Use of artificial intelligence has grown enormously in recent years. A decade ago, machine learning was a new and exotic technology—at least, for mainstream commercial applications—with few companies patenting ML-based solutions. Since then, use of AI has gradually become mainstream, with more and more companies employing ML solutions in their inventions by the mid-to-late 2010s, and concepts like neural networks becoming household names. Indeed, interest in AI has become so keen that questions previously found only in works of science fiction have begun to receive more serious consideration. One such question is the possibility of according rights in intellectual property—formerly accorded only to natural persons—to AI software, if the AI software indeed generated the IP in question. Recently, for example, a patent application was filed listing AI software “DABUS” as the sole inventor. The patent examiner rejected the application, and in a decision on April 22, 2020, the U.S. Patent and Trademark Office upheld the rejection, arguing that under the current law only natural persons may be considered inventors.
This decision based on current law, however, leads to more theoretical questions regarding how AI should be treated under the IP laws, such as the theoretical plausibility, and practical value, of granting inventorship and/or ownership in IP to AI software. Indeed, the USPTO had previously solicited input on these very questions. In August 2019, the USPTO issued a request for comment on the patenting of AI inventions. The 12 questions posed by the USPTO were wide-ranging and comprehensive, addressing not only hot-button issues such as patent eligibility under § 101 of the Patent Act (question 5), but also questions of inventorship and ownership (questions 3 and 4), whether AI impacts the level of a person of ordinary skill in the art (question 8), and prior art (question 9). They also addressed more foundational questions, such as what constitutes an “AI invention” (questions 1 and 2) and whether the patent laws should be revised to better address AI issues, such as inventorship by AI (question 3). Other questions addressed the required degree of written description and enablement (questions 6 and 7), and whether non-patent forms of protection of AI-related data would be beneficial (question 10).
Several months later, in October 2019, the USPTO additionally issued a request for comment on non-patent IP issues for AI, such as copyright, trade secret and trademark. This article will focus solely on the patent issues, however.
Areas of Agreement
The responses received from the public included 56 from individuals and 43 from organizations. Perhaps surprisingly, there was near-unanimity on a number of issues: AI is merely a tool, like any other tool. The popular imagination is captivated by the vision of an artificial general intelligence, a strong AI with intellectual capability equivalent to that of a human being, and the USPTO’s questions appear to some degree to be influenced by that vision. In contrast, the public responses by both corporations and individuals unequivocally stated that artificial general intelligence is currently of merely philosophical interest, and that in practice AI is simply a tool used by humans.
Hostility to the concept of bestowing legal personhood upon AI. Although many areas of the law have granted quasi-personhood to entities such as corporations, commentators were opposed to legislation under which an AI program could invent or own intellectual property. Many commentators argued that awarding inventorship to an AI program would not comport with the constitutional requirement to “promote the progress of science and useful arts,” since machines lack the volition that would allow them to act based on incentives to create. It was additionally noted that of the IP5—the five leading IP jurisdictions—only Europe fails expressly to define inventorship as being limited to human persons.
Need for analytical distinctions between AI categories. Commentators noted that it is important to distinguish between the degrees to which AI is involved in an invention when analyzing the relevant legal issues. For example, is the invention directed to an improvement in fundamental AI techniques themselves, or are existing AI techniques merely being used to solve a particular problem in a particular domain? Such distinctions may lead to sharply different analyses of the issues posed by the USPTO.
Eligibility concerns. AI is almost invariably implemented in software, and the law of eligibility post-Alice can be problematic for software, particularly for AI software that can be considered to represent mathematical processes, one of the categories considered by USPTO guidance to constitute an ineligible abstract idea. Some responses argued that Federal Circuit eligibility law should be revised, e.g., to focus primarily on preemption, as in the Federal Circuit’s McRO case.
Areas of Disagreement
Commentators were more divided on other issues. For example, they split over the issue of whether existing forms of IP protection are sufficient for protecting AI inventions. Several responses opined about possible need for sui generis protection of certain data used in AI inventions, such as training databases or outputs. Others, however, suggested that AI can be adequately protected by existing forms of IP or technology, such as training databases being protected by trade secret law (or perhaps even copyright law if the requisite creativity is present), machine learning models being protected by technological protections such as DRM and machine learning model output being protected by trade secret or licensing law.
Relatedly, although commentators generally agreed that existing IP standards can be properly applied to AI in the same way that they applied to other technologies, there was some disagreement about how patent disclosure requirements should be implemented. Specifically, commentators differed on the degree of detail needed to establish written description and enablement under § 112 of the Patent Act, given the black-box nature of trained models. Some (including the American Bar Association’s Intellectual Property Law section and an individual commentator) opined that vigorous enforcement, including requiring disclosure of low-level details such as model weights, may be required; others, in contrast, were more concerned about the difficulties that such disclosure would pose in the case of technologies such as deep-learning systems.
Several commentators cited a need for weakening of IP protections in certain cases to foster competition. For example, the “large data collectors” that have access to huge training datasets due to their ability to read network data or other user data were cited as guarding access to those datasets for anticompetitive reasons, rather than reasons of user privacy, leading to a recommendation that there be “controlled sharing” of the datasets.
Considerations for Those Applying for AI-Related Patents
If the public comments are representative of the future state of the law, there will be little, if any, move toward the establishment of AI personhood for purposes of inventorship or ownership, nor will there likely be significant changes to the IP laws specifically to address AI issues. And as noted on Fenwick’s Bilski Blog, it seems unlikely that patent eligibility law will be altered any time soon, either judicially or legislatively. Regarding disclosure questions, applicants for AI-related patents will continue to need to weigh the risk of disclosing aspects such as parameter values that may ultimately be considered required for enablement and written description under § 112 against the option to keep such details as trade secrets.