created by a human and those created by AI. He concluded that the UK has the best system to deal with AI-generated works under section 9(3) CDPA 1988 which provides that copyright in a computer-generated work goes to the person by whom the arrangements necessary for the creation of the work are undertaken. Similar provisions also exist in Ireland, NZ and in India.
Kats, of course, dream of regular sheep (and mice)
Tobias McKenney, Associate Copyright Counsel, Google, noted that the text and data mining ("TDM") exception was a very important touchpoint when considering whether the underlying data should be open to all. Questions that have to be borne in mind are: is AI safe and is it discriminatory? How do you answer those questions? Looking at the process does not answer the question of whether the output is safe. The big debate at the moment is how to make sure that AI is explainable. He gave the example of a complaint that had been made about Quick, Draw!, a game where the computer tries to guess what you are drawing. The more it is used, the better the machine learning gets. At one point, and to Google's surprise, a user complained that it was biased because, although it was good at recognising drawings of sneakers, it was bad at recognising high heeled shoes and ballerina shoes. But bias can be much more serious than that and Tobias gave us three points to consider:
- If data in the public domain, such as literary works written before 1870, are used in the input to an AI system, think about what those works may say about gender and race.
- If you need to demonstrate safety and reliability of an AI system, you cannot have a TDM exception which requires you to destroy data immediately after it has been used.
- Datasets have to be observable and if copyright laws make that an infringement, then, again, how can you show that an AI system is not biased?
AI and data
Erich Andersen, Corporate VP and Chief IP Counsel, Microsoft, started the debate on data protection and free flow of data by pointing out that we needed to think about privacy law as a 'new' protection for data. Microsoft had drafted three data sharing agreements which had been annotated with lots of helpful legal points in an effort to remove barriers and empower people and organisations to share and use data more effectively.
Jonathan Osha, Reporter General of AIPPI, had asked AIPPI's members whether a new right was needed to protect data used in AI systems; and to his surprise they were split 50:50. To get to a solution, therefore, he realised that the question had to be more specific i.e. are we talking about vast pools of unstructured data or training data? Each scenario raises different IP questions and the balancing of different interests such as increasing innovation, societal benefits and privacy issues. The question of whether to create a new right is going to be the subject of an AIPPI study question for next year. AIPPI's resolution on copyright in AI-generated works is about to be published (spoiler alert - see the IPKat's report of the debate at the AIPPI Congress on this topic).
Andreas Wiebe, Chair for Civil Law, Intellectual Property Law, Media Law and Information and Communications Technology Law, University of Göttinger, pointed out that there was
no copyright in raw data, although it could be protected by confidentiality. Data is produced whether there is a right in it or not therefore there was no need to incentivize its creation. The question was whether we need a new right to protect data on disclosure. The other big issue was allowing access to data. He was of the view that a compulsory licence was going too far and that competition law would probably not help as it was difficult to say whether data gives market power/ dominance. He therefore queried what could be done now to support markets in their developments of AI.
Virginie Fossoul, Legal and Policy Officer, European Commission Directorate-General for the Internal Market, Industry, Entrepreneurship and SMEs noted the importance of open data. She warned that creating a new right was very complicated as it required a balancing exercise between competing interests and warned against rushing into legislation. It was an extreme option to say there should be compulsory licencing as in the payment services directive. She also referred to the directive on reuse of public sector data, but noted that it was very difficult to find a horizontal solution across all sectors. It was better to foster sharing without legislation until we were better able to understand the markets.
Erich then picked up on the TMD exception in the directive on copyright and related rights in the digital single market. Data is the fuel for AI and therefore it is important that society has access to data, particularly in the medical field where a number of breakthroughs have been made by looking at big datasets. Ursula Von der Leyen had made the fuelling of modern innovations one of the planks of her agenda for her presidency of the European Commission, whilst also noting that a balance needed to be struck between the free flow of data and privacy issues. To this end, an enormous amount of research is going into anonymising data and Microsoft had just announced that they were working on an open data differential privacy platform. (This is where the technology does not strip out personal data, rather it does something similar to inserting white noise into the data so that it is anonymised). He finished by noting that researchers need to work with Governments to find safe-harbours so that they can be confident that they are working within the law.
The final word went to Andreas who turned to a very apt quote from Stephen Hawking: "Whereas the short-term impact of AI depends on who controls it, the long-term impact depends on whether it can be controlled at all."
In closing the day's conversation, Francis Gurry thanked the presenters for their invaluable input to the debate and stated that there would be questions on these issues published for consultation in either October or November of this year.
[Guest Post] IP and AI - the debate continues, this time at WIPO Reviewed by Alex Woolgar on Monday, October 14, 2019 Rating: