Two interwoven challenges come to mind when considering how to successfully patent AI technologies. The first of these challenges is drafting claims whose infringement is detectable despite the black box nature of AI technologies. The second challenge is drafting claims that are directed towards subject matter that the USPTO will find patentable (which, for the purposes of this article, covers both a determination of patent eligibility under 35 U.S.C. 101 and patentability under 35 U.S.C. 102, 103 and 112). These two challenges often are at odds with each other because claims are more likely to be patentable if they are directed towards specific implementations, but as claims become increasingly specific, it becomes increasingly difficult to detect whether a competitor has infringed the claims. This article delves into each of these challenges individually, and explores strategies that can improve a company’s chances of getting the USPTO to grant patents that can effectively be asserted against infringers.
Patentable subject matter satisfies the USPTO's main requirements for granting a patent: (1) that the claims are directed to a patent-eligible process, machine, manufacture, or composition of matter that does not recite a judicial exception, such as an abstract idea, or if they do, that the claims recite additional elements that integrate the exception into a practical application of the exception (35 U.S.C. 101); (2) that the claims are new and non-obvious relative to prior art (35 U.S.C. 102 and 103); and (3) that the claims meet certain requirements regarding enablement, possession of the invention, written support and clarity from the rest of the application. Given the "black box" nature of many AI based systems, extra care should be taken to address potential rejections under each of these requirements.
1. Detecting Infringement of Black Box AI Systems
AI systems can generally be described as including a “black box” in which input data is provided to an algorithm to generate an output, but the inner workings of the algorithm may not be readily apparent (or easily substitutable with similar alternatives). Given the black box nature of many AI systems, drafting claims that describe these inner workings may not be as valuable as claims that are directed to the use of the algorithm without specifying such details.
|Generally, detectability refers to how easily you can determine that someone is infringing the claims of your patent. Infringement can ideally be detected using only publicly available materials, including statements or documents made by the potential infringer, such as marketing statements, tech data sheets, white papers, or other information that is openly published. On the other hand, the more you have to rely on inferential reasoning -- or, later during litigation, discovery -- to confirm exactly how another party is performing certain parts of your claims, the more difficult (and thus expensive) it will be to assert and enforce your patent or use the patent as an effective deterrence tool.
As an example, a claim directed towards making personalized clothing recommendations using a model that was trained using prior purchase histories of consumers is likely to be more valuable than a claim that describes a specific type of classifier (for example, a Bayesian network, a neural network, SVM, etc.) that is used to make personalized clothing recommendations. The former is more valuable because most companies do not describe or publish details about their AI models but rather simply focus on the inputs and outputs of their AI models. For example, a company that has a system to recommend clothing to users based on the users’ past purchases may publicize that they use “machine learning” to provide recommendations more accurately targeted for what users are looking for, but may never have a need to actually explain how the machine learning is being implemented or what data (such as the user’s past purchases) is being used to train and operate underlying algorithms. As such, detecting infringement of a claim directed towards the use of an AI model is easier than a claim requiring specific details about the inner workings of the AI model.
Although it is always more desirable to pursue claims whose infringement is more readily detectable, in order to get such claims granted in the first place it is equally important to ensure that the claims are directed towards patentable subject matter, which brings us to the other biggest challenge in patenting AI technologies.
2. Drafting Claims Directed to Patentable Subject Matter
Strong patents, which can proceed easily through examination and resist validity challenges by opposing parties, tell a clear and compelling story about what makes your technology innovative. This result can often be achieved by explaining a technical problem (which may underlie, but not necessarily drive, a business problem or customer need) and how your innovation provides a technical solution to that problem. For example, if existing technologies are too slow for practical use, try to identify structural or functional deficiencies that result in the slowness, and then explain how your invention’s structures and functions enable improved speed.
With respect to AI technologies in general, consider explaining the improvements your technology provides, such as:
- the selection or training of the machine learning model results in improved speed or accuracy, or the ability of a computer to perform a function it could not previously perform;
- the generation or filtering of training data results in models that require fewer computing resources or increase processing speed; or
- the identification of salient parameters, features or thresholds that are more important to decision making than others, which improve the processing speed or reduce network latency of the AI technology.
Many AI based systems are designed to perform actions that humans previously performed. Claims directed to such AI systems may be determined by the USPTO as being directed to a mental process. To overcome such assertions by the USPTO, the application should explain how the AI based system may perform the action differently than a human and elaborate on some of the improvements the AI based system can achieve relative to humans.
One improvement which, at first glance, might seem easy to rely on, is that the speed at which the AI based system can process the data will be greater than if a human were to perform a similar analysis. However, such an improvement is likely not going to be successful if it merely implies that the AI is used for its conventional purposes (without adding details to your story about how the AI has been uniquely customized or the actions that the improved speed allow you to perform that could not be performed previously). Rather, it can be better to focus on improvements such as the elimination of subjectivity and the introduction of an objective decision making process based on certain rules or policies of the AI based system or describing multiple approaches in which the AI based system can perform the action and describing the pros and cons of each approach.
Providing information about the technical improvements an AI based system such as the ones described above in your patent application can persuasively demonstrate how your invention--and the engineering decisions along the development pathway to your invention--is innovative. This approach will not only help overcome any rejections under 35 U.S.C. 101 pertaining to patent eligible subject matter but will also likely help in overcoming art-based rejections under 35 U.S.C. 102 and 35 U.S.C. 103. It can be easy to fall into a trap in which your claims are repeatedly rejected under these standards if your application only talks about the business problems your invention solves and describes the end result of your invention without describing how the invention is implemented or the benefits enabled by your engineering decisions. By providing a compelling story about your AI based system’s improvements and how it achieves those improvements, you can increase the likelihood of streamlined examination, which will be less expensive and result in a stronger patent consistent with your overall IP strategy.