Drafting claims with infringement in mind has always been a challenge. For instance, claims should be drafted to ensure that they can be infringed by a single party in order to address divided infringement issues. Similarly, it may be useful to draft claims in a way that avoids requiring end-user infringement as end-users may not be the best target when considering litigation. In the same vein, it is important to draft claims in a way that infringement is detectable, as a patent owner must have facts that provide a plausible entitlement to relief. This means that the patent owner must have some basis to allege that the patent claims are being infringed. If the claims include features that are not readably detectable, the claims may, in effect, be useless when considering infringement. This is especially true in Artificial Intelligence where many features are difficult to easily detect without intimate knowledge of the AI system. In fact, AI systems are often considered black box systems in which the inner workings are not evident to the outside and sometimes not even to the operator of the AI system.
This characteristic of AI systems should be carefully considered when drafting and prosecuting AI patent applications. For example, it is often more useful to draft claims that are directed to the inputs and outputs of a system then to explicitly recite the specific type of machine learning algorithm used in the system. This of course sounds easier in principle than it is in practice as broad claims directed to AI systems often run afoul of §101 or are simply rejected based on prior art. Nevertheless, considering how AI systems will be discussed in promotional material or in publicly available materials is a very important part of generating claims that are powerful when considering infringement.
When considering these issues, it may be helpful to look at an example of an AI invention directed to a system for detecting images of animals, which includes a convolutional neural network and a large, tagged dataset of different animals. Even if this exemplary system includes novel aspects directed to the operation of the convolutional neural network, a claim directed to this invention should be careful to avoid too narrowly reciting these features in the independent claims because it may be difficult to detect such features. Instead, the claim should attempt to focus on features which would more likely be advertised or discussed publicly by a potential infringer. For instance, the infringer may describe that the infringing system is able to detect different images of animals even when the animals are captured from different angles. These functional features of the convolutional neural network could be claimed in place of explicitly claiming the underlying algorithm. For instance, the claim could recite a classifier configured to apply a neural network that compensates for the angle at which the image is captured. By focusing on features that are easily detectable the claims can ensure that the claims remain a powerful tool against potential infringers.Although there are many issues that make drafting AI claims challenging, it is important not to overlook detection of infringement as this issue could become very important if the claims are ever to be asserted.