chevron-down Created with Sketch Beta.
December 01, 2019 Feature

Informed Drafting and Prosecution of Software and AI Patents

Kate Gaudry and Rodney Rothwell

©2019. Published in Landslide, Vol. 12, No. 2, November/December 2019, by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association or the copyright holder.

Making an informed decision requires a circumstance in which actions have predictable results. For example, suppose that Action A is 90 percent likely to result in Result A and 10 percent likely to result in Result B, while Action B is 100 percent likely to result in Result B. A decision maker can then weigh the cost of Action A, the potential value of Result A, the cost of Action B, and the potential value of Action B to determine which action to take. In some instances, result predictions depend on a set of variables. Knowing these variables increases the degree to which an informed decision can be made. When these variables depend on luck or future events, risk-adverse decision makers will be more likely to avoid costly actions.

To give an analogy, suppose a student is enrolled in a calculus course and is determining whether to stay in the class. If she stays in the class and gets a good grade, it may dramatically help the strength of her transcript. If she drops the class, she can enroll in an easier class to avoid the potential of a poor grade. Staying in the calculus class would require substantially increased studying time compared to the easier class. Nearly any student in this situation would want to know how tough of a grader the calculus professor is prior to deciding whether to drop the class.

In the patent world, the decision at hand can be whether to file a patent application, and the results can indicate whether a patent is secured. In this instance, the result depends on how the U.S. Patent and Trademark Office (USPTO) interprets patent requirements. For example, whether a patent application is determined to comply with the nonobviousness requirement will depend on how broadly claim terms are interpreted and a degree to which a solid reasoning is required to combine disclosures from multiple references. If one could predict how strictly requirements would be applied, the strength of a given invention could then be assessed to determine whether to file an application. However, what happens when there is a broad distribution of interpretations? This will make it much more difficult to assess outcome probabilities for a given invention, and thus making an informed decision becomes more difficult.

Understanding the sequence of events that give rise to how examiners interpret requirements can shed light on why high variability may be observed. The fundamental requirements of securing a patent are defined by Congress. For example, 35 U.S.C. § 101 sets forth positive requirements for securing a patent, in that the invention must be a new and useful process, machine, manufacture, or composition of matter, or a new and useful improvement thereof. Courts are then faced with particular fact patterns to determine whether a patent has complied with the statutes, and new case law then develops. In fact, courts have interpreted the above patentability requirements to further include negative limitations, in that abstract ideas, laws of nature, and natural phenomena themselves cannot be patented. In subsequent decisions, the same or different courts interpret the new case law in view of different fact patterns (e.g., to define what constitutes an “abstract idea”). The USPTO regularly reviews court decisions. When precedential decisions are issued that are at odds with current examination guidelines, the agency updates the guidelines and trains the examiners. Supervisors at the USPTO take note of how the updated guidelines pertain to the technology types associated with their art unit, and they further train examiners in the art unit. Examiners then use the recent training (and guidelines) to examine individual applications. Thus, there are many stacked layers of interpretation, in which many individuals interpret the decisions and guidelines of people before them. Given that it is unlikely that any complex text will be interpreted precisely in the same manner by multiple people, this stacked interpretation frequently results in a high degree of variability at the final examiner layer.

Ripple Effects of Court Decisions through Parts of the USPTO

Another consequence of the stacked interpretation is that a new decision by a top court is likely to have effects that ripple through the courts and USPTO agencies. This ripple effect was triggered in 2014 when the U.S. Supreme Court issued its decision in Alice v. CLS Bank.1 In that decision, the Court held that the patent eligibility requirement was not satisfied when claims were directed to an abstract idea and were not significantly more than the abstract idea. Each of the four patents that were invalidated had been examined and issued by a business method art unit at the USPTO.

Shortly after the decision, examination statistics in the business method art units (and in the bioinformatics art unit) dramatically changed.2 Specifically, patent eligibility rejections became much more prevalent and allowance rates plummeted. Thus, a decision maker may have made a well-reasoned decision to file a patent application on a business method invention in 2013, though the tide change in 2014 would mean the prospects for securing a patent in the coming years were drastically reduced.

As mentioned, precedential decisions must be abided by other courts (and the same court). Thus, the Federal Circuit has had to adjust its statute interpretation in view of Alice and apply the new case law to different fact patterns. One such instance arose in Electric Power v. Alstom.3 The patents evaluated in this case related to techniques for detecting and analyzing events on an interconnected electric power grid. The patents had been examined in art units 2121 and 2125, which are in a different technology center than that of the business method art units. The Federal Circuit held the patents to merely be directed to collecting information, analyzing it, and displaying certain results, which it deemed to be an abstract idea. The artificial intelligence (AI) technology class corresponds to art units within technology center 2100 (with the prominent art unit being 2122). Though examination for this class was unaffected by Alice, examination statistics drastically changed after the 2016 Electric Power decision, with eligibility rejections increasing and allowance rates plummeting.

As a final exemplary “ripple,” the USPTO issued new eligibility guidelines in January 2019.4 The guidelines define an abstract idea as a mathematical concept, certain method of organizing human activity, or mental process. If an examiner were to contend that a given claim is directed to an abstract idea outside these three categories, approval from the technology center director is required. Further, the guidelines specify that even if a claim is directed to an abstract idea, if the idea is integrated into a practical application, the claim is to be found eligible. Allowance rates increased and eligibility rejections decreased following these new guidelines.

Statistics show how patent prospects can change over time. The unpredictability of these shifts can make it difficult to make decisions. Further complicating matters is variability across decision makers, as each examiner may have a slightly different interpretation of case law and agency guidance.

Even within Individual Art Units, Examiner Stats Can Vary Widely

We examined the percentage of issued office actions that include an eligibility rejection against the percentage of total actions that are allowances in the following art units: the bioinformatics art unit, the prominent AI art unit, an image-processing art unit, and a business method medical software art unit.

Prior to the Alice decision, there was a very wide spread across examiners within individual art units with respect to both variables. For example, the allowance prevalence within the AI art unit ranged from 0 percent to 60 percent, and the prevalence of eligibility rejections within the business method art unit ranged from under 10 percent to over 75 percent.

Following Alice, for the vast majority of examiners in the represented bioinformatics/AI and business method art units, allowances were very rare, and eligibility rejections were very common. In 2018, the AI examiners were issuing few allowances and many eligibility rejections. Notably, as the art unit statistics moved to one extreme of the range, the examiner spread shrank.

Like changes in case law, a decision maker cannot know at the time of a filing decision to which examiner an application will be assigned and the statistics of that examiner. For example, AI art units generally have relatively high allowance rates, which may be a consideration when determining whether to try to patent an AI innovation. However, even when case law shifts have not reduced allowance prospects, assignment to some examiners may mean that it will be very difficult to secure a patent.

Evaluating whether particular claims satisfy patent requirements is far from an exact science. Nonetheless, now that we have a better understanding of why high variability is observed in predicting whether a patent is secured, what lessons can be learned from post-Alice drafting and prosecution of software applications in the business method and bioinformatics art units to make more informed decisions on future filing and prosecuting of software patent applications?

Practice Tips in View of Examination Unpredictabilities

What Works Today May Not Work Tomorrow

Consider strategies employed by practitioners to overcome § 101 rejections in business method and bioinformatics art units, which have significantly evolved over time. Practitioners should not simply craft claims based on what is working at the moment but instead try to anticipate what may happen down the road and include support in applications for these “what if” moments. This can facilitate not only changes that may occur during prosecution of an application but also post-allowance changes that may affect validity of a resulting patent.

Track Case Law and Anticipate Examination Shifts

As illustrated by the above data, new case law, new USPTO guidelines, and new USPTO training all create shifts in how statutes are interpreted and the allowance rates of applications. By following early events such as new legislation and case law, practitioners can anticipate changes downstream in examination stringency and quickly adapt drafting and prosecution strategies. For example, it appears as though the USPTO attempts to specifically focus case-triggered updates to guidance and examiner training to art unit groups that correspond to those of patents reviewed in the cases. This may be due to logistical constraints or due to attempts to avoid overextending early case law. In any event, practitioners should attempt to understand trends in case law and to predict whether analysis from a given case may (through subsequent case law) affect validity or patentability assessments from other types of technology not at issue for the given case.

Know Your Audience

There is widespread variability in allowance prospects across art units and examiners. Practitioners should be armed with statistical information (e.g., allowance rates, average number of final office actions, average number of appeals, reversal rates for appeals, allowance rates after an examiner interview, etc.) pertaining to the art unit, the examiner, and the supervisory patent examiner (SPE) assigned to their application. Knowing this statistical information should allow practitioners to better devise prosecution strategies that have higher chances of securing an allowance with that particular art unit, examiner, and/or SPE. For example, overall reversal prospects on appeals across tech centers have been relatively predictable at less than 45 percent (split) and less than 35 percent (full), but drastically decrease when certain issues are raised such as § 101 (with some variation from art unit to art unit). Practitioners aware of this discrepancy can strategically select rejections with higher prospects for succeeding on appeal.

Delay Can Be a Viable Strategy

A best-case scenario is that practitioners prepare their applications and claims to withstand the most stringent type of foreseeable examination. However, this type of preparation is not always possible. In certain instances, the value of protecting the technology may be sufficiently high to justify relatively high prosecution costs or long prosecution time based on a prospect of securing limited market exclusivity. In these instances, delaying prosecution may be a viable strategy, particularly when current obstacles pertain to a volatile or uncertain issue (e.g., as patent eligibility has been across recent years). Delaying prosecution may allow other cases to be decided that add clarity about the legal interpretation of a given requirement, to allow newer guidance to pave the way for applying recent case law to different fact circumstances, and to allow examiners and supervisors time to gain comfort in applying new guidance in different ways based on different fact patterns. Delay tactics can include extending office action response deadlines (which requires a fee payment and can hasten any patent’s expiration) and appealing a rejection (given that the appeal board frequently takes a substantial amount of time to issue a decision on an appeal).

Exemplary Strategies in the AI Space

There are many lessons to be learned from the post-Alice evolution of case law and USPTO guidance and training. We would like to focus this section on how practitioners can transfer the lessons learned to a new context of AI. AI is a much newer technology type as compared to business methods. Societies are only now starting to adjust to the potential impacts of this technology. How long will it take for the law to come to grips with this new type of technology? For example, suppose an inventor devises a new type of machine-learning algorithm that is well designed to generate solutions to a vexing problem.

  • What can or should be patented? The algorithm or the solutions? Which better complies with existing patent requirements? And which better serves the objectives of the century-old patent system?
  • To what extent should or must the inventor understand why a machine-learning model outputs various results? Should or must the inventor understand how various combinations of different input data types influence outputs? Given the potential complexity of models, this understanding might be difficult or impossible to gain.
  • What are the advantages of seeking protection on each of: (1) training the model and running the trained model; (2) running a model that has been trained in a given way (without claiming the training); or (3) running a model that has particular parameters or characters determined as a result of training (again, without claiming the training)?

Vigilant practitioners can make efforts to be well positioned to predict how patent law may evolve in the context of AI technology, such that the practitioners can prospectively tailor drafting and prosecution techniques. Currently, very little case law exists that pertains to AI patents. Even though Electric Power seemingly influenced examination in the AI art units, it would be a stretch to say that the claims at issue in the case focused on AI technology.

However, many precedential cases pertain to software. These cases and the USPTO’s and examiners’ interpretations of the cases can be used to guide technique development for AI patenting. For example, as noted above, in January 2019, the USPTO interpreted various cases to indicate that claims merely directed to a mathematical concept, certain method of organizing human activity, or mental process are not patent eligible. The cases giving rise to this guidance generally did not relate to AI. However, the cases and guidance may indicate that an applicant trying to protect an AI innovation will be well served when a patent application illustrates why using the AI technique goes beyond what humans can do. Mere automation or reduced processing time is insufficient to demonstrate eligibility, but frequently AI techniques achieve higher accuracies and discover key predictors that are beyond what existing technologies and/or professionals can do. Emphasizing these achievements can facilitate showing that claims comply, not only with the eligibility requirement but also with the novelty and nonobviousness requirements.

Thus, general strategies can be shaped based on cases and USPTO guidance. More specific strategies for securing an allowance may further be developed to handle an individual patent application or respond to an individual examiner. These specific strategies may be developed based on information indicating the examiner’s allowance tendencies, any recent changes in allowance rates, and information about what types of claims the examiner has recently found to be allowable. Some of this data is available through examiner statistics software programs, and some can be collected via conversations with the examiner. For example, the examiner may personally be more persuaded by arguments that illustrate computational resource efficiencies achieved by an invention relative to accuracy improvements achieved by the invention. Practitioners may be able to the concentrate or focus amendments accordingly and may even provide a declaration by an expert that explains the advantage secured by the technique.

Practitioners may also be well-advised to keep in mind the pace at which AI is growing in its ubiquity. There were times at which having a computer perform a “simple” computation was exceedingly difficult, determining how to organize parallel computer processing was challenging, and dynamically availing virtual computer computation was complex. Now, a practitioner would have a difficult time arguing that a claim is patent eligible or nonobvious if the claim merely relates to implementing a known type of solution using any of these techniques. Judges and the USPTO have made particular note of this when assessing business method claims, as illustrated by some of the data presented above.

There is a parallel for AI innovations: now most of the public does not understand what neural networks are, how training and validation can be performed, what data standardization means, etc. But these concepts are quickly becoming more familiar and understood. Thus, it may soon be (and perhaps already is) insufficient to argue that claims are nonobvious and patent eligible in the AI space just because the invention is using a machine-learning technique for a different purpose. Rather, practitioners may soon need to more specifically indicate why AI innovations amounted to more than plugging a given data type into a standard model to generate an output. For example, perhaps a new type of preprocessing was performed to transform the data’s dimensionality in a manner well suited for the use case. Or perhaps a new type of accuracy assessment was performed, in which suboptimal results would selectively trigger new model execution using a different model type, different input data combinations, or different model parameters (selected in a new way). Rarely can machine-learning models be used in a plug-and-play fashion. A patent application today may be assessed in court 19 years from now. Concentrating on the complications encountered by the developer (and how they were overcome) may prevent a jury or judge from trivializing the innovation.

Allowance rates in the AI art units were quite high before Electric Power, but they seem to be returning to those levels. However, allowance prospects can drastically change over time, there is high variability in allowance rates across examiners, and new case law can quickly shift allowance prospects, as well as patents’ values and validity outlooks. Case law and examination statistics tell us that it is unlikely that one can secure a valid patent when an innovation merely includes using a computer to perform a technique that was previously performed by humans to get the same type of result. Practitioners and applicants in the AI space may want to revisit these decisions and mentally replace “computer” with “machine-learning model.” As AI technology becomes increasingly used and taught, patents that merely rely on the new combination of a given type of data with a given type of machine-learning model may end up being successfully challenged. It is becoming increasingly common to emphasize the details and complexities of selecting the given type of model, configuring it appropriately, and identifying pre- and post-processing techniques, along with emphasizing the technical advantages of these details.


As technologies, patent drafting styles, and patent prosecution approaches continue to evolve, so will the case law and examination approaches. A patent issued today may reach its peak value in well over a decade. However, it is not possible to predict how patent requirements will be defined and interpreted at that time. We recommend preparing for the best and hoping for the worst. If practitioners and applicants envision a time during which use of machine-learning models is as common as use of a computer is today, patent applications and prosecution may emphasize the complications, clever configurations, and advantages of their invention. Though depending on an examiner assignment, current guidance, and current case law, these explanations may turn out to be well beyond what was needed to secure an allowance. However, if the pendulum swings at the USPTO or at the courts, these details may turn out to be just what was needed to overturn a rejection or withstand a validity challenge.


1. Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 573 U.S. 208 (2014).

2. Statistics were generated based on raw data provided by LexisNexis® PatentAdvisorSM.

3. Elec. Power Grp., LLC v. Alstom S.A., 830 F.3d 1350 (Fed. Cir. 2016).

4. 2019 Revised Patent Subject Matter Eligibility Guidance, 84 Fed. Reg. 50 (Jan. 7, 2019).

The material in all ABA publications is copyrighted and may be reprinted by permission only. Request reprint permission here.

Kate Gaudry is a senior patent attorney at the Washington, D.C., office of Kilpatrick Townsend & Stockton LLP. She focuses her analysis on patent prosecution and counseling, with an emphasis on quantitative analysis of patent portfolios and strategic options and on prosecution for the software, computer systems, and quantitative biology industries.

Rodney Rothwell is counsel at Kilpatrick Townsend & Stockton LLP in Washington, D.C. He specializes in bioinformatics and machine learning patent prosecution and counseling.