Intelligently Protecting AI: The Do's and Don'ts Under U.S. Patent Law
Artificial intelligence (AI) is a broad spectrum of technologies in both software and hardware. We provide a brief overview of patenting AI inventions in the U.S. patent system as well as practical advice on obtaining future protection of AI inventions.
The High Level Expert Group on AI established by the European Union Commission defines AI as systems designed by humans that “act in the physical or digital dimension by perceiving their environment through data acquisition, interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the information, derived from this data and deciding the best action(s) to take to achieve the given goal.”[1]
For example, a subset of AI is machine learning where a system can learn to solve complex problems, make predictions, or undertake tasks that require human-like sensing such as vision, speech, and touch.[2] Moreover, AI’s impact in commerce, daily life, and patent filings are becoming increasingly significant over this past decade. In this article, we provide a brief overview of patenting AI inventions within the U.S. patent system as well as practical advice on obtaining future protection of AI inventions.
Identifying the first AI patent is difficult, in part, due to a lack of a dedicated AI patent class at the United States Patent & Trademark Office (USPTO). The most commonly assigned classification for AI inventions is under Class G: Physics, specifically the subclass G06F.
However, recent work at the USPTO – conducted using AI methods – indicates that while the number of AI related patent applications prior to 2010 was insignificant, applications grew at a 30% annual growth rate between 2011 and 2015 and at 45% annual growth rate between 2015 and 2018, to reach a total of 16.5% of all applications in 2018.[3]
Within the broad scope of AI as single subcategory, machine learning inventions represent about 89% of all AI patent applications. Machine learning may include supervised or unsupervised learning, reinforcement learning, and various hybrid methods for training a machine learning model to perform a novel real-world application. This machine learning model may operate as a “black-box” where the underlying rules of the model may not be explicitly known, but are instead “learned” from training data. Possible patent claims for this group may seek to protect the preparation of training data, whether through special data acquisition, augmentation, or generation of synthetic training data. Another possible patent claim may include the actual training of a machine learning model using the training data.
Moreover, a novel machine learning architecture may be patentable through claiming the prediction of outputs based on a particular set of inputs and the hidden layers within the architecture. Finally, the use of a machine learning model in predicting data for real-world applications provides another claiming strategy.
In addition to method type claims, many AI inventions are implemented as software manifested as instructions in a computer readable medium or computer system. Thus, AI inventions may be claimed algorithmically or as a portion of a workflow. Similarly, AI inventions can also claimed as individual components or a special apparatus within a larger system especially when integrated into a real-world application. Sometimes, the AI component is simply another widget in a series of widgets for implementing a grand system. Where the other widgets are old, the AI component may be the novel, nonobvious, and nonroutine widget that presents the path to patentability.
AI patent applications are subject to the same novelty, non-obviousness, and claim definiteness requirements as patent applications in other technology arts. Similarly, AI patent specifications should also contain a written description of the invention “to enable any person skilled in the art” to make and use the invention.
However, AI patents may pose particular questions and issues. For example, the interdisciplinary nature of AI causes many inventions to be unique combinations of computer science, mathematics, and one or more domain-specific technologies, such as automotive engineering, logistics planning, or geoscience. Thus, it is reasonable to ask “which art” is the relevant art for an AI invention, the computer science art or the automotive engineering, logistics planning, or geoscience art? Furthermore, is AI a “predictable” art or an “unpredictable” one, where either type may have different legal standards for determining patentability? And, will a non-human entity ever be classified as an inventor under U.S. law?
While we can’t currently answer all legal questions that may arise in connection with AI patents, we can provide some useful observations and practice pointers for obtaining and enforcing some AI patents. For illustration, specifications on machine learning inventions should contain a sufficient description of the machine learning components selected by the inventors, such as particular activation functions, number and types of hidden layers, the input features obtained at the input layer, and any labels or other classifications predicted at an output layer.
Moreover, because the underlying models in machine learning continue to be updated, refined, and changed through the industry, it is especially important to “future proof” AI patent applications. Whether it is desired to protect a preexisting framework in a new useful application, or pursue a patent on one of the trending models in machine learning, a good machine learning description beyond the current methodology should be disclosed. The history of machine learning is full of colorful examples where dead-end technologies were revived through individual breakthroughs in other technologies. For example, the advent of graphical processing units (GPUs) for modern video gaming enabled the efficient training of convolutional neural networks.[4] Consequently, it is very important for patent applicants to sufficiently cover multiple alternative technologies, such as supervised algorithms, unsupervised algorithms, reinforcement learning algorithms, and hybrid architectures. Because the next big wave in AI may arrive very fast, patent applicants cannot claim what they do not include in their specification.
And no article on software inventions would be complete without at least some discussion of the patent eligibility problem. AI is typically a very respected technology in computer science and is thereby easily integrated into real-world applications for articulating technical improvements. Therefore, many “abstract idea” rejections are relatively straightforward for patent agents or patent attorneys. One approach is to claim the novel internal AI framework in connection with both hardware and software components disposed outside the AI framework. Likewise, the architecture of machine learning inventions may be claimed alone or in conjunction with their training processes. For patent eligibility, the USPTO contends that such a training process may fall outside of the mathematical concepts and mental processes groupings used for rejections under 35 U.S.C. 101. For example, in the Subject Matter Eligibility Examples published simultaneously with the January 2019 Patent Eligibility Guidance,[5] the USPTO included a patent eligible claim based on a training algorithm. As such, the USPTO indicated that training functionality falls outside of the mathematical concept and mental process groupings for rejecting claims for lacking patent eligible subject matter.
On the other hand, what should patent applicants avoid in their patent applications for AI inventions? Firstly, while mathematics is a highly respected discipline, it is best left out of patent claims. Claim drafters must walk a fine line between claiming general and functional concepts based on mathematical principles, on the one hand, and directly claiming the actual mathematical concepts, on the other hand. Because “the easiest patent rejection to overcome is the one that is never asserted,” equations should not be recited in either independent or dependent claims, generally speaking. Where certain techniques can only be defined using mathematics, such as various statistical analyses, the claim drafter should refrain from including such claim limitations. Likewise, it is very important to look at a claim through the eyes of a patent examiner who personally dislikes software inventions. Patent applicants can fine-tune claim terminology to make their claims look a little more like hardware and little less like software. For example, rather than the step of “send” data, the step might be drafted as “transmit a signal” even if that signal is just a “data” signal. There is little difference between “compute,” “calculate,” and “determine,” so the use of “determine” without its associated mathematical connotations might be advantageous in some cases.
Even if an AI invention has wonderful business advantages, it is not necessary to claim the invention as a “business method.” The difference between a good patent drafter and a mediocre one is that the good one will often do enough small things that the patent application begins in a very favorable position at the start of patent prosecution.
While obtaining an AI patent is good, being able to enforce an AI patent is even better. The traditional shortcoming of many software patents is that patent infringement is not always easy to detect by product outsiders. In particular, the “black-box” nature of machine learning may make it exceptionally difficult to identify what type of model is being used in a commercialized product. One past solution to detecting infringement in software was to obtain as broad patent coverage as possible for the independent claims, even if the claims might be susceptible to invalidity challenges later on. Another solution is knowing what software documentation is publicly available in a particular industry, whether in published FDA premarket approval submissions, press releases, or privacy policy disclosures. Knowing which types of AI inventions are “more” enforceable than others is a valuable consideration for directing future patent resources.
In conclusion, AI and machine learning inventions present unique challenges to both patentability and patent enforcement. While the rapid advancement of AI technologies in the past fifteen years has made this technology area exciting and popular, fast changing technologies also present special legal challenges in protecting such inventions. Accordingly, patent applicants should be aware that their inventions may become obsolete soon after filing if they fail to take adequate steps when drafting their patent applications. Similar to playing a good game of chess, inventors in the AI space always need to look at least three moves ahead at the changing patent landscape.
[1] Available at https://www.aepd.es/sites/default/files/2019-12/ai-definition.pdf.
[2] NIST (2019), 7-8. In their leading textbook Artificial Intelligence: A Modern Approach (Pearson, 2016), S. Russell and P. Norvig define AI broadly as the development of machines capable of undertaking human activities in four areas: thinking humanly, acting humanly, thinking rationally, and acting rationally.
[3] Giczy, Alexander & Pairolero, Nicholas & Toole, Andrew, “Identifying artificial intelligence (AI) invention: a novel AI patent dataset.” The Journal of Technology Transfer (2021) (DOI: 10.1007/s10961-021-09900-2).
[4] For example, while neural networks had been implemented since the 1970s, it wasn’t until a convolutional neural network, Alexnet, won several international competitions during 2011 and 2012 that the value of deep neural networks was clearly shown to the machine learning community.
[5] See “Example 39 - Method for Training a Neural Network for Facial Detection,” available at https://www.uspto.gov/sites/default/files/documents/101_examples_37to42_20190107.pdf.
