In September 2019, WIPO started what it called a "conversation" on the impact of AI on IP policy. The second session in this conversation was held virtually between 7 – 9 July 2020, with more than 50 speakers commenting on WIPO's Revised Issues Paper. Katharine Stephens and Toby Bond summarise the main points arising from this session.
The first day was devoted to consideration of a number of fundamental issues relating to IP protection for AI-generated and AI-assisted works and inventions. Everything from general policy considerations relating to the patent system, inventorship and ownership of patents, copyright and related rights, bias and deep fakes was touched upon. Representatives from a number of national patent offices and the EU Commission noted that, in addition to WIPO's conversation, they were consulting on the issues raised by AI. Nevertheless, there was wide support for WIPO's conversation and the opportunity it gave to develop international consensus and harmonisation of laws. Due to the number of issues involved, WIPO was encouraged by some speakers to develop a list of priorities and appeared ready to do so. Of the topics discussed over the three days, copyright authorship and ownership, patent inventorship and ownership, sufficiency of description in patents and the use of datasets for training AI systems appeared to be some of the most pressing issues.
It was noticeable that there was a tension between those speakers who urged caution against hasty policy decisions and those who considered that there was a need for urgency. In support of the later proposition, one speaker pointed out that AI assisted inventions have been with us for a number of years and inventorship was already an issue.
Despite this, many speakers considered that present IP laws were adequate to deal with today's narrow AI systems. However, there was less agreement about how AI generated works and inventions should be dealt with. Many speakers urged that no protection, whether patents or copyright, should be granted to AI generated works. Others pointed out that, absent any legislative provisions, there was a danger of someone wrongly taking credit as the inventor or author of an AI generated work. There was also the danger that the invention could be lost or fall into the public domain. In other words, there were diverse views on how and who the IP system should incentivise. Further, some speakers urged the need for a a new sui generis right to resolve the situation, whilst others warned that a new IP right could restrict competition.
Again, on the second day, a number of fundamental questions were raised, in particular, on the role the patent system should play in relation to AI inventions. Judge Klaus Grabinski of the Federal Court of Justice, Germany, opened the proceedings with a short master class on AI's interaction with patentable subject matter, obviousness and disclosure. On the first issue, he noted that AI may be of an abstract mathematical nature or a computer program and therefore not patentable. However, when AI solves a technical problem by achieving a technical result, it is to be considered patentable subject matter. In relation to obviousness, today's AI is a tool in the hands of the person skilled in the art, with that person still in command. Therefore, although this may lead to a rise in the threshold for inventive step, it is entirely different to the considerations needed in the future where AI autonomously generates an invention. In such (future) cases, the AI would be acting without any intervention from a (natural) person skilled in the art, whether posing a technical problem or finding a solution with reference to the state of the art. He preferred to leave open how this situation would be judged. Finally, in relation to disclosure, Judge Grabinski pointed out that patents should not be granted for "black box inventions". Instead, the specification should disclose how the AI is structured and how it works and interacts with the environment so that the person skilled in the art can reproduce the invention.
Many of the participants considered that a redesign of the current legislation on patentability was not required, but that the guidelines on how to deal with AI related inventions may need changing. Certainly, having clear examination standards and example cases would be useful for both examiners and users.
There was some debate on the issue of inventive step, with suggestions that the person skilled in the art may have to be redefined. However, it was pointed out that if the present approach is to be varied, there is an immediate difficulty. Under the present system, there is a clever person, the inventor, and his or her invention is judged against the knowledge of the average skilled person. If AI is both inventor and average skilled person, how can the inventive step be judged?
The third issue relating to sufficiency of description generated the greatest debate. On the one hand, there was a concern about black box inventions but on the other a concern that if the requirements on disclosure were too strict, they could conflict with trade secrets and might stand in the way of innovation. One speaker pointed out that, in fact, there might be no need to disclose AI's use in the inventive process: if the AI was already published and readily available, arguably any invention would be obvious as it would only involve routine experimentation; but if not and if AI was used to create an invention, that would only be one way of making the invention and therefore it would not have to be disclosed, nor would AI's involvement in the inventing process have to be identified. Of course, if the scope of protection included AI, then it would have to be disclosed in sufficient detail and this led back to the black box problem. A number of speakers suggested a deposit system might be a solution to the issue for both algorithms and training data. The Budapest Treaty on the deposit of microorganisms could serve as a precedent, although the issue was predicated on what disclosure was needed. Then there were the questions of how people would have access to the algorithm and training data and how much such a system would cost.
Day three focused on issues relating to the use of data to train AI systems. A number of contributors agreed that facilitating greater access to data to train AI systems was a general public good and would stimulate AI based innovation. Greater access to data would also reduce the potential for bias, by allowing AI systems to be developed using more diverse and representative data sets. In this context much of the discussion focused on the application of existing exceptions to copyright infringement to the use of copyright works in the process of training AI systems, often referred to as text and data mining (TDM). A number of jurisdictions have general "fair use" type doctrines which have already been applied to permit the use of copyright works to train AI systems. However, it was noted that fair use does not provide a blanket exception to all possible uses of copyright works to train AI systems. In particular fair use exceptions include their own checks and balances and will need to comply with the three conditions set out in Article 9 of the Berne Convention.
The discussion also highlighted that the approach to exceptions permitting TDM has been quite mixed in jurisdictions which don’t already have general fair use provisions. Some jurisdictions have research and temporary copies exceptions but speakers questioned how effective these are at facilitating AI based research where researchers are restricted in their ability to retain and share copies of their training data with others. TDM specific exceptions have also started to appear in a number of jurisdictions and permit a party to use copyright works lawfully in their possession for the purpose of TDM, subject to certain safeguards for the rights holder. Articles 3 and 4 of the EU's Digital Copyright Directive are well known examples but other jurisdictions are also moving in this direction, with Singapore being a recent example of a jurisdiction planning to introduce a permissive TDM exception. A number of contributors suggested that achieving the right balance between the interests of AI developers (and those who benefit from their systems) and the interests of rights holders was an area where WIPO could play an important role.
While much of the discussion focused on copyright in AI training data, a number of contributors highlighted that copyright was only one of a number of rights which can subsist in the data used to train AI systems. They stressed that it was important to take a holistic view of the rights which can apply to data, including trade secrets, privacy, contractual rights, unfair competition and the application of technical measures to protect data. One of the issues WIPO intends to consider is whether further rights in data are desirable. Some contributors suggested that the rights currently available to protect certain categories of data (e.g. sensor data) are not sufficient and, given the economic value of this data, it would be desirable to have a new right to facilitate licensing. Others suggested that existing rights are sufficient and introducing a new right could upset the balance between protecting investments made in relation to data and unlocking further value through allowing the re-utilisation of that data. Rights to access and rights of remuneration for providing access were suggested as one policy mechanism which could assist in certain contexts.
The development of AI technology clearly requires high quality data inputs. The contributions during day three made clear that challenge for IP policy is to encourage and reward the generation of those inputs without stifling their use by AI developers. As data becomes fundamental to our society the rights which can control access, use and dissemination of data also have a much wider implications beyond the AI context and touches on broader legal issues including data privacy, competition law and cybersecurity. If the discussions during day three are anything to go by there is still much ground to cover before we arrive at a unified approach to rights in data. Of interest to those in the UK was an indication from the UKIPO that they intend to consult on the IP issues arising from AI systems in the autumn of 2020.