Protecting your AI innovation: Strategies for startups (Second part)

For some AI startups, having a moat of patents will be the best way to protect their AI innovations.  But most AI startups are unaware that there are different types of patents for different purposes.  For example, there are design patents and utility patents.  Design patents protect the ornamental design of a product.  Design patents used together with trade secret protection provide a good combination of publicly available protection and secrecy.  Several famous Apple design patents are set forth below in Figure 3.

As shown in Figure 3, one of Apple’s most famous design patents, D604,305, covers the appearance of the iPhone’s GUI. AI startups should take note that design patents can be strategically employed to protect the ornamental appearance of their user interface or some other aspect of their product.    

In contrast to design patents, utility patents cover the novel and innovative functions of a device or a method, for example. However, the general rule in America is that software and AI are not patentable. Although there are many exceptions to this general rule, AI startups need to have a basic understanding of whether their inventions fit with one or more of the exceptions.  

First, a brief overview of what is patentable in America is in order. In America, the universe of patentable subject matter is very large, but there are 3 subject matter areas that are not patentable. These 3 areas are laws of nature, natural phenomena and abstract ideas, as shown below in Figure 4.  

Software, data structures and AI are considered abstract ideas, therefore, they are generally not patentable in America. However, all is not lost, as there are many exceptions to this rule that I will discuss below.    

In order to determine whether an AI innovation fits within an exception and is therefore patentable in America, the Patent Office and Courts employ a framework consisting of 2 main inquiries. This framework is set forth below in Figure 5, and it is called the Alice-Mayo test, for 2 famous cases decided by the U.S. Supreme Court.  

As shown in Figure 5, the Step 1 of the Alice-Mayo test analyzes whether the claims of a patent (or patent application) concern a law of nature, a natural phenomenon, or an abstract idea.  As most AI innovations will probably be characterized as abstract ideas, we move to Step 2. Step 2 of the Alice-Mayo test determines whether the claims contain additional elements that transform them from patent ineligible to patent eligible. Understanding the qualities and attributes of the additional elements that will transform a patent claim from patent ineligible to patent eligible is key to understanding whether an AI invention will be granted a patent in America. 

Usually, the “additional elements” identify a real world problem and provide a novel, technological solution to that problem. For example, the real world problem might be that a particular LLM is too slow, requires too much training data, consumes too much power, is too difficult to train, hallucinates frequently, etc. The novel, technological solution might be how to develop a particular type of synthetic training data, a novel, efficient arrangement of APIs that is faster or saves power, an easier or more accurate to way to train an LLM, or a way to detect or prevent an LLM from hallucinating. 

A recent case decided by the Court of Appeals for the Federal Circuit, provides a good example of one type of AI subject matter that is not patentable. The name of this case is Recentive Analytics, Inc. v. Fox Corp. and it was decided by the Federal Circuit on April 18, 2025. In this case, the patent owner Recentive sued Fox Corp. for infringing 4 of its patents which concern the use of AI to create an optimum television broadcast schedule for a live event.  

By way of illustration, one of Recentive’s patents-in-suit, U.S. Patent No. 11,386,367, claimed the following 4 steps:  (i) collecting data concerning a live event to be televised; (ii) training a generic, machine learning model to recognize relationships in the collected data; (iii) using a machine learning model to generate an optimized television schedule for the live event; and (iv) detecting changes to live event data to generate a revised, optimized television schedule for the live event. These steps are illustrated in the block diagram set forth in Figure 6 below. 

Recentive argued that its AI inventions were patentable because they dynamically adjusted the training of its machine learning model to generate a revised, optimum TV schedule. In other words, Recentive asserted that its patents-in-suit pass Step 2 of the Alice – Mayo test because its dynamically adjusted training is a novel, additional element that transforms an otherwise abstract idea (using AI to generate a television schedule) into a patentable invention. See Fig. 5, Step 2, supra.  

The Federal Circuit disagreed, explaining that “dynamic adjustments based on real-time changes are incident to the very nature of machine learning.” Recentive Analytics, slip op. at 12 (citation omitted). Therefore, Federal Circuit ruled that “[a]n abstract idea does not become nonabstract by limiting the invention to a particular field of use”. Recentive Analytics, slip op. at 14 (citation omitted). The holding of the Recentive case is summarized in Figure 7, below.     

Next, I will provide a positive example of the type of AI invention that is patentable in America. As many people know, the IBM Watson Project broke many barriers in the AI field, so it should not be surprising that this patent covered one of many inventions that came out of the Watson Project. IBM Watson’s U.S. Patent No. 11,475,331 (hereafter, “the ’331 Patent”) concerns the problem of bias in dataset, which can skew results and compromise accuracy. The invention described in the ’331 Patent detects bias in a dataset, identifies the biased data entrees, and surgically removes the biased data entrees to create an unbiased dataset. A block diagram illustrating the ’331 Patent’s bias detection and removal is set forth below in Figure 8.  

As shown in Figure 8 supra, Bias Detection Tool 310 detects bias in Input Dataset 305.  If bias is identified, the Source of Bias Identification Tool 320 locates the biased data entrees.  Finally, the De-Biasing Engine 340 removes the biased data entrees from Input Dataset 305, to create De-Biased Dataset 380.   A simplified summary of the invention claimed in the ’331 Patent is set forth below in Figure 9. 

As shown in Figure 9, the ’331 Patent easily passes the Step 2 test of the Alice-Mayo test (depicted in Figure 5, supra) mainly because it provides a novel, technological solution to a real world problem: bias in a dataset.  

Hopefully, this example (and the previous, counter-example) provides some guidance to AI startups on the types of AI innovations that are patentable in America. As I explained above, disclosing your company’s valuable ideas by filing patent applications is not always the correct choice. Please carefully consider trade secret protection, design patents and utility patents before deciding on IP strategy for your AI startup.

AI startups, particularly AI startups in Japan, often have great ideas, boundless energy and eternal optimism, yet they often fail to consider how best to protect their IP. It is my sincere hope that this paper causes some AI startups to give serious consideration to how best to protect their ideas and thereby advance their business.  Spending a modest amount of time and money to protect their innovations will undoubtedly benefit their company in the future.  

Eric D. Kirsch is a partner with Rimon, P.C. and a permanent resident of Japan.  Eric was a successful IP litigation partner in New York City before moving to Japan in 2010 to become Nikon’s Chief Intellectual Property Counsel.  In 2023, Eric joined Rimon and opened Rimon’s Tokyo Office.  If you have any questions about this article or a U.S. legal issue, Eric can be reached at eric.kirsch@rimonlaw.com

Top photo: Envato

For inquiries regarding this article, please contact jstories@pacificbridge.jp

Click here for the Japanese version of the article