November 5th, 2020
Artificial intelligence (“AI”) is an emerging technology field that presents unique challenges in Intellectual Property (“IP”) protection and enforcement. Recognizing these challenges, the United States Patent and Trademark Office (“USPTO”) issued a Request for Comments on Patenting Artificial Intelligence Inventions on August 27, 2019 and another Request for Comments on Intellectual Property Protection for Artificial Intelligence Innovation on October 30, 2019, which covered non-patent IP protections, such as trademark, copyright, and trade secret. Each Request garnered nearly a hundred comment submissions from various respondents, including foreign patent offices, bar associations, trade associations, companies, and law firms. On October 6, 2020, the USPTO issued a report entitled Public Views on Artificial Intelligence and Intellectual Property Policy (“the Report”), which summarizes the submitted comments from each Request and corresponding USPTO guidance.
In particular, the Report identifies various themes that emerged from reviewing the submitted comments. For example, many commenters noted that AI has no universally-recognized definition and therefore urged caution when proceeding with specific IP policymaking with respect to AI. Further, while commenters recognize three categories of AI inventions – inventions directed to advances in AI, inventions that apply AI, and inventions produced by AI – commenters noted that defining AI should be avoided, as the AI field is dynamic and may continue to undergo fundamental changes. In addition, commenters noted that the current state of the art is limited to “narrow” AI and that artificial general intelligence (“AGI”), akin to intelligence possessed by humans, is still only a theoretical possibility. As such, the majority of commenters opined that AI presently cannot invent nor author without human intervention. Thus, the majority of commenters agreed that, for now, existing US IP laws are sufficient to address IP for AI.
However, respondents did not find consensus for all issues. For example, some commenters voiced concerns regarding enablement of AI inventions, while others suggested new classes of IP rights for a more robust IP system for AI. Similarly, some commenters raised questions regarding how AI may impact the legal standard of a “person having ordinary skill in the art.” In addition, many commenters expressed concern whether the use of copyrighted materials to train AI would be infringing, while others stated that it should be allowed under fair use.
The Report explores a number of current issues regarding the patenting of AI inventions and provides insights which may be useful to consider during procurement or enforcement of AI patents. One such issue is whether there are any patent eligibility considerations unique to AI inventions. Commenters were split on this issue, with some asserting that AI inventions should not be treated differently from other computer-implemented inventions, which is the status quo for AI inventions currently before the USPTO. Other submissions indicated there may be increased risk of ineligibility, as AI inventions may be characterized as methods of organizing human activity, mental processes, or mathematical concepts.
Thus, as with other computer-implemented inventions, the Report underscores the importance of drafting patent claims which satisfy the test for patent eligibility under 35 U.S.C. § 101. In particular, patent applications directed to AI or machine learning (ML) subject matter should, when possible, include examples of how the invention can be incorporated into practical applications (e.g., specific use cases, data modalities, or tasks), stress any resultant improvements in computer or technical performance (e.g., reduction in transmitted/stored data or improved processing efficiencies), and set forth detailed descriptions of specific inputs/outputs and system or model architecture(s).
On the issue of the written description requirement under 35 U.S.C. § 112(a), most commenters indicated that AI inventions should not be subject to any unique requirements beyond those already in place. However, one response submitted by International Business Machines Corporation (IBM) noted that “AI inventions can be difficult to fully disclose because even though the input and output may be known by the inventor, the logic in between is in some respects unknown.” Report at pages 9-10. Perhaps in view of such difficulties, the USPTO-promulgated Report advised that “applications for AI inventions…should disclose the computer and the algorithm (e.g., detailed steps or procedures, formulas, diagrams, and/or flowcharts) that perform the claimed function in sufficient detail such that one of ordinary skill can reasonably conclude that the inventor possessed the claimed subject matter.” Report at page 9 (parenthesis in original).
With respect to the enablement requirement under 35 U.S.C. § 112(a), the Report reiterates that there is an inverse relationship between the level of predictability in the art and the amount of information that is required to be explicitly stated in the specification. Report at page 10 (citing MPEP § 2164.03). However, commenters provided inconsistent answers on the level of predictability associated with current AI systems. In particular, some comments asserted that AI systems were highly predictable, while others noted that AI systems include inherent randomness that make them unpredictable. As enablement can be fact-specific and there does not currently appear to be a general consensus regarding the level of ordinary skill in the art and predictability of AI systems, it may be beneficial to err on the side of caution when enabling the disclosure.
Thus, as with other computer-implemented inventions, to satisfy the written description and enablement requirements, it is important for a patent application to include descriptions of the hardware and the software involved, as well as the detailed algorithms for executing the invention. For example, patent applications directed to AI or ML subject matter should, when possible, include descriptions of the types, sources, or structures of data on which training or inference may be performed, the style or structure of the process used to perform model training (e.g., including description of any relevant loss functions or objective functions), and the types, architectures, or arrangements of models which may be trained or used in accordance with the invention. Similarly, the description should include any important downstream uses or applications of the AI or AI-generated outputs.
Another issue where responses were inconsistent related to implications of training AI using copyrighted data. In particular, typical ML processes for training AI include ingestion and processing of a large amount of data. As such, the Report also explored whether ingestion of copyrighted data constituted copyright infringement. The comments diverged on this issue, with some asserting that such ingestion should be considered fair use, citing, among others, Fox News Network, LLC v. TVEyes, Inc., 883 F.3d 169 (2d Cir. 2018). However, other respondents asserted that copyright holders should be compensated and suggested blanket licensing schemes. As recognized by the conflicting responses discussed in the Report, the implications of using copyrighted data when training AI or ML inventions are still unresolved. Therefore, those practicing AI and ML inventions should carefully consider the sources of the data they are leveraging and the potential consequences of using such data.
In sum, the comments in the Report generally expressed satisfaction with mechanisms in place at the USPTO for IP protection of AI inventions. However, because AI is a quickly developing field, many issues remain unresolved and still other issues are likely to arise over time. As such, practitioners and applicants should keep up-to-date on ongoing developments and best practices.