Claim drafting strategies for artificial intelligence innovations

Artificial intelligence is one of the fastest growing technologies in terms of sheer volume of patent filings at the United States Trademark and Patent Office. Between 2002 and 2018, annual AI patent applications increased by more than 100 percent.  This increase evidences the importance for applicants to develop a strategy for building a patent portfolio around their artificial intelligence technology.  In this article, we take a look at a few considerations when crafting claims for a patent filing which may help ensure that an AI patent application stands out.

In the wake of the Supreme Court’s 2014 decision in Alice, software-based innovations are subject to higher scrutiny during examination.  Because of this, it is prudent for practitioners to draft claims in a way that anticipates an eligibility challenge.  Depending on your innovation, a variety of strategies are available.

The latest eligibility guidelines from the USPTO limit the examiner’s ability to issue an eligibility challenge by specifying that the abstract idea must fall under one of several enumerated subject matter groupings in order to be rejected under §101.  These enumerated groupings include mathematical concepts, certain methods of organizing human activity, and mental processes.

One way to avoid the subject matter groupings is to describe a training process for the prediction model.  The USPTO has provided that such training process falls outside of the mathematical concepts and mental processes groupings.  For example, in the Subject Matter Eligibility Examples published simultaneously with the January 2019 Patent Eligibility Guidance, the USPTO included a machine learning based claim for consideration.  The USPTO provided that such training functionality falls outside of the mathematical concept and metal processes grouping.  While the training step may not be absolutely necessary to overcome an eligibility challenge, due to the low bar for identifying the abstract idea and the nature of an AI-based claim, omitting the training step could result in an examiner pigeonholing the claim into the mathematical concepts or mental processes subgroupings.

At times, including the training process may not be enough to avoid the subject matter groupings.  For example, the certain methods of organizing human activity grouping, an examiner typically considers the task for which the AI is optimized.  In such situations, the training process recited in the claim may not be sufficient to avoid the certain methods of organizing human activity grouping.  Instead, an applicant may need to go a step further and show that the claims include elements that integrate the abstract idea into a practical application.  To do so, it would be prudent to ensure that your claims capture the improvement as set forth by the specification.  For example, if your innovation is focused on a new architecture or combination of models, an applicant may include the particularly novel details about the new architecture in the claims.  The addition of the improvement, as described in the specification, into the claims may transform an ineligible claim into an eligible claim.

Once eligibility has been granted, it is also essential to consider enforceability of the claims.  While avoiding or overcoming an eligibility challenge is a noteworthy accomplishment, if it is achieved at the expense of enforceability, it could result in a very expensive piece of paper.  Applicants should make sure that the claims are drafted with a level of generality that covers the applicants’ invention and limits the opportunity for competitors to design around.

One of the strategies described above for avoiding an eligibility challenge involves incorporating a training process into the claims.  Depending on the owner of the patent, this strategy could result in issues down the line for patent holders.  In some circumstances, the party developing or training the AI innovation may be different from the party deploying the AI innovation.  In other circumstances, the training may be a highly manual process.  In both situations, applicants may have a claim requiring two separate actors, thus resulting in potential divided infringement issues.  One way to avoid this may be to cast the training process as a learning process.  For example, as a model is deployed, its use may result in the model’s continuous learning.  Additionally, even if the training is a highly manual process, it is the model that is learning from the training process.

With all these different strategies for claiming an AI innovation, applicants can build a diverse portfolio of filings that cover AI innovations from a variety of levels.  An applicant could draft a first claim set focused on use of a deployed AI model, a second claim set focusing on training the model for deployment, a third claim set focused on a combination of both processes, and a fourth claim set casting the training process as a learning operation, focusing on the point of view of the AI model.  An applicant may continue to build out the portfolio by changing the scope of the claims around each iteration.  This overall strategy may help guard against possible invalidity or non-infringement challenges when asserted.

As the number of filings on AI innovations continues to increase at the USPTO, developing a claim strategy that captures your innovation from a variety of angles can help applicants carve out a bigger space for themselves in an increasingly crowded technological field.