Navigating a Patent Gap for AI and Machine Learning Algorithms

The Bottom Line

• An April Federal Circuit decision offers an initial perspective into the benchmarks for successful artificial intelligence and machine learning algorithm patent claims.
• Technological improvements appear to be an important factor for a successful patent claim, though it is still unclear whether improvements in machine learning on their own suffice.
• Claims that focus on how an algorithm can be applied, beyond what improvements it offers in machine learning, are more likely to be successful given the current state of the law.

Patent applications on artificial intelligence and machine learning have soared in recent years, yet legal guidance on the patentability of AI and machine learning algorithms remains scarce.

The US Court of Appeals for the Federal Circuit’s first significant AI ruling in Recentive Analytics v. Fox and the US Patent and Trademark Office’s most recent guidance on eligibility of machine learning models still leave a critical question unanswered: Can an invention be patent eligible if it uses an improved (that is, non-generic) machine learning model but makes no other technical contributions?

Patent eligibility of software in general has been in flux since the US Supreme Court’s Alice Corp. Pty. v. CLS Bank Int’l opinion in 2014. In Alice, the court explained that the use of a “generic computer” doesn’t transform an otherwise abstract or non-technical idea (such as a business method) into a patent-eligible one.

At that time, some practitioners feared this was the end of all software patents. But in 2016, the Federal Circuit explained in Enfish, LLC v. Microsoft Corp. that software can still be patent eligible if it’s used to improve computer functionality. In Enfish, the software was found to have improved how “a computer stores and retrieves data in memory” with a new “data structure.”

That same year, the court’s decision in McRO, Inc. v. Bandai Namco Games Am. Inc., explained that software also can be patent eligible if it provides an improvement in another technical field, such as “3–D computer generated lip-synchronization” in computer animation. This line of case law instructed that claims providing “a specific, technological solution to a technological problem” are eligible for patenting.

But where does machine learning fit into this patent eligibility landscape? Will the Federal Circuit simply categorize it as conventional software, or recognize it as having a distinct nature?

Machine learning models inherently are more technically advanced than conventional software programs. Unlike conventional software programs, which include a predetermined set of instructions based on a given set of inputs, machine learning algorithms use training data and iterative optimization steps to improve the model’s ability to respond to a given set of inputs.

A human can’t trivially predict what a machine learning model will output with a given set of inputs. The technologically complex nature of machine learning algorithms raises a critical question: What threshold of technical improvement must such algorithms cross to be patent eligible?

AI Guidance

In July 2024, before the Federal Circuit’s Recentive decision, the PTO issued patent eligibility guidance for AI inventions based on existing software jurisprudence, and provided examples of what would be a patent eligible use of machine learning:

• One example illustrated that patent applicants can’t claim the use of an artificial neural network (a type of machine learning model) for general detection of data set anomalies. But, according to the PTO, applicants can get a claim for using an artificial neural network to detect and block anomalies in network traffic (preventing malicious attacks).
• In another example, the general use of a deep neural network to solve a mathematical formula related to speech analysis wasn’t deemed eligible. But using the same machine learning algorithms to improve speech analysis by separating desired speech signals from extraneous or background speech was.

Considering the Federal Circuit’s software eligibility cases available at that time, the guidance provided nothing surprising. Consistent with cases such as McRo, the PTO found patent claims eligible if the machine learning algorithm provides a technical improvement in another technological field—network traffic and speech signal analysis, for example.

The guidance is agnostic to innovation in machine learning. It says nothing about whether an improvement in the performance or structure of a machine learning model would result in an eligible claim.

Recentive Analytics

Despite the PTO’s guidance seemingly to the contrary, patent applicants continued to file claims for non-technical uses of machine learning models. The Recentive patent was one such example—using known machine learning techniques to optimize advertising revenue for live event and television scheduling.

At oral argument, Recentive focused not on revenue optimization but on the complexity of machine learning itself. It emphasized that a computer executing a machine learning model is a “special purpose computer” that finds connections between inputs that a human wouldn’t.

The Federal Circuit wasn’t convinced: Claims “are not rendered patent eligible by the fact that (using existing machine learning technology) they perform a task previously undertaken by humans with greater speed and efficiency than could previously be achieved.”

The court found the claims ineligible because they did “no more than claim the application of generic machine learning to” tasks that have “existed for a long time.” The outcome therefore closely tracked Alice—“stating an abstract idea,” such as optimizing advertising revenue, “while adding the words ‘apply it with a computer’” isn’t enough.

Applying a generic (non-innovative) machine learning model to a well-known task didn’t meet the Alice threshold.

Though closely tracking Alice in its holding, the Federal Circuit’s focus on the machine learning innovation—or lack thereof—in the Recentive patent eligibility inquiry was nevertheless notable. The court was keenly interested in the technical improvement that the machine-learning algorithm itself could have provided.

Multiple times, the court juxtaposed the idea of a technological improvement with the machine learning model. For example, the court explained that the “requirements that the machine learning model be ‘iteratively trained’ or dynamically adjusted in the … patents do not represent a technological improvement.”

The court also was concerned that the patents “do not claim a specific method for ‘improving the mathematical algorithm or making machine learning better’” and thus “the claims do not delineate steps through which the machine learning technology achieves an improvement.”

This language in the opinion seems to indicate that improvement in a machine learning algorithm itself could be the type of technological improvement that would render the patent eligible. But neither the Recentive opinion nor PTO’s earlier AI guidance explicitly discuss what types of improvement or innovation in machine learning algorithms could rescue an otherwise an ineligible abstract idea.

Improved Model Gap

There appears to be a critical gap in current machine learning eligibility guidance about what kind of improvement would constitute a sufficient “technical improvement.”

Could someone claim an improved machine learning model that didn’t necessarily have any corresponding improvement in overall computer (or other technical) functionality?

For example, is it enough for patentability if a machine learning model used an architecture or training algorithm (such as an optimization method or loss function) that was theoretically or empirically linked to improvements in the model’s performance metric (in accuracy, training or inference efficiency, precision or recall, robustness against hallucination, and the like)? Or would the model’s improved performance need to be reflected in the improved performance of the computer itself to attain eligibility?

Best Practices

Given the ambiguity that persists after Recentive Analytics and the PTO’s recent AI guidance, practitioners should adopt strategic approaches when drafting machine-learning-related patent applications. There are several practices that can maximize the likelihood of patent eligibility while navigating these uncertain waters.

Draft claims strategically by focusing on a technical application. To the extent possible, patent practitioners should emphasize the “technical improvement” in their claims by reciting specific downstream technical applications of improved machine learning models.

An approach that may be prudent in many cases is to focus the independent claims on the core machine learning innovation, with dependent claims capturing some important applications. This preserves broader protection for the core innovation in the independent claim, while including dependent claims that are more likely to survive eligibility challenges under current eligibility law.

Describe machine learning performance improvements. Both courts and examiners demonstrate clear preference for detailed technical descriptions over generic improvement pronouncements. The specification should provide a comprehensive description of how the claimed innovation improves machine-learning-related performance metrics.

The specification also should establish a clear nexus, using more than broad and conclusory statements, between the machine learning algorithm’s innovative features and the machine-learning-related performance improvements.

Connect machine learning improvements to broader technical advancements. Machine learning improvements frequently translate into secondary technical benefits beyond the system itself. These connections can strengthen eligibility arguments by linking the machine learning innovation to established technical domains with clearer eligibility precedents.

The specification should articulate, in sufficient detail, how the AI improvements contribute to technical improvements, such as improvements in computational efficiency, network security, data integrity, operation of hardware systems, or technical problems such as image processing and signal analysis.

Maintain pendency of patent families. An applicant should consider keeping the application’s family pending through continuation applications to benefit from potential favorable changes in the law.

For example, an applicant forced to narrow claims by including specific downstream applications may later be able to secure broader protection for the core machine learning innovation if the law evolves to recognize machine learning improvements as inherently technical.

This approach requires careful attention to written description support for broader claims. The specification should comprehensively describe the machine learning innovation independent of its applications, while also detailing specific implementations to support narrower claims if necessary.

While the Federal Circuit’s Recentive decision suggests that advancements in a machine learning algorithm may be sufficient to render a claim patent eligible, it does not provide a definitive answer to this question.The PTO’s most recent guidance on AI is silent on this issue.

In light of this lack of clarity, practitioners should try to highlight both improvements in the AI model itself and any associated technical improvements in computer functionality (or another technical field). Such an approach increases the likelihood of allowance and benefits from any potential changes in the law—for which Recentive provided only hints.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.

Author Information

Lidiya Mishchenko, Ph.D., is special counsel at MoloLamken, where she focuses on intellectual property and technology-related litigation, with an emphasis on appeals.

Pooya Shoghi is of counsel and head of the AI committee at Lee & Hayes, where he specializes in patent prosecution and strategic counseling for software and electronic technologies.

Write for Us: Author Guidelines