Subject Matter Eligibility of Artificial Intelligence and the Problems with a Functionally Drafted Patent

Michael Kiklis

Introduction

Although Artificial Intelligence has been the focus of research since the 1950s,[1] only recently has it become a mainstream topic, discussed and utilized by many.  In fact, no one can doubt that AI has become the big hope for future economic gains.[2] 

Artificial Intelligence

There are many definitions for AI, but most include simulating human activity (such as robotics) and providing decision support (such as suggesting solutions to network connectivity issues).  This article focuses on the patentability of decision support systems, which include machine learning (ML).  Decision support systems usually analyze data, which could be a static data set or dynamically derived data, like network activity.  They then provide recommendations by, for example, displaying them or even automatically implementing the recommended changes.  Neural networks are a common tool for use in such systems.

Patentable Subject Matter Overview

The statute governing patentable subject matter is very broad, for it states “Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.”[3]  The Supreme Court has recognized several judicial exceptions that limit patent eligibility:  “Excluded from such patent protection are laws of nature, natural phenomena, and abstract ideas.”[4]  The rationale for these exclusions is that “these exceptions are ‘part of the storehouse of knowledge of all men … free to all men and reserved exclusively to none.’”[5]  When and how these exceptions are applied are perhaps the most challenging part of a § 101 analysis.

Patentability of Computer-Related Inventions

The Supreme Court first dealt with computer-related inventions in a trilogy of cases.  First, in Benson, the Supreme Court struck down a binary-coded decimal to pure binary conversion program that could be performed by a human and that was “so abstract and sweeping as to cover both known and unknown uses” of the conversion, ultimately leaving it up to Congress to decide whether to patent these inventions.[6]  Next, Flook dealt with a method for updating alarm limits using a mathematical formula and again struck down the invention because the only difference “between the conventional methods of changing alarm limits and that described in respondent’s application rests in the second step—the mathematical algorithm or formula.”[7]  In this “point-of-novelty” test, the formula (a law of nature) is ignored and the rest of the claim must be novel.[8]  The last in the trilogy of cases was Diehr, where the Court rejected the point-of-novelty test, stating that “The ‘novelty’ of any element or steps in a process, or even the process itself, is of no relevance” in a § 101 inquiry.[9]  Then, decades later, the Supreme Court in Mayo seemingly resurrected the point-of-novelty test but left Diehr intact.[10]  This is the cause of much difficulty today.[11] 

The current § 101 analysis—as explained by the Supreme Court in Alice—is a two-part test:  (1) determine if the claims are directed to one of the judicial exclusions, such as an abstract idea; and if so, (2) “consider the elements of each claim both individually and as an ordered combination to determine whether the additional elements transform the claim into a patent-eligible application” by reciting an inventive concept, which must be more than a generic computer.[12]  This test caused much concern for software patents because the novelty in those inventions is found in the processing of the software itself, rather than in the hardware.  Luckily, in Enfish, the Federal Circuit clarified that software is not “inherently abstract” and can be patent eligible.[13] 

Patent Eligibility of AI Inventions

As the Federal Circuit’s case law evolved after Alice, it generated yet another hurdle for decision-support systems in the Electric Power Group (EPG) line of cases.[14]  In EPG, the Federal Circuit held that collecting information, analyzing that information, and presenting the results is an abstract idea.[15]  This means that many decision support systems may fail step one in the § 101 analysis, requiring that such systems recite an inventive concept—more than a generic computer—to be patent eligible.  EPG is therefore dangerous for AI inventions. 

EPG has become one of the Federal Circuit’s most often called upon weapons to find patents ineligible.  In fact, the Federal Circuit has used EPG at least six times this year alone to kill patents.[16]  All hope for AI inventions, however, is not lost. 

The courts have found patent eligibility in a number of situations potentially applicable to AI inventions.  First, both the Supreme Court and the Federal Circuit have found patent eligibility where the invention improves the functioning of the computer itself.[17]  Second, the Federal Circuit has found patent eligibility when the claimed invention provides a technical improvement over the prior art.[18]  Third, the Federal Circuit has found patent eligibility where the claims recite a technical solution to a technical problem,[19] and fourth, the Federal Circuit has found patent eligibility when conventional elements are arranged in a non-conventional way.[20]  Should an AI invention fit into one of these categories, it will likely be found patent eligible as long as those features are both described in the patent specification and claimed.

There are other concerns beyond the Electric Power Group line of cases for AI inventions.  For example, the Federal Circuit is very willing to strike down functionally drafted claims regardless of their subject matter.[21]  Also, the Federal Circuit has found no business method patents after Alice satisfy § 101.

Recentive Case Study

In Recentive Analytics,[22] another EPG case, the Federal Circuit has now applied these principles to strike down claims directed to machine learning (ML).  The patents involve scheduling of live events (kind of a business method or at least an age-old human activity), and the ML is both generically and functionally described and claimed.  In fact, the specification describes the ML as a black box:  “[T]he performer model 116 includes or utilizes a machine learning model 210 or other predictive tool …. Any suitable machine learning technique can be used, such as, for example….”[23]  The specification generically and functionally describes the ML model’s operation as training, receiving input parameters, and then scheduling the live events, just about as high level as possible and not specific to any ML model.[24]  What few technical steps exist are all described functionally, without describing “how” anything is done.  Indeed, each step is described as a mere possibility by using the verb “can.” 

At one point, the specification claims to achieve “a significant improvement in computer functionality” through improved accuracy, automation, and being able to automatically update the schedule based on changes to the input parameters.[25]  However, Recentive had to admit that its functionally drafted patents “do not claim a specific method for ‘improving the mathematical algorithm or making machine learning better.’”[26]  The court noted that even if Recentive had not made this admission, “neither the claims nor the specifications describe how such an improvement was accomplished.”[27]  Ultimately, the court held that the patents are not eligible simply by applying machine learning to a new field of use.[28]  In light of the Supreme Court’s Alice decision (2014) and the Federal Circuit’s EPG decision (2016), this is not surprising.  In fact, it is to be expected.

Although the result is not surprising, the court could be seen as possibly raising the bar on software inventions.  In its opinion, the court states:

In the context of software patents (which includes machine learning patents), the step one inquiry determines ‘whether the claims focus on “the specific asserted improvement in computer capabilities … or, instead, on a process that qualifies as an abstract idea for which computers are invoked merely as a tool.”’[29] 

This statement seems rather strong in view of the Federal Circuit’s case law:  either the claimed software improves the computer capabilities or else it is ineligible.  Certainly, there are software patent cases where specific details or a claimed solution necessarily rooted in computer technology were enough, e.g., DDR Holdings.[30]  Indeed, a review of the case to which the court cites shows the Recentive court did perhaps move the bar a bit.  There, the court made no such ultimatum:

In cases involving software innovations, this inquiry often turns on whether the claims focus on “the specific asserted improvement in computer capabilities ... or, instead, on a process that qualifies as an abstract idea for which computers are invoked merely as a tool.”[31]

It is unfortunate that the Federal Circuit’s first patentable subject matter test for AI involved these facts.  But, the outcome of Recentive is not surprising and fits comfortably within the Federal Circuit’s § 101 caselaw. 

Another Trap for Functionally Drafted Patents

Yet another invalidity trap awaits functionally drafted patents:  35 U.S.C. § 112(f).  This section of the Patent Act governs functionally drafted claims:

(f) Element in Claim for a Combination.—

An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.[32]

Of course, a patent drafter can intentionally invoke this provision by using “means for” or “step for” language, but even in the absence of such language, a court may still construe a claim as invoking § 112(f).  In that case, if the specification does not disclose the corresponding structure, material, or acts, the claim is invalid.[33] 

Fintiv is instructive.  There, the court was construing the term “payment handler service.”[34]  The court acknowledged the rebuttable presumption that § 112(f) does not apply because the word “means” did not appear, but nevertheless found the presumption was overcome because “the payment-handling terms recite function without reciting sufficient structure for performing that function.”[35]  In dismissing structural arguments based on extrinsic evidence, the court stated, “[t]he determination of whether the payment-handler terms invoke § 112 ¶ 6 turns on how those terms are used in the asserted patents.”[36]  Having found that § 112(f) was invoked, the court performed the second step of the analysis, “identifying the corresponding structure defined in the specification.”[37]  Importantly, the court noted the linking requirement:

Structure disclosed in the specification qualifies as corresponding structure if the intrinsic evidence clearly links or associates that structure to the function recited in the claim.[38]

The court cautioned that the specification’s disclosure of corresponding structure must be adequate to achieve the claimed function.[39]  The court then provided the general requirement for software patents:

If the function is performed by a general-purpose computer or microprocessor, then the second step generally further requires that the specification disclose the algorithm that the computer performs to accomplish that function.[40] 

The court then explained the invalidity trap, “the claims are indefinite if the specification fails to disclose adequate corresponding structure to perform the claimed function.”[41]   

Appellant’s problem, and the reason why the claims were found invalid, was that the specification merely provided the same level of disclosure as the claim term, which was insufficient:

Fintiv's purported two-step algorithm merely recites the asserted claims’ language. The specifications do not provide additional disclosures other than reciting the function of the payment-handler terms using generic terms without providing any details about an algorithm to carry out the functions of using APIs of different payment processors including one or more APIs of banks, credit and debit cards processors, bill payment processors, and exposing a common API for interacting with different payment processors.  We have held that describing “the results of the operation of an unspecified algorithm” is not sufficient to transform the disclosure of a general-purpose computer into the disclosure of sufficient structure to satisfy § 112 ¶ 6.[42]

Fintiv and similar cases demonstrate that practitioners should proceed with caution by drafting full, detailed specifications.  Also, avoid claiming software claim limitations in the abstract (e.g., An apparatus comprising an event handler…”).  Rather, perhaps a better alternative is to recite “An apparatus comprising … a memory encoded with program instructions that [perform some function]” and then a processor that executes those instructions.  This should provide sufficient structure to avoid § 112(f) interpretation.  But, support is important.  The specification should show the program in memory with a processor executing those instructions and flowcharts explaining the algorithm. 

Additionally, the Federal Circuit has looked at the patent drafter’s intent in determining whether to invoke § 112(f) or not.[43]  Therefore, a patent drafter would be wise to include a means-plus-function claim (i.e., using “means for…”) to show that they knew how to draft one and therefore the other apparatus claims should not be so construed.  Also, a mention during prosecution that your apparatus claims were drafted so as not to invoke § 112(f) can only help your case.

It is worth noting that apparatus claims are not the only claims that are subject to invalidity via § 112(f).  Method claims can also be tripped up by § 112(f)’s step plus function, as this author cautioned over 25 years ago.[44]  Practitioners should use caution and follow this issue. 

Conclusion

AI inventions may face more patent-eligibility difficulties than other technologies based on the EPG line of cases.  But where the specification and claims describe the technological advantages and other features that the Federal Circuit equates with patent eligibility, AI inventions should pass § 101 muster.  

Patent drafters should avoid functional claiming and avoid drafting functional specifications for the best chance of surviving § 101 challenges, but also to avoid potential § 112(f) issues.  Specifications should be detailed with (1) a system diagram showing the software programs in memory with a processor that executes those programs; (2) flowcharts explaining at a detailed level what each program does and how it does it; (3) diagrams and description explaining the major interfaces; and (4) diagrams and description explaining the major data structures.