AI Visualize and the Eligibility of Innovative AI Systems

by Dennis Crouch

The recent eligibility decision in AI Visualize v. Nuance, __ F.4th __ (Fed. Cir. 2024), gives me pause to consider more general eligibility issues of AI Inventions. When does the design or creation of AI system elements qualify as an eligible invention?  In his recent article, Prof. Nikola Datzov wrote what we have all been thinking: “Innovative applications of AI are everywhere we look [and are] revolutionizing our society.”  Nikola L. Datzov, The Role of Patent (In)Eligibility in Promoting Artificial Intelligence Innovation, 92 UMKC L. REV. 1, 4 (2023).

In AI Visualize, the Federal Circuit sided with the accused infringer in finding the asserted claims ineligible under the two-step Alice framework.  AI Visualize had asserted four related patents that facilitated use of a low-bandwidth web portal for visualizing 3D/4D medical scans. The key here is to use virtual views and a system to determine which views have already been downloaded. Some claims require a unique identifiable key for each view; others use a tiered approach – first sending lower-quality frames for immediate viewing followed by higher-quality frames. U.S. Patent Nos. 8,701,167; 9,106,609; 9,438,667; and 10,930,397.  The district court dismissed the case void from the outset, finding asserted claims directed to patent-ineligible subject matter under 35 U.S.C. § 101. AI Visualize, Inc. v. Nuance Commc’ns, Inc., 610 F. Supp. 3d 638 (D. Del. 2022). On appeal, the Federal Circuit affirmed.

At step one, the court concluded the claims were directed to the abstract idea of “retrieving user-requested, remotely stored information.”  Although the claims reciting the creation of “on the fly” virtual views, the court reasoned that “the claim language makes clear that virtual view ‘creation’ is achieved by the manipulation of a portion of the existing [data set].”

At step two, the Federal Circuit agreed with the district court that “the asserted claims involved nothing more than the abstract idea itself” and conventional computer technology. The court found that the creation of a virtual view was an abstract idea that was also known in the art, as conceded in the patent specifications and later at oral arguments.  Therefore, that limitation could not supply the inventive concept required to transform the claims into patent-eligible subject matter.

This holding, focused on how the claim language defined the “creation” of virtual views, has interesting implications for the eligibility of AI systems, particularly generative AI (GenAI) technologies. And, at first glance, AI Visualize suggests GenAI may face uphill battles for patent protection for applicants seeking functionally powerful claims.

A key factor will likely be how the “generation” of the new AI content is defined in the claims. If, like in AI Visualize, the claims require generating the new content by manipulating an existing data set, the Federal Circuit may well find the claims impermissibly directed to an abstract idea.  The solution though is to recite a particular innovative technological solution for the creation of new content.  The closer the claims come to reciting a concrete technical solution for how the AI generates the content in a way that improves computer functionality, the better the prospects for eligibility.

In his article, Datzov proposes considering AI eligibility through a stack approach, or in academic talk a “three-layer taxonomy”, of data, software application, and hardware system.  As you might imagine, eligibility becomes as you delve deeper into the stack; although many computer hardware system claims have been deemed ineligible when they relied upon conventional computer components and where the true innovation was found in the functional data output.  In other words, at all three levels, the key for eligibility is to expressly claim how the technology works, rather than just its function or objective.  I have heard from patent attorneys that they are  more than ever sending invention disclosures back to the inventors for more development until the technological improvement becomes apparent.

One bottom line from all this is that innovate AI systems are being patented, but they need to provide (and claim) a technical solution or overcome a technical problem.  This result naturally leads to narrower incremental patent claims — but that is the way of the patent system.  We don’t get 11 million patents that are each an archetypal constellation.  Still, a second reality is that the current approach means that application layer developers are less likely to achieve patent eligible inventions, whereas those building the foundation models as well as those coordinating the linkage and layering of AI systems have a much greater chance.  Finally, we have to recognize that AI continues to be a buzzword without tight definitions — by some counts more than 10% of US patent applications are AI related.  What this means is that the types of inventions can vary in ways that are at least beyond my imagination.  The result then is that creative and skillful patent drafting is more important than ever.