Drafting AI Patents: Challenges and Solutions

By Aastha Uppal
August 21, 2021

“Realizing that every approach has certain drawbacks and benefits, one should adopt a strategy to protect the AI invention in light of future planned research and the business strategy of the assignee and/or inventors.”

https://depositphotos.com/207844888/stock-photo-digital-artificial-intelligence-text-hologram.htmlArtificial Intelligence (AI) is the latest buzzword across all sectors. Every tech and non-tech company is vigorously filing, strategizing or planning to enter the AI patent domain. However, the journey is not as easy as it may seem. While drafting AI-based patent applications, drafters often face challenges in formulating the right strategy for writing claims and identifying the correct scope of the application. Thus, it’s important to know the challenges in detail and to develop practical solutions for drafting a patent-worthy application.

Before moving on to the challenges, let us first understand the key aspects of AI-based inventions or patent applications. It is imperative to identify the correct combination of these aspects so that an invention is suitably protected. Not all AI patent applications emphasize all three aspects, although they might be indirectly related. The three aspects are:

  • Algorithm – An AI related model (or algorithm) carries out intelligent processing.
  • Training Methodology – A training methodology and its underlying training dataset enables the “intelligence” in the AI related model.
  • An application or product – Uses trained AI model for carrying out a real-world or “tangible” phenomena with improved efficiency.
[[Advertisement]]

Challenges Related to Drafting AI Patent Applications

While drafting, it is vital to identify which of the three aspects should be covered solely or in combination when seeking legal protection. Depending on the approach chosen, the patent drafter may face one or more of the three main challenges:

  • Patent Ineligibility: The invention or process stands the risk of being considered as abstract and patent-ineligible. This implies that it may get classified as an “abstract idea” by virtue of implementing a series of steps or operations that could possibly have been carried out in “mind” or by using “pen and paper”. For example, complex mathematical calculations.
  • Divided Infringement: Divided infringement refers to the issue wherein multiple parties are required to infringe on all the clauses mentioned in an independent claim. Thereby, it makes the invention or the process less profitable for the assignee.
  • Infringement Detectability: Detectability refers to the ease with which infringement of a patent by a product or service can be detected. Difficulty in detecting the infringement lowers commercialization potential of the underlying patent.

Understanding Challenges with the Help of Case Studies

Case Study 1: Assuming an invention pertains to the training of a neural network and novelty lies in the way it is trained.

In this scenario, from a novelty perspective, a claim covering the training procedure should be framed. However, this approach has few shortcomings that should be considered:

  1. Claim infringement:

Conventionally, the training of the neural network is done offline in research labs, not in real-time. For example, let’s assume a neural network is trained for picking up an object and is further implemented on a robot. In this case, what one can see in real-time is the robot picking up an object. Here, the task performed is the result of the invention, not the invention itself.

So, one can’t detect infringement because the claimed training is generally done offline. Therefore, the difficulty in detecting infringement makes it hard to file a lawsuit.

  1. Patent-eligible subject matter:

The claim plainly covering the training procedure is not considered to be patent-eligible subject matter. Such a claim is highly prone to 35 U.S.C. § 101 rejection in the United States. The United States Patent and Trademark Office (USPTO) may assert that the claim is directed towards an abstract idea.

In order to mitigate such problems, one must frame the claim in accordance with the USPTO’s example 39. Similarly, in specifications, one must meticulously describe how the claimed training procedure solves problems related to training of the neural network. For example, it requires less training data, can be trained with different data (e.g., labelled and unlabelled), and allows the neural network to be trained faster. Lastly, describe the practical application of the trained neural network, i.e., showcase its utility.

While the aforesaid solutions may solve the patent-eligible subject matter problem, the shortcoming in terms of claim infringement persists. This is the drawback of this case.

Case Study 2: Let’s assume that the invention related to training of the neural network and execution of the neural network for a specific application is claimed.

Using this approach has certain advantages. First, the utility of the neural network is directly outlined in the claim, which might make it patent-eligible subject matter. Second, when the applicant claims the execution of the neural network, the infringement detectability is improved. This is because it can be seen in real-time. However, there are certain drawbacks of this approach as well:

  1. Inactive recitation of active steps of invention

Generally, in such a claim, the training procedure is claimed using “wherein clauses”. This is because the training is done offline and the execution takes place in real-time. For example, “wherein the neural network is trained based on [enter training aspects].”

It is evident that the training process, where the novelty lies, is not claimed actively. Furthermore, claiming inactively yields the impression that there already exists a pre-trained neural network that is now being used to execute a particular workflow. As a result, the claim interpretation may incline towards the way of executing the neural network rather than the way of training the neural network.

  1. Difficult infringement enforcement due to divided infringement

If one claims execution of the neural network, along with actively claiming the training aspect, it may lead to divided infringement. For instance, let’s assume that the neural network is trained by company X, and it is integrated with a robot manufactured by company Y. This implies that the robot manufactured by company Y is receiving and processing the input with the neural network produced by company X. So, it results in divided infringement. The involvement of two entities makes the enforcement of infringement difficult and expensive.

Case Study 3: Let’s assume that the invention is related to execution of a neural network and a claim covering proposed execution of neural network is framed.

Take, for example, that the invention may be related to a particular way of using a known neural network (e.g., autoencoder). It can be used for a specific application or modification to the structure of neural network. Furthermore, it can be utilized for transforming input data to a new form, such that the transformed data can be used with known neural networks for a specific application. In this case, a claim that covers proposed execution of the neural network can be framed.

The purpose behind using such a claim is that it clearly outlines practical application and is less susceptible to a Section 101 rejection. Also, the claim includes aspects that are detectable in real-time, which resolves the infringement detection problem.

However, this claim may also suffer from the limitation of being too narrow, as it is focused on a specific application using a specific type of neural network only.

Final Thoughts

There is no one-size-fits-all solution when it comes to drafting claims of a patent application in the AI realm. A practitioner must weigh the pros and cons of each approach outlined above, and then choose the right one, keeping in mind their client’s strategy, budget and core competencies.

Realizing that every approach has certain drawbacks and benefits, one should adopt a strategy to protect the invention in light of future planned research and the business strategy of the assignee and/or inventors. Furthermore, always find where the novelty lies and formulate a strategy to claim the invention accordingly. Lastly, no matter what kind of AI invention it is, always describe the practical application of the invention in claims and specifications. In addition to that, describe a technology-centric solution to a problem or technical improvements provided by the invention to make it patent-eligible subject matter.

Image Source: Deposit Photos
Author: sdecoret
Image ID: 207844888 

The Author

Aastha Uppal

Aastha Uppal is a Consultant, ICT Drafting and Prosecution, with Sagacious IP. Aastha is an IP professional with 12+ years of experience in the IPR domain. Her work profile spans across multiple IP related services including Patentability Assessment Searches, Technology Landscapes, Infringement/ Overlap Studies, Patent Drafting and Technical Content Curation and Data Scraping and Analysis.

Warning & Disclaimer: The pages, articles and comments on IPWatchdog.com do not constitute legal advice, nor do they create any attorney-client relationship. The articles published express the personal opinion and views of the author as of the time of publication and should not be attributed to the author’s employer, clients or the sponsors of IPWatchdog.com. Read more.

Discuss this

There are currently 3 Comments comments. Join the discussion.

  1. Anon August 21, 2021 8:35 pm

    I am curious to see how George feels about this subject.

    🙂

  2. Mark Nowotarski August 22, 2021 7:17 pm

    Great summary of the critical issues. Effective AI patents are indeed a challenge.

  3. ipguy August 24, 2021 2:03 pm

    Does the latest AIPLA Economic Survey indicate how much is being charged for drafting an AI patent application?

Post a Comment

Respectfully add to the discussion.

Name *
Email *
Website